-
扩散模型持续推动着文生图领域的进步。然而,将近期进展归因于特定的建模和数据选择颇具挑战性:最先进的开放权重模型提供的消融研究有限,且不公开其训练数据和完整的训练细节。研究社区需要完全开放(权重、数据和代码)的模型作为进一步研究的基础;然而,现有的完全开放模型在性能上仍远落后于领先模型。在本项目中,我们通过 300 多次受控实验(总计超过 70 万 TPU v6e 小时),对文生图扩散训练和推理中的建模与数据设计选择进行了系统性研究。我们的实验突出了若干实证发现(例如,在混合精选数据集时,等权加权是一个强大的默认设置)和简单的设计决策(例如,更大的文本编码器适配器能以极少的参数增加来提升性能),用于训练强大的模型。在这些见解的指导下,我们仅使用公开可用的数据集训练了 i1,一个拥有 30 亿参数的文生图扩散模型。i1 在五个代表性基准测试(GenEval、DPG、PRISM、CVTG-2K 和 LongText)上与领先模型具有竞争力,并且在平均得分上比现有最好的完全开放模型高出 29.5 个绝对百分点。我们提供了 i1 的检查点、训练和推理代码以及数据处理流程。我们的发现与 i1 方案共同为未来文生图扩散模型的开放研究奠定了实践基础。
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
目录
第一节 引言
自 DALL-E 2(ramesh2022hierarchical)、Imagen(saharia2022photorealistic)和 Stable Diffusion(rombach2022high)等早期模型以来,基于扩散的模型凭借其生成具有精细细节的逼真图像的强大能力,推动了文生图领域的重大进展(wu2025qwen; labs2025flux; cai2025z; gao2025seedream)。然而,尽管当今最先进的模型具备卓越能力,我们往往难以厘清究竟是哪些建模和数据选择真正推动了性能提升。这种不清晰性源于两个因素。
首先,领先模型通常不公开其训练数据和完整的训练方案(wu2025qwen; cai2025z; qin2025lumina; cai2025hidream),即便它们公开发布了模型检查点。这限制了可复现性,并阻碍了在其设计基础上开展受控分析和后续工作。虽然存在完全开源(权重、数据和代码)的模型(chen2025blip3; chen2025blip3o; ma2026deco; wang2026pixnerd),但它们在性能上远不及领先模型。
其次,领先模型通常不对其设计选择进行详尽的消融实验(wu2025qwen; cai2025z; cai2025hidream; gao2025seedream; ryu2025flite; fang2026flux)。在实践中,许多模型将大量架构、训练和数据决策捆绑到单一方案中,使得难以将性能提升归因于任何特定因素。因此,现代文生图扩散模型在诸多重要设计选择上仍缺乏共识(例如,使用单个文本编码器(qin2025lumina; wu2025qwen)与使用多个文本编码器(esser2024scaling; cai2025hidream))。
为了更深入地理解现有及新型架构与数据设计在文生图扩散模型中的影响,我们开展了一系列受控实验,主要聚焦于 256×256 低分辨率预训练阶段。从一个简单的基线模型(yao2025reconstruction)出发,我们首先探索了来自文本编码器的文本条件注入策略,以及噪声/时间步条件注入策略(sun2025noise)。随后,我们识别出能带来更强性能的主干架构设计(bao2023all;esser2024scaling)。最后,我们比较了高质量图文数据集筛选与推理时提示词增强方面的设计选择,以及混合图像数据集的策略。
在建模方面,我们发现:(1)使用单个强大的文本编码器配合更大的适配器,比组合多个文本编码器效果更佳;(2)时间步/噪声条件注入与自适应层归一化(peebles2023scalable)在我们的文生图设定中收益甚微;(3)采用长跳跃连接(bao2023all)的双流 DiT 架构(esser2024scaling)是一种强大的主干设计。
在数据方面,我们发现:(1)使用长描述进行训练能产生比短描述训练更强的模型,但会导致模型在短提示词上表现不佳,这一问题可通过推理时的提示词改写来缓解;(2)合成描述生成器的选择对下游性能至关重要;(3)在混合精选数据集时,对每个数据集按重复次数计算、使用相同数量的图像进行训练(下文称为“数据集间等权重分配”)是一种强大的默认策略;(4)在多样化数据集混合的情况下,重复训练数据仅会造成轻微的性能下降;(5)从低分辨率模型获得强大的高分辨率生成能力,并不需要广泛的高分辨率数据覆盖。
为了给未来的开放研究提供一个强有力的基线,我们利用受控实验中的洞察,在公开可用的数据集上训练了 i1——一个拥有 30 亿参数的文生图扩散模型。在 1024 分辨率下,i1 在完全开放模型中达到了最先进的性能,并在多个代表性基准测试中超越了多个参数量大得多的领先开放权重模型(例如,170 亿参数的 HiDream-I1 (cai2025hidream) 和 120 亿参数的 FLUX.1 [Dev] (labs2025flux))。
i1 表明,仅使用规模适中的公开图像数据集也能实现强大的性能,并凸显了仔细探索设计空间的价值:如图 4 所示,i1 没有引入任何显著的新网络模块,而是识别了先前工作中存在但未被充分利用的设计(例如,长跳跃连接),并对标准组件进行了简单的修改(例如,使用更大的文本编码器适配器)。我们提供了模型权重、代码、数据集以及模型训练和评估的详细配方。我们的发现与 i1 配方共同为开放的文生图研究奠定了实践基础,既提供了一个强大的完全开放基线,也为构建能力更强的模型提供了设计思路。
第 2 节 预备知识
在本节中,我们提供必要的术语和背景知识,以帮助理解和推动后续的受控实验设置(第 3 节)以及建模与数据设计实验(第 4 节和第 5 节),重点关注骨干架构、文本与噪声条件机制,以及现有的开放训练数据配方。
文生图骨干架构
尽管存在其他范式(chang2023muse; sun2024autoregressive; zhou2024transfusion),但大多数领先的文生图系统都采用基于流匹配(lipman2022flow)训练的扩散Transformer(DiT)(peebles2023scalable)。根据文本特征融入方式的不同,当前的扩散模型通常可分为三类:交叉注意力模型(chen2024pixart; xie2025sana; ryu2025flite)、单流模型(qin2025lumina; cai2025z; chen2025dit)和双流MMDiT模型(esser2024scaling; wu2025qwen)。交叉注意力模型通过交叉注意力层注入文本嵌入向量,而单流和双流模型则将图像与文本的token序列进行拼接。双流模型对图像和文本token使用各自模态专用的注意力机制和MLP参数,而单流模型则在各模态间共享注意力机制和MLP参数。长跳跃连接是一种架构上的改进,它在早期层与后期层之间添加了快捷连接。早期工作(bao2023all)曾对此进行过探索,但在现代文生图模型中并未得到广泛应用。
文本与噪声条件机制
在当前的模型中,输入提示词由一个(cai2025z; qin2025lumina; wu2025qwen)或多个(esser2024scaling; cai2025hidream)文本编码器进行编码。由此产生的文本特征通常会经过一个线性层(labs2025flux; cai2025hidream; wu2025qwen; cai2025z)或MLP(xie2025sana)适配器,将其映射到扩散模型的隐藏维度。在各类主干架构中,自适应层归一化(AdaLN)(peebles2023scalable)被普遍用于注入时间步信息。AdaLN学习一个从时间步嵌入向量到注意力机制和MLP输入的缩放因子与偏移因子的线性投影,以及针对其输出的门控因子。在某些模型中,时间步嵌入向量会与池化后的文本嵌入向量通过逐元素相加的方式结合,然后再用于AdaLN条件处理(esser2024scaling; cai2025hidream; labs2025flux)。
开放文生图数据配方
许多领先模型(wu2025qwen; cai2025z; qin2025lumina; cai2025hidream; labs2025flux)会公开其权重,但并未披露训练数据的配方。除少数模型(qin2025lumina; ryu2025flite)外,甚至连训练数据集的来源和规模也仍未公开,这限制了开放研究社区对如何构建高质量文生图训练数据的理解。完全开放的模型(chen2025blip3; chen2025blip3o; sehwag2025stretching; ma2026deco; tong2026scaling; wang2026pixnerd)在性能上通常仍落后于领先系统,且其数据集往往来自相似且有限的来源(例如 JourneyDB (sun2023journeydb)、SA-1B (kirillov2023segment) 和 CC12M (changpinyo2021conceptual))。此外,完全开放的配方几乎不涉及数据平衡技术的探索。
第 3 节 受控实验的基线
在第 2 节介绍现有设计的基础上,我们通过第 4 节和第 5 节中 256 分辨率预训练阶段的受控实验,对文生图扩散模型的设计空间进行研究。对于每组实验,我们都从相同的强基线出发,并独立改变单一设计选择(即,各实验间的修改不进行累积)。性能得到提升的设计将在第 6 节中组合,以构建我们的最终模型 i1(见图 21)。下文描述了基线设置,图 5 提供了高级示意图,详细架构则绘制于附录 A.2 中。
模型
如图5所示,我们的骨干架构基于LightningDiT-XL/2(yao2025reconstruction)。LightningDiT是一种现代DiT架构(peebles2023scalable),它融合了用于提升性能的常见设计(例如RoPE(su2024roformer)、RMS归一化(zhang2019root)、SwiGLU FFN(shazeer2020glu))。我们添加了QK归一化(dehghani2023scaling)以稳定训练。为确保后续消融实验能与强基线进行对比,我们还应用了长跳跃连接(bao2023all)(见第2节),这是一种不太常用的设计,我们在第4.2节中重新审视了它,并发现它有助于提升性能。
默认情况下,我们使用交叉注意力机制来注入文本嵌入。在某些实验中,我们还会在骨干网络的单流和双流变体(见第2节)上进行额外验证,以确保我们发现的通用性。我们使用AdaLN将模型条件设置为时间步嵌入与池化文本嵌入(通过对文本嵌入token求平均计算得出)之和。对于单流架构,我们遵循Lumina-Image 2.0(qin2025lumina)的做法,在骨干网络前添加两个特定于模态的精炼模块。对于单流和双流骨干网络,我们都采用了多模态RoPE(wang2024qwen2)。默认情况下,我们使用T5Gemma-2B的编码器部分作为文本编码器,并使用FLUX.2 VAE。
| ImageNet-22K | YFCC | RedCaps | Megalith |
| Places | Pexels | iNaturalist | FLUX-Reason |
| Midjourney v6 | GPT-Edit | TextAtlas | RenderedText |
数据
我们仅使用公开可用的图像数据集,包括 7 个真实图像数据集(ImageNet-22K、YFCC100M、RedCaps、Megalith、Pexels、iNaturalist 2024、Places365-Challenge 2016)、3 个合成数据集(GPT-Image-Edit-1.5M、FLUX-Reason-6M 和 Midjourney v6),以及 2 个文本渲染数据集(RenderedText 和 TextAtlas)。默认情况下,我们简单地将这些数据集中的 1.68 亿张图像合并,不进行加权处理(这一设计选择将在后续第 5.2 节中重新审视)。在预训练中,所有图像均被中心裁剪为正方形,并调整大小为 256x256。我们在图 6 中展示了每个数据集的示例图像。我们使用 Qwen3-VL-30B-A3B,在 FP8 精度下,通过提示词“用一段话详细描述该图像。”为每张图像生成一条长的合成描述(该视觉语言模型接收中心裁剪为正方形、若原图大于 512x512 则调整大小为 512x512 的图像)。更多信息见附录 E。
| 基线变体 | DPG | PRISM | LongText |
| 交叉注意力 | 84.66 | 56.4 | 0.211 |
| 单流 | 85.89 | 55.6 | 0.293 |
| 双流 | 86.82 | 58.3 | 0.439 |
训练与推理
我们使用流匹配目标训练模型,共进行 50 万次迭代(即最终 i1 模型 200 万步 256 分辨率预训练阶段的 25%),批次大小为 512,学习率为 1e-4。我们使用 250 步欧拉积分器进行采样,无分类器引导尺度设为 12。更多细节见附录 A.1。
提示词:在一张反光的金属桌子上,有一个色彩鲜艳、带有花卉图案的手提包,旁边是一个刚切开的牛油果…… 餐具和一个透明玻璃水瓶整齐地摆放在牛油果旁边……(77 词)

提示词:Lindsey Wixson 自信地站在一片金色麦田中,头戴宽檐草帽,佩戴醒目的红色太阳镜,身穿一件鲜艳的红色毛边上衣,搭配一条闪耀的钻石项链,尽显夏日优雅。

提示词:一款名为“DailyFlow”的优雅、专业外观的移动端生产力与习惯追踪应用界面。应用名称以粗体圆润字体、舒缓的蓝绿色显示,醒目地位于顶部……(170词)

评估
。我们使用三个广泛采用的基准测试来为控制实验提供信号:DPG-Bench、PRISM-Bench 和 LongText-Bench。这三个基准测试均使用 VLM 作为评估器。DPG 和 PRISM 衡量跨多样化提示词的细粒度提示跟随能力,其中 PRISM 额外评估图像美学。LongText 专门评估文本渲染能力。我们在图7中展示了每个基准测试的示例提示词。我们使用 DPG 和 LongText 的原始提示词,并使用 Qwen3-4B 配合第5.1节中的简单元提示词重写了 PRISM 的提示词。表1报告了我们的基线模型在这些基准测试上的性能。
第4节 建模
我们研究了对第3节中介绍的基线模型进行的建模设计修改。我们首先重新审视文本和噪声调节机制,包括多个文本编码器和 AdaLN,并确定了更强的替代设计方案。然后我们探索了主干架构的选择。更多结果见附录C。
4.1 文本与噪声调节
现有方法已探索了将文本和噪声条件融入骨干扩散模型的各种方式。有些方法使用单一文本编码器(cai2025z; qin2025lumina; wu2025qwen),另一些则拼接来自多个文本编码器的特征(esser2024scaling; cai2025hidream)。此外,噪声水平的嵌入向量通常通过 AdaLN 注入模型(peebles2023scalable),有时还会与池化后的文本嵌入向量一起使用(esser2024scaling; cai2025hidream; labs2025flux)。我们研究了这一广阔的设计空间,并表明,与其组合多个编码器,不如使用一个更强的单一文本编码器配合更大的适配器。我们还发现,基于 AdaLN 的噪声水平和池化文本嵌入条件对于文生图模型可能并非必要。
文本编码器
早期模型(例如 SD 1.5(rombach2022high))主要使用 CLIP 风格的文本编码器。后来的模型采用了编码器-解码器模型 T5(raffel2020exploring; esser2024scaling; cai2025hidream)。最近,模型常使用仅解码器的大语言模型或视觉语言模型(wu2025qwen; qin2025lumina),这一趋势通常归因于它们强大的推理能力和复杂指令遵循能力(xie2025sana)。
在此,我们比较了(1)一个基于现代 CLIP 的编码器(FG-CLIP 2(xie2025fg)),(2)两个现代编码器-解码器模型系列,T5Gemma(zhang2025encoder)和 T5Gemma2(zhang2025t5gemma),(3)一个现代仅解码器大语言模型系列(Qwen3(yang2025qwen3)),以及(4)一个现代仅解码器视觉语言模型系列(Qwen3-VL(bai2025qwen3))。除非另有说明,我们在可用时使用指令微调后的检查点,否则使用相应的基础检查点。图 8 报告了使用每个模型作为文本编码器时的文生图性能(其他设置下的比较见附录 C.3 和 C.4)。
我们观察到,指令微调的影响微乎其微(例如,T5Gemma-2B 与 T5Gemma-2B(基础版)对比),并且更大的模型并不一定表现更好(例如,T5Gemma-2B 与 T5Gemma-9B 对比)。最重要的是,编码器-解码器模型(T5Gemma 和 T5Gemma2)取得了最佳的整体性能,超越了 FG-CLIP 2 以及仅解码器的大语言模型和视觉语言模型。这引出了影响我们设计的第一个发现:
结合文本编码器
。许多近期模型(esser2024scaling; cai2025hidream)会结合来自多个编码器(例如 CLIP、T5 和大语言模型)的文本特征。在此,我们实验了图 8 中评估的不同文本编码器组合。先前的工作采用不同策略来组合来自不同编码器的嵌入向量(例如,沿嵌入维度拼接与沿序列维度拼接)。在本工作中,我们沿序列维度拼接文本特征,并为每个编码器使用独立的适配器,以适应它们不同的嵌入维度。这避免了将文本特征填充到统一序列长度的需要——当沿嵌入维度拼接特征时,通常需要这样做。
我们将图 8 中最强的文本编码器之一 T5Gemma-2B 与另一个额外的编码器相结合。如表 2 所示,将其与 T5Gemma2-1B 或 FG-CLIP 2 结合能取得最佳性能。然而,将三者(T5Gemma-2B、T5Gemma2-1B 和 FG-CLIP 2)全部结合并未带来显著的进一步提升。
| 类型 | T5G-2B | T5G2-1B | T5G2-4B | Qwen3-VL-2B | FG-CLIP 2 | DPG | PRISM | LongText |
| 基线 | ✓ | 84.66 | 56.4 | 0.211 | ||||
| +1 编码器 | ✓ | ✓ | 85.62 | 58.4 | 0.303 | |||
| ✓ | ✓ | 84.86 | 55.8 | 0.270 | ||||
| ✓ | ✓ | 84.80 | 57.4 | 0.264 | ||||
| ✓ | ✓ | 85.72 | 57.7 | 0.285 | ||||
| +2 编码器 | ✓ | ✓ | ✓ | 85.37 | 58.8 | 0.272 | ||
| ✓ | ✓ | ✓ | 85.28 | 58.1 | 0.351 |
尽管结合多个文本编码器能提升性能,但这种提升究竟源于不同编码器提供的多样化表征,还是仅仅来自序列长度增加以及适配器引入的额外参数?为探究此问题,我们构建了两个基线,对 T5Gemma-2B 文本嵌入进行重复:第一个基线对两份相同的嵌入副本使用两个独立的适配器(从而同时增加序列长度和适配器参数),第二个基线则使用共享适配器(因此仅增加序列长度)。如表 3 所示,使用两个独立适配器重复嵌入带来了显著提升,而使用共享适配器的结果与未重复的基线相似。这表明,结合多个文本编码器带来的增益可能主要源于额外的适配器,而非多样化的文本编码器特征或更长的序列。
| 文本编码器 | DPG | PRISM | LongText |
| T5Gemma-2B | 84.66 | 56.4 | 0.211 |
| 重复 w/1 MLP | 84.93 | 55.8 | 0.225 |
| 重复 w/2 MLP | 85.09 | 56.5 | 0.309 |
我们通过比较来自每个适配器的 DPG-Bench、PRISM 和 LongText 提示词的嵌入,来验证两个 MLP 适配器确实学习了不同的特征。我们测量每个 token 的两个嵌入之间的余弦相似度,并对每个提示词中所有 token 取平均。如图 9 所示,得到的分布大部分为负值,表明两个适配器确实捕捉到了不同的表征。
更大的文本编码器适配器
为了进一步探究组合多个文本编码器带来的性能提升是否主要源于额外的适配器参数而非多样化的文本编码器特征,我们将默认设置(第3节)中使用的小型MLP适配器(260万参数)替换为与主干模块宽度相同的大型Transformer适配器(每块1720万参数)。如图10所示,尽管参数量增加幅度不大,但提升适配器容量在所有主干架构上均能持续改善性能。不过,将适配器扩展到两个Transformer模块以上后,带来的额外收益就非常有限了。
此外,如表4中“default”和“+2 encoders”两行所示,当使用更大的适配器时,组合多个文本编码器在所有主干上带来的收益都大幅减小,尤其是在DPG和PRISM指标上。这进一步表明,多个文本编码器的优势可以通过增加单个文本编码器的适配器容量来获得。重要的是,使用多个编码器会增加文本序列长度,从而大幅提升内存和计算成本,而使用更大的适配器则不会。
| MLP适配器(默认) | Transformer适配器(1个模块) | ||||||||
| 参数量 | DPG | PRISM | LongText | 参数量 | DPG | PRISM | LongText | ||
| 交叉注意力 | |||||||||
| 默认 | 0.89B | 84.66 | 56.4 | 0.211 | 0.91B | 86.33 | 58.7 | 0.414 | |
| +2个编码器 | 0.90B | 85.37 0.71 | 58.8 2.4 | 0.272 0.061 | 0.94B | 86.47 0.14 | 59.5 0.8 | 0.491 0.077 | |
| 无池化嵌入 | 0.89B | 85.98 1.32 | 57.5 1.1 | 0.391 0.180 | 0.91B | 86.37 0.04 | 59.4 0.7 | 0.446 0.032 | |
| 无时间步 | 0.89B | 82.58 2.08 | 54.7 1.7 | 0.185 0.026 | 0.91B | 84.71 1.62 | 58.9 0.2 | 0.418 0.004 | |
| 无AdaLN | 0.66B | 84.99 0.33 | 57.4 1.0 | 0.351 0.140 | 0.67B | 85.13 1.20 | 59.7 1.0 | 0.413 0.001 | |
| 单流 | |||||||||
| 默认 | 0.82B | 85.89 | 55.6 | 0.293 | 0.83B | 87.64 | 60.0 | 0.472 | |
| +2个编码器 | 0.83B | 84.89 1.00 | 56.3 0.7 | 0.439 0.146 | 0.87B | 87.29 0.35 | 59.0 1.0 | 0.428 0.044 | |
| 无AdaLN | 0.57B | 87.38 1.49 | 59.0 3.4 | 0.390 0.097 | 0.58B | 87.39 0.25 | 59.5 0.5 | 0.410 0.062 | |
| 双流 | |||||||||
| 默认 | 1.24B | 86.82 | 58.3 | 0.439 | 1.25B | 87.67 | 60.7 | 0.576 | |
| +2个编码器 | 1.25B | 87.34 0.52 | 59.6 1.3 | 0.514 0.075 | 1.29B | 87.76 0.09 | 60.8 0.1 | 0.588 0.012 | |
| 无 AdaLN | 1.01B | 87.82 1.00 | 60.3 2.0 | 0.508 0.069 | 1.02B | 87.38 0.29 | 60.7 0.0 | 0.554 0.022 | |
移除 AdaLN 条件化
自适应层归一化(AdaLN)(peebles2023scalable)是现代文生图扩散模型(labs2025flux; wu2025qwen; qin2025lumina; cai2025hidream)中的标准组件。它通常用于将时间步嵌入和池化文本嵌入注入到主干网络中。最近一项研究(sun2025noise)表明,在类别条件图像生成中,移除噪声条件化对性能的影响极小,尤其是对于流匹配模型。如果能在不损害性能的情况下移除 AdaLN,模型将变得更加参数高效。
有趣的是,如表 4 所示,当文本编码器适配器是一个小型 MLP 时,从默认设置(第 3 节)中移除 AdaLN 能持续提升性能。然而,当使用更大的 Transformer 适配器时,这种效果变得非常微弱。为了更好地理解这一行为,我们对交叉注意力主干网络进行了额外的消融实验,其中 AdaLN 仅对池化文本嵌入或仅对时间步嵌入进行条件化,而非对两者之和进行条件化。我们分别使用小型和大型适配器评估了这些变体。
我们发现,当适配器较小时,对池化文本嵌入进行 AdaLN 条件化会降低性能(DPG 上 84.99 vs 82.58),但当适配器较大时,其影响则小得多(DPG 上 85.13 vs 84.71)。这表明,移除 AdaLN 带来的性能提升可能主要源于使用小型 MLP 适配器时特征学习不佳。然而,即使使用更大的适配器,通过 AdaLN 对池化文本和时间步嵌入进行条件化处理,其带来的额外收益仍然微乎其微。
4.2 主干架构
在本小节中,我们重新审视了长跳跃连接设计,并基于第 3 节的基线设置,对主流主干架构系列进行了受控对比。我们在附录 C.1 和 C.2 中包含了关于位置嵌入、归一化和 VAE 的额外分析。
长跳跃连接
在早期层和后期层之间添加捷径。它们最初由 U-Net(ronneberger2015u)推广,后来被应用于 U-ViT(bao2023all)中的扩散模型。虽然已有研究表明它们能提升性能(bao2023all; li2024hunyuan; liu2024playground),但它们尚未被广泛应用于现代文生图模型。在图 11 中,我们通过训练基线的双流变体(第 3 节)来重新审视这一设计,这些变体在多种模型宽度(1152、1296、1440、1584 和 1728)下,分别采用和不采用长跳跃连接,同时保持所有其他配置不变。我们发现,长跳跃连接能持续提升不同规模模型的性能,这可能是由于增强了模型的表达能力。基于 FLOPs 的额外分析见图 44,其他主干架构的结果见附录 C.7。
主干架构系列
当今领先的模型在主干网络的选择上各不相同:有的采用交叉注意力,有的采用单流架构,还有的采用双流架构(详见第 2 节)。我们在多种模型宽度下(对于所有三种主干网络家族,宽度分别为 1152、1296、1440、1584 和 1728),同时保持所有其他模型配置不变,测量了交叉注意力、单流和双流主干网络的模型性能。图 12 绘制了模型性能与参数量的关系。我们观察到,双流主干网络实现了最佳的性能-参数量权衡。
第 5 节 数据
除了模型架构之外,高质量的图像-标题数据对于文生图训练也很重要。在本节中,我们首先研究合成标题的设计,并表明在长标题上训练能产生更强的模型,但可能导致在短提示词上表现不佳,我们通过在推理时进行提示词重写来缓解这一问题。接着,我们探索了数据集混合,并发现跨数据集等权重分配是一个稳健的默认方案。
5.1 合成标题与提示词重写
文生图模型中的提示词遵循能力从根本上依赖于训练数据中高质量的图像-标题对。早期的工作,如 Parti (yu2022scaling) 和 DALL-E 3 (betker2023improving),表明在由视觉语言模型生成的高度描述性合成标题上进行训练可以显著提升性能。自那以后,大多数文生图模型 (chen2024pixart; esser2024scaling; qin2025lumina; xie2025sana) 都在训练中利用了合成标题。
在此,我们探讨了合成字幕生成中的若干设计选择及其对模型性能的影响(更多结果见附录 D.1)。我们特别发现,使用长合成字幕进行训练能产生更强的模型,但这些模型在短提示词上可能表现不佳,因此需要在推理时进行提示词重写。为降低字幕生成的计算成本,本节所有实验均在 ImageNet-22K 数据集上进行,而非默认基线设置(第 3 节)中使用的完整训练集。
为探究字幕质量如何影响下游文生图性能,我们使用五种视觉语言模型生成字幕:Qwen2-VL 2B、Qwen2.5-VL 3B、Qwen3-VL-2B、Qwen3-VL-4B 和 Qwen3-VL-30B-A3B。如图 13 所示,合成字幕生成器的选择对下游文生图性能有显著影响。我们注意到,LongText 上的微小差异主要源于在 ImageNet-22K 上训练的模型整体文本渲染能力较差,因为该数据集包含的文本丰富图像很少。因此,这一结果并不意味着字幕生成器的质量对文本渲染不重要。
默认情况下,我们仅使用长合成字幕训练模型(见第 3 节)。虽然我们的模型在原始 DPG 和 LongText 提示词上能取得强劲性能,但在原始 GenEval 提示词上表现不佳。我们发现,这可能是由于 GenEval 中的提示词比 DPG 和 LongText 短得多(见图 35):仅将 GenEval 提示词重复 12 次就能带来性能的大幅提升(0.17 提升至 0.49)。这一观察结果表明,在原始短 GenEval 提示词上的糟糕表现可能源于仅使用长字幕进行训练。
为了进一步理解这一点,我们使用提示词“用一句话描述该图像”生成了另一组短描述。提示词长度的分布如图14所示。我们将这些短描述与原始长描述按不同采样权重混合,并在表5中报告了由此得到的GenEval评分。我们观察到:(1) 主要使用短描述进行训练(例如,长描述占比0%或20%)能提升在原始短GenEval提示词上的表现;(2) 当GenEval提示词被重复时,训练数据中长描述比例越高的模型表现越好。
| 训练描述中长描述的占比 | GenEval提示词上的表现 | ||||
| 原始提示词(短) | 重复提示词 | 重写提示词(长) | |||
| 4 | 12 | 20 | |||
| 0% | 0.47 | 0.55 | 0.34 | 0.24 | 0.60 |
| 20% | 0.47 | 0.54 | 0.53 | 0.50 | 0.67 |
| 40% | 0.35 | 0.59 | 0.55 | 0.54 | 0.70 |
| 60% | 0.37 | 0.60 | 0.57 | 0.54 | 0.73 |
| 80% | 0.26 | 0.57 | 0.54 | 0.47 | 0.73 |
| 100% | 0.17 | 0.48 | 0.49 | 0.46 | 0.73 |
虽然重复短提示词可以恢复性能,但这会引入不自然的提示词结构。为解决此问题,我们改用一个大语言模型(Qwen3-4B),通过以下元提示词对GenEval提示词进行重写:
“我有一个简短的文生图提示词 {prompt}。请将其扩展为一段描述性文字,同时确保生成的图像仍然清晰地包含原始提示词中提到的所有内容。请仅输出重写后的提示词,不要输出其他任何内容。”
如表 5 最右列和图 15 所示,改写 GenEval 提示词能显著提升模型性能。值得注意的是,使用长描述进行训练并在改写后的提示词上进行评估(0.73),其表现明显优于使用短描述训练并在原始、重复或改写后的提示词上进行评估。这表明,即使推理提示词原本较短(例如 GenEval),也更倾向于使用长描述进行训练,并通过增加推理提示词长度(例如通过提示词改写)来匹配训练分布,而不是使用短描述训练来匹配原始的推理提示词长度。
| 提示词:一张葡萄酒杯和一只熊的照片 | ![]() | ![]() | ![]() | ![]() |
| 提示词:一张停车计时器右侧斑马的照片 | ![]() | ![]() | ![]() | ![]() |
| 训练:短描述,测试:原始(短描述) | 训练:长描述,测试:原始(短描述) | 训练:长描述,测试:重复 12 次(长描述) | 训练:长描述,测试:改写后(长描述) |
5.2 数据混合
到目前为止的所有实验都是将所有数据集简单组合,没有进行显式的数据集级别加权。由于我们的训练语料库(见第 3 节)高度不平衡(例如,YFCC 在 1.68 亿张图像中贡献了 9800 万张),我们隐式地将更大的权重分配给了少数几个大型数据集,这可能会主导训练信号。在本小节中,我们研究数据集组成和数据集级别重新加权如何影响性能。
数据集组成部分的贡献
为了理解每个数据集对性能的贡献,我们在每个数据集上分别训练了一个模型。评估结果如图 16(a) 所示。在真实图像数据集中,ImageNet-22K 和 YFCC 取得了最佳的整体性能,而 iNaturalist 的表现则明显较差,这可能是由于其领域范围狭窄所致。FLUX-Reason 和 GPT-Edit 在 PRISM 上的表现尤为出色。除 TextAtlas 外,每个数据集在 LongText 上的得分都很低,这与真实和合成数据集中包含文本的图像稀缺的情况一致。这表明文本渲染能力依赖于专门的、富含文本的数据集。为了控制数据集大小,我们进一步在每个数据集的随机 100 万子集上训练模型。如图 16(b) 所示,相对性能趋势与完整数据集的情况基本保持一致。
此外,我们测试了是否可以在不损害性能的情况下,从基线(第 3 节)中移除真实数据、合成数据或文本渲染数据。如图 17 所示,移除真实图像会损害 DPG 性能,而移除合成图像则会损害 PRISM 性能。此外,LongText 性能与文本渲染数据的比例直接相关(“完整”为 10.4%,“移除真实”为 66.7%,“移除合成”为 10.9%,“移除文本”为 0%)。这些结果表明,这三组图像提供了互补的益处。
数据集等权重
默认情况下(第 3 节),我们简单地将所有数据集合并,不进行显式的数据集级别加权,因此每个数据集的有效采样权重就是其图像数量。受 VLM 训练中对单一来源数据点数量设置上限的数据平衡策略启发(tong2024cambrian),我们使用四个手动选取的阈值对每个数据集的采样权重进行上限设定。图 18 的结果显示,阈值为 120 万(即所有数据集权重相等)时,整体性能表现强劲。
鉴于数据集等权重分配的有效性(见图 18),我们进一步探索了两种简单变体。首先,我们逐一移除低质量的真实数据集,同时保持其余数据集权重相等。表 6 显示,移除 iNaturalist 在所有基准测试中均带来了明显提升,而进一步移除其他真实数据集则没有带来实质性改善。其次,在移除 iNaturalist 之后,我们测试了是否应通过将某个数据集的权重提高 3 倍或 5 倍,同时保持其余数据集权重相等,来强调该单一数据集。如图 19 和图 46 所示,提高任何单一数据集的权重都无法超越完全平衡数据集的性能。
| 数据集 | DPG | PRISM | LongText |
| 完整 | 85.14 | 58.2 | 0.335 |
| 移除 iNaturalist | 85.56 | 58.7 | 0.384 |
| 移除 iNaturalist + Megalith | 85.13 | 59.0 | 0.438 |
| 移除 iNaturalist + Megalith + Places | 85.18 | 57.9 | 0.453 |
数据规模
图 16 提供了初步证据,表明对数据集进行子采样通常对模型性能影响甚微。为了探究性能在多大程度上依赖于训练集中独特图像的数量,我们还在 ImageNet-22K 的随机子集上进行了训练。如图 20 所示,尤其是在每张图像使用 5 条描述的情况下,将图像数量从 1370 万子采样至 40 万仅会导致轻微的性能下降。只有当图像数量缩减至 10 万时,我们才观察到显著的性能下降。由于我们的 50 万步训练方案已经将完整的 ImageNet-22K 数据集重复了 18.7 次,这些结果表明,对于文生图扩散模型而言,使用更少的独特图像并更频繁地重复它们,可能不会显著损害模型性能。
| 每个数据集的子集大小 | 已见独特图像数量 | DPG | PRISM | LongText |
| 完整 | 8810 万∗ | 85.56 | 58.7 | 0.384 |
| 100 万 | 1100 万 | 85.34 | 57.7 | 0.384 |
| 40 万 | 440 万 | 84.67 | 57.7 | 0.382 |
| 10 万 | 110 万 | 84.71 | 57.4 | 0.349 |
我们进一步将单个数据集上的数据子采样实验扩展到了数据集混合场景。具体来说,我们从一种数据混合方案开始,该方案对除 iNaturalist 之外的 11 个数据集赋予相同权重。对于混合中的每个数据集,我们随机对其进行子采样,使其恰好包含 1.0M、0.4M 或 0.1M 张图像,同时保持各数据集之间的采样权重相等。由此产生的模型性能如表 7 所示。即使每个数据集缩减到 0.4M 张图像(导致看到的唯一图像数量为 4.4M,而非 88.1M),各基准测试的性能下降也微乎其微。这表明,在多样化的数据集混合中,重复训练数据在文生图扩散模型训练中仅会导致边际性能损失。
第 6 节 i1-3B:完全开放模型中的顶尖性能
在前面的章节中,我们探索了能够提升文生图性能的建模和数据处理方案。基于这些见解,我们训练了 i1,一个拥有 3B 参数的模型,其在多个代表性基准测试中与领先模型性能相当。111我们还在额外训练一个 1B 模型,并将很快发布。在本节中,我们将描述最终的预训练、高分辨率训练和推理设置及实验,并展示评估结果。
6.1 低分辨率预训练
模型
i1 的架构如图 21 所示。它采用带有长跳跃连接的双流 MMDiT 主干网络、FLUX.2 VAE 以及 T5Gemma-2B 作为文本编码器,并配有一个由两个 Transformer 模块组成的大型适配器。i1 移除了所有 AdaLN 参数,因此不使用噪声条件化。此外,我们同时使用了正弦位置嵌入和 RoPE 位置嵌入,并在文本流和图像流之间共享 Sandwich 归一化(相应的对照实验见附录 C.1)。
数据
我们采用了第 5.2 节中确定的最佳数据混合方案,其中对 6 个真实图像数据集、3 个合成数据集和 2 个文本渲染数据集分配了相同的权重。我们使用 Qwen3-VL-30B-A3B 为每张图像生成多个长合成描述。由于资源限制,我们为 ImageNet-22K、Pexels、RenderedText、GPT-Edit、RedCaps、FLUX-Reason、TextAtlas 和 Midjourney v6 每张图像生成五条合成描述,为 YFCC 每张图像生成两条,为 Places 和 Megalith 每张图像生成一条。
| 提示词:阿根廷足球巨星莱昂内尔·梅西在 2022 年 FIFA 世界杯决赛对阵法国队的激烈比赛中。他……正准备用左脚射门……(240 词) | ![]() | ![]() | ![]() | ![]() |
| 提示词:一张吸引人的海报……宣布一场民谣音乐会活动……在顶部中央,写着诱人的短语“让原声旋律启迪你的灵魂”……(109 词) | ![]() | ![]() | ![]() | ![]() |
| 10 万次迭代 | 20 万次迭代 | 50 万次迭代 | 200 万次迭代 |
训练
我们将默认训练方案(第3节)中的训练迭代次数延长至200万步,同时保持所有其他超参数不变。我们以256分辨率训练i1,直到性能在大约200万步时趋于稳定,如图22所示。我们还在图23中展示了示例生成图像,并观察到基准测试的改进伴随着图像质量的提升。训练设置和计算资源的详细信息见附录A.1。
6.2 高分辨率训练
数据与建模
。为了构建512和1024分辨率的训练集,我们仅保留短边分别至少为512或1024像素的图像。如果过滤后的集合包含少于30万张图像,我们将完全移除该数据集(每个数据集的分辨率统计信息见附录E.2)。基于我们在第5.2节中的发现,我们进一步对每个超过100万张图像的数据集进行子采样至100万张,并为每个数据集分配相等的采样权重。在1024分辨率下,我们因质量低下而丢弃了RenderedText数据集(见图16)。遵循esser2024scaling的方法,我们在512和1024分辨率训练期间执行位置索引插值和时间步调度偏移(详细信息见附录A.1)。
结果
。我们在512分辨率下训练模型50万步,在1024分辨率下训练30万步。训练期间的基准测试性能趋势见附录A.1,最终的1024分辨率检查点在6.3节中进行评估。如图24所示,512分辨率训练显著提高了LongText分数(0.75 → 0.92)。我们进一步在图25中通过定性示例展示了文本渲染方面的改进。
我们还研究了不同数据集组成部分对 512 分辨率训练的贡献。从 256 分辨率检查点出发,我们分别使用仅真实图像数据集、仅合成图像数据集或仅文本渲染数据集在 512 分辨率下训练单独的模型。如图 24 所示,仅使用真实或合成图像数据集进行训练,在 LongText 指标上取得的提升与使用完整数据集训练相当,尽管这两个子集包含的富文本图像数量有限。这表明,强大的高分辨率生成能力并不需要高分辨率训练数据来匹配低分辨率预训练数据的全部广度。
![]() | ![]() |
![]() | ![]() |
6.3 推理与评估
推理设置
。在推理过程中,我们使用 CFG 缩放系数 12,并应用 Rescale CFG 技术(lin2024common),重缩放强度为 1。与之前的方法(wang2024emu3; deng2025emerging; pan2025transfer)针对特定基准应用提示词重写不同,我们使用单个元提示词(详见附录 B.3)来重写所有输入提示词,使其匹配训练提示词长度,其动机已在第 5.1 节中阐述。
基准测试
。我们在近期图像生成模型技术报告(cai2025z; qin2025lumina; cai2025hidream; cui2025emu3)中常用的五个代表性基准上评估我们的模型:GenEval(ghosh2023geneval)、DPG-Bench(hu2024ella)、PRISM-Bench(fang2026flux)、CVTG-2K(du2025textcrafter)和 LongText-Bench(geng2025x)。GenEval 专注于以物体为中心的图像生成,并评估一组固定的物体属性和关系。DPG-Bench 和 PRISM-Bench 提供对通用提示词遵循能力的细粒度评估,其中 PRISM-Bench 还额外评估图像美学。CVTG-2K 和 LongText-Bench 评估模型生成包含可检测文本且该文本与输入提示词描述相匹配的图像的能力。
我们注意到,先前已有研究指出 GenEval 可能与人类判断存在偏差(kamath2025geneval),并且与人类感知的模型能力相关性较差(cao2025hunyuanimage)。此外,当前模型(chen2025blip3; ma2026deco; wang2026pixnerd)中普遍存在在 BLIP3o-60K(chen2025blip3)上进行微调的做法,这可能会虚高 GenEval 分数,因为研究发现 BLIP3o-60K 微调能显著提升 GenEval 分数,但对其他基准测试并无帮助(wu2025openuni)。因此,我们报告 GenEval 结果仅出于完整性考虑,并指出这些结果可能无法准确反映模型能力。
模型 #参数量 GenEval DPG-Bench PRISM CVTG-2K LongText-Bench 仅 API 调用 GPT Image 1 [High] (gptimage1) - 0.84* 85.15* - 0.8569* 0.956* Seedream 3.0 (gao2025seedream) - 0.84* 88.27* - 0.5924* 0.896* 仅开放权重 FLUX.1 [Dev] (labs2025flux) 12B 0.66* 83.84* 65.1 0.4965* 0.607* SD3 Medium (esser2024scaling) 2B 0.62* 84.08* 61.9 0.4037 0.322 Janus-Pro-7B (chen2025janus) 7B 0.80* 84.19* 60.0 0.0667 0.019* BAGEL (deng2025emerging) 14B 0.88* 85.44 61.8 0.3642 0.373* HiDream-I1-Full (cai2025hidream) 17B 0.83* 85.89* 66.1 0.7738 0.543* Lumina-Image 2.0 (qin2025lumina) 3B 0.73* 87.20* 63.5 0.1577 0.088 Z-Image (cai2025z) 6B 0.84* 88.14* 74.2 0.8671* 0.935* Qwen-Image (wu2025qwen) 20B 0.87* 88.32* 73.9 0.8288* 0.943* 开放权重 + 数据 + 训练代码 BLIP3o-4B (chen2025blip3) 4B 0.77 79.73 53.2 0.0353 0.023 PixNerd (wang2026pixnerd) 1B 0.73* 80.9* 53.3 0.0006 0.020 DeCo (ma2026deco) 1B 0.86* 81.4* 53.1 0.0014 0.003 BLIP3o-N-S (chen2025blip3o) 3B 0.87 81.98 56.8 0.2493 0.110 BLIP3o-N-G-G (chen2025blip3o) 3B 0.90 81.93 57.5 0.2442 0.114 BLIP3o-N-G-T (chen2025blip3o) 3B 0.86 79.77 56.8 0.3330 0.153 i1 (Ours) 3B 0.84 86.73 70.1 0.8531 0.922
结果
。我们在表 8 中将 i1 模型与领先的图像生成系统进行了比较。除 GenEval 外,i1 在所有五个基准测试中均达到了完全开放模型中的最先进性能。它还超越了多个领先的仅权重模型,包括 Lumina-Image 2.0、HiDream-I1 和 FLUX.1 [Dev]。i1 的强劲性能反映了我们整个研究中确定的建模和数据选择的综合效果。
第 7 节 讨论与结论
完全开放的配方支持文生图建模中的累积性研究。
当前文生图研究中的一个挑战是,强大的模型通常作为不透明的端点发布,而非可作为科学制品进行检验。因此,进展往往难以归因于各种(可能未公开的)设计因素。我们的研究倡导完全开放的配方,旨在理解哪些设计选择确实重要。通过发布 i1 背后的模型、代码、数据配方和消融实验,我们不仅希望提供一个强大的基线,也希望为更具累积性和可复现性的研究提供一个参考点。
强大的性能并不需要复杂的设计。
近期文生图模型的强劲性能可能会造成一种印象,即前沿能力需要日益专业化的架构、专有数据或高度工程化的配方。我们的研究提供了一个反例:使用规模适中(例如 440 万,见第 5.2 节)且公开可用的数据集,并结合对当前建模设计空间的仔细探索,即可实现强劲性能。我们认为这对开放研究来说是令人鼓舞的,因为具有竞争力的文生图模型无需从难以获取的数据或未公开的训练流程起步。
局限性与未来工作
本工作存在若干局限性。首先,我们的评估主要依赖自动化基准测试,这些测试侧重于提示词遵循能力,而非人类偏好。因此,尽管 i1 在这些基准测试上接近领先的纯权重模型(例如 Qwen-Image),但其生成的图像在整体视觉质量上仍明显逊色(我们在附录 B.5 中展示了失败案例)。其次,受资源限制,所有实验均使用约 3B 参数或更小的模型进行。需要进一步实验来确定我们的发现在更大规模下是否仍然成立。第三,我们的探索仅覆盖了文生图扩散模型设计空间的一个子集:诸如多宽高比训练(我们正在开发多宽高比模型,并将很快发布)、数据过滤(startsev2026alchemist)、解码器专用 LLM 与扩散 Transformer 在文本编码方面的深度融合(liu2024playground; shi2026lmfusion)以及强化学习(wallace2024diffusion; liu2026flow)等设计均未涉及。未来的工作可以将我们的方案扩展到更大的模型,并在保持整体流程简洁性和开放性的同时,进一步探索设计空间。
致谢
我们衷心感谢 Google TPU 研究云(TRC)项目为本项目提供主要计算资源。普林斯顿大学的普林斯顿研究计算资源提供了额外支持,这些资源由普林斯顿计算科学与工程研究所(PICSciE)和研究计算部门领导的联合小组管理。我们要感谢 Liang-Chieh Chen、Ishan Misra、Kaiming He、Yida Yin、Haozhe Chen、Wenhao Chai、Linrong Cai、Linzhan Mou 和 Xingyu Fu 提供的宝贵讨论和反馈。我们还感谢 Yufeng Xu、Shengbang Tong、Yiyang Lu 和 Hanhong Zhao 在 TPU 方面的有益讨论。我们感谢 Cihang Xie 的研究小组分享他们的 JAX DiT 代码库,该代码库是我们研究的起点。
参考文献
附录
附录 A 实现细节
在本节中,我们提供关于建模和训练配置的更多细节。
A.1 配置
硬件
我们的模型训练和推理在 TPU v4、v5p 和 v6e 上使用 JAX(jax2018github)进行。基准评估则在 NVIDIA A100、H100 和 H200 GPU 上执行。
通用配置
我们的默认基线模型主要沿用此前扩散模型(peebles2023scalable; yao2025reconstruction)中使用的 XL/2 模型配置,即隐藏层大小为 1152、16 个注意力头、MLP 比率为 4.0、补丁大小为 2。不过,与这些模型不同的是,我们使用了 29 层而非 28 层。默认情况下,在训练和推理过程中,我们将文本编码器保持为 bf16,同时将所有其他参数保持为 fp32。为确保模型能够装入内存,我们使用 JAX pjit/GSPMD(xu2021gspmd)跨设备对模型参数和优化器状态进行分片,遵循 ZeRO 风格的全分片数据并行(rajbhandari2020zero)。
| 配置 | 值 |
| 优化器 | Adam |
| 学习率 | 1e-4 |
| 权重衰减 | 0 |
| 优化器动量 | |
| 批量大小 | 512 |
| 学习率调度 | 常数 |
| 梯度裁剪 | 1 |
| 训练目标 | 流匹配 |
| 训练步数 | 500K |
| 训练时间步分布 | lognorm(0, 1) |
| 推理时间步偏移值(esser2024scaling) | 0.3 |
| 推理步数 | 250 |
| CFG 缩放(ho2022classifier) | 12 |
| CFG 重新缩放强度(lin2024common) | 0 |
| CFG 区间(kynkaanniemi2024applying) | [0, 1] |
256 分辨率受控实验
表 9 总结了第 4 节和第 5 节中所有受控实验的训练配置。所有模型均训练 500K 次迭代,但由于不同实验使用了不同的模型组件,训练时间有所不同。对于交叉注意力基线,在 TPU v6e-64 机器上,500K 步需要 31.0 小时。
| 训练阶段 | 图像数量 | 训练步数 | 批量大小 | 训练时间步偏移值(esser2024scaling) | TPU v5p-128 小时 |
| 256 分辨率 | 162.9M | 2.0M | 512 | 不适用 | 383.0 |
| 512 分辨率 | 9.7M | 0.5M | 512 | 不适用 | 174.4 |
| 1024 分辨率 | 4.3M | 0.3M | 128 | 3.33 | 150.9 |
i1 训练
在表10中,我们详细列出了最终i1模型在每个训练阶段的训练配置和计算资源。所有未指定的配置均与表9保持一致。高分辨率训练阶段各迭代轮次的基准性能趋势如图26所示。我们观察到,512分辨率训练显著提升了PRISM和LongText上的性能,而1024分辨率训练的效果较小,其性能与初始化所用的512分辨率检查点相近。对于1024分辨率训练,我们额外比较了使用时间步偏移值3.33训练的模型与未使用时间步偏移训练的模型,发现应用训练时间步偏移能持续提升性能。

(a) 512分辨率训练

(b) 1024分辨率训练
A.2 基线架构
如第3节所述,我们在第4节和第5节中的控制实验均基于固定的基线架构。我们每次只改变一个设计选择,同时保持所有其他配置与基线相同。虽然我们在基线中默认使用交叉注意力主干,但我们也在单流和双流主干上额外验证了一些设计选择。在本节中,我们提供三种主干架构的图示。
交叉注意力主干
通过插入在自注意力和前馈网络层之间的交叉注意力层,将文本条件信息传递给主干。该架构如图27所示。
单流主干
将文本特征和带噪图像特征沿序列维度拼接,并使用一组主干权重处理整个序列。该架构如图 28 所示。
双流主干
将文本特征和带噪图像特征沿序列维度拼接,但对文本 token 和图像 token 使用不同的主干参数。该架构如图 29 所示。
A.3 文本编码器详情
模型版本。对于所有 T5Gemma 模型,我们使用 UL2(tay2023ul)变体,因为它具有更好的编码器表示(zhang2025encoder)。对于 T5Gemma-9B 模型,我们使用带有 2B 解码器的变体,而非带有 9B 解码器的变体。对于 FG-CLIP 2 模型,我们使用“长”模式。
截断。遵循主流实现(esser2024scaling; blackforestlabs_flux2_2025; cai2025z; wu2025qwen),我们对文本分词器采用右侧截断。除 FG-CLIP 2 外,所有文本编码器均截断至 256 个 token;FG-CLIP 2 截断至 196 个 token,因为其训练数据最多包含 196 个 token。
隐藏状态。
. Following mainstream implementations (esser2024scaling; blackforestlabs_flux2_2025; cai2025z; wu2025qwen), we use right truncation for the text tokenizers. We truncate to 256 tokens for all text encoders except FG-CLIP 2, which we truncate to 196 tokens because it is trained on up to 196 tokens.
Hidden states
对于编码器-解码器模型(即 T5Gemma 和 T5Gemma2 系列),我们使用编码器最后一层的隐藏状态作为文本 token 特征。对于仅解码器模型(即 Qwen3 和 Qwen3-VL 系列),我们使用最后一个 Transformer 层的隐藏状态作为文本 token 特征。默认情况下,我们直接将文生图提示词输入文本编码器以获取特征。此前的一些工作(ma2024exploring; xie2025sana; wu2025qwen)对 LLM/VLM 文本编码器应用了系统提示词;我们在附录 C.4 中消融分析了系统提示词的影响。
附录 B 推理与评估补充信息
B.1 与 Stable Diffusion 3 Medium 的定性对比
在图 2 和图 3 中,我们展示了由 i1 模型生成的精选示例图像。在图 30 和图 31 中,我们额外提供了四个精心挑选的模型生成示例,并将其与使用相同提示词由 Stable Diffusion 3 Medium 生成的图像进行了对比。
提示词:美国女演员维罗妮卡·莱克(1922-1973)坐在扶手椅上,身穿白色围裙裙搭配红色衬衫,正全神贯注地阅读一本书,营造出约 1955 年的场景氛围。
提示词:一间明亮、温馨的面包店内部,在温暖柔和的晨光中拍摄,展示了一块诱人的、质朴木框黑板菜单,醒目地放置在白砖墙前。黑板中央,用优雅的手绘大号字体清晰写着“今日特供:酸面包与肉桂卷”。在这条中央信息正下方,较小的文字工整地标注着“每日清晨新鲜烘焙”。黑板右上角,以俏皮的草书字体隐约写着“用心手工制作”。菜单边框周围,略带褪色、复古风格的小麦穗和糕点插画,巧妙地勾勒出文字,增强了手工面包店的氛围。主黑板旁边立着一块较小的木牌,上面手写着“免费试吃供应!”,并配有一个装饰性箭头,引导顾客前往展示柜台。文字元素清晰、别致,且为自然手写风格,唤起一种真实的手工感,完美衬托了这家面包店诱人的氛围。
提示词:一幅由 David Forks 创作的名为“午夜之月”的宁静油画,捕捉了在明亮满月下,一个孤独的身影立于岩石海岸的场景,以戏剧性的光影和深蓝色调渲染,唤起一种沉思且富有氛围的情绪。
提示词:一幅当代艺术风格电影海报,采用极简且富有冲击力的文字排版。顶部居中位置,以粗体、流畅字体书写的标题“最后的航程”置于一艘古老帆船迎向汹涌波涛的微妙剪影之上。剪影正下方,以稍小字体呈现一句引人入胜的宣传语:“当勇气意味着驶向未知”。海报下半部分中央,以清晰的水平排列方式,整齐地列出了关键信息:“由获奖导演亚历克斯·里弗斯执导”、“艾米莉·克拉克与雅各布·班尼特联袂主演”、“丹尼尔·哈珀原创配乐”。在最底部,以简洁的大写字母清晰分隔,上映信息写道:“2023年10月6日 全球影院上映”。左下角有一行较小的斜体细线文字:“你敢于踏上这段旅程吗?”
B.2 在不同推理设置下评估其他模型
i1的推理设置(见第6.3节)使用了12的CFG尺度,这高于许多现有文生图扩散模型的默认值。例如,PixArt-(chen2024pixart)使用的默认CFG尺度为4.5,Lumina-Image 2.0(qin2025lumina)使用4,SANA(xie2025sana)使用4.5,Stable Diffusion 3(esser2024scaling)使用7。此外,我们使用了一个自定义的元提示词用于推理时的提示词重写。这些选择可能引发疑问:我们的推理设置是否使我们的方法相比基线模型具有不公平的优势。为了检验这一点,我们在替代推理设置下评估了Lumina-Image 2.0和Stable Diffusion 3 Medium,包括更大的CFG尺度以及使用我们的元提示词重写后的提示词(见附录B.3)。结果如表11所示。我们发现,无论是增大CFG尺度还是使用重写后的提示词,都未能显著提升这两个模型的性能。
| 提示词 | CFG尺度 | DPG | PRISM | LongText |
| 原始 | 4 | 87.20 | 63.5 | 0.088 |
| 8 | 87.39 | 60.7 | 0.100 | |
| 12 | 87.56 | 60.9 | 0.101 | |
| 16 | 87.84 | 58.6 | 0.107 | |
| 重写后 | 4 | 85.37 | 63.1 | 0.092 |
(a) Lumina-Image 2.0
| 提示词 | CFG尺度 | DPG | PRISM | LongText |
| 原始 | 7 | 84.08 | 61.9 | 0.322 |
| 8 | 85.49 | 61.2 | 0.341 | |
| 12 | 84.94 | 56.2 | 0.361 | |
| 16 | 82.82 | 50.1 | 0.368 | |
| 重写后 | 7 | 84.43 | 61.0 | 0.313 |
(b) Stable Diffusion 3 Medium
B.3 用于提示词重写的元提示词
在第 5.1 节中,我们发现,在短描述上训练会导致整体模型较弱,而在长描述上训练则能产生更强的模型,但会导致在短提示词上表现不佳。提示词重写可以通过扩展短推理提示词来缓解这种训练-推理提示词长度不匹配的问题,使得即使在原始推理提示词较短的情况下,在长描述上训练也优于在短描述上训练。对于表 5 中的实验,我们使用了一个简单、极简的元提示词,将较短的 GenEval 提示词扩展为更长的提示词。
然而,始终指示提示词重写 LLM 扩展输入提示词可能并非最优方案,因为推理提示词的长度是可变的,甚至可能比训练描述还要长。因此,我们设计了一个更全面的元提示词,指示模型根据输入提示词的复杂程度遵循两套不同的指导原则。在元提示词的末尾,我们还额外包含了 20 对人工编写的原始提示词与重写提示词作为上下文示例,以引导 LLM。
包含这 20 个上下文示例的元提示词如下所示。
B.4 推理步数的消融实验
在第 4 节和第 5 节的受控实验以及表 8 中对 i1 的评估中,我们将推理步数固定为 250。然而,我们注意到,要实现强大的模型性能,250 步并非必需。在图 32 中,我们使用 5、10、20 和 50 个推理步数评估了 i1 模型的性能,并在图 33 中展示了一个定性示例。我们观察到,可以将推理步数减少至低至 20 步,而不会显著损害生成质量。
| 10 | 20 | 250 |
![]() | ![]() | ![]() |
B.5 i1 的失败案例
尽管 i1 性能强劲,但我们注意到该模型仍存在若干重要失败案例,如图 34 所示。如图 34(a) 所示,特别是在需要生成群体场景中的多个小人像时,i1 有时会生成保真度较低的人脸、不自然的面部表情,以及畸形的手部或肢体。此外,i1 并不总能遵循物理属性,有时会生成物理上不合理的图像。图 34(b) 提供了这样一个例子:i1 未能捕捉镜子的物理特性,因为反射图像显示镜子与车窗平行,这在物理上是不一致的。
B.6 各基准测试的提示词长度分布
在图 35 中,我们可视化了 i1 最终评估(见表 8)所用基准测试的提示词长度分布。我们观察到,GenEval 的提示词比其他基准测试短得多。这促使我们在分析仅使用长描述训练的模型在短提示词上表现不佳时,重点关注 GenEval,以及相应的缓解技术(见第 5.1 节)。
附录 C 关于建模设计的额外结果
在第 4 节中,我们通过控制实验得出了关于建模设计的几个结论。在此,我们在附录 C.1 和 C.2 中提供额外的实验结果,这些结果支撑了最终 i1 模型中使用的位置编码、归一化和 VAE。在其余小节中,我们提供额外的结果和分析,以验证我们在第 4 节中得出的结论在替代设置下依然成立。
C.1 位置编码与归一化
在我们的最终 i1 模型中,我们同时使用了正弦位置编码和 RoPE 位置编码,采用了三明治归一化,并在 MMDiT 的文本和图像流之间共享归一化层。我们在下面描述了促使这些设计选择的实验。
位置编码
当前的文生图扩散模型通常只使用一种位置嵌入(例如正弦位置嵌入(esser2024scaling)或旋转位置编码(cai2025z; wu2025qwen; qin2025lumina))。受 LightningDiT(yao2025reconstruction)设计的启发,我们探索了将正弦位置嵌入和旋转位置编码相结合是否能提升扩散 Transformer 的文生图性能。如图 36 所示,将两者结合能显著提升交叉注意力模型和双流模型的基准测试性能。因此,我们在最终的 i1 模型中同时使用了正弦位置嵌入和旋转位置编码。
归一化
尽管前归一化(xiong2020layer)(即对注意力模块和前馈网络模块的输入进行归一化)一直是扩散 Transformer 的主流选择,但早期工作(ding2021cogview)引入了三明治归一化(即同时对注意力模块和前馈网络模块的输入和输出进行归一化)以稳定训练,该方法近期已被 Z-Image(cai2025z)采用。我们将三明治归一化引入基线模型,如图 37 所示,并发现它能在不同骨干架构和基准测试中稳定地提升性能。
| 模型 | DPG | PRISM | LongText |
| 共享归一化 | 86.82 | 58.3 | 0.439 |
| 独立归一化 | 86.16 | 57.6 | 0.415 |
此外,虽然为文本和图像模态使用独立的归一化层是现有 MMDiT 模型(esser2024scaling; cai2025hidream; wu2025qwen)的默认做法,但我们探索了跨模态共享归一化层所带来的更统一的特征分布是否有利于模型性能。如表 12 所示,共享归一化确实能持续提升性能。
C.2 变分自编码器
| VAE | DPG | PRISM | LongText |
| FLUX.2 | 84.66 | 56.4 | 0.211 |
| Qwen-Image | 83.29 | 54.3 | 0.266 |
| VA-VAE | 85.07 | 55.5 | 0.126 |
我们将 VA-VAE(yao2025reconstruction)与前沿模型 FLUX.2(blackforestlabs_flux2_2025)和 Qwen-Image(wu2025qwen)中使用的 VAE 进行了比较。总体而言,FLUX.2 在所有基准测试中取得了最均衡的性能。可能由于其在训练过程中与预训练语义特征的对齐,VA-VAE 在 DPG-Bench 上取得了最强性能,该基准侧重于生成图像与提示词之间的语义对齐。然而,在评估细粒度文本渲染的 LongText 基准上,它的表现远逊于其他模型。这可能是由于其较低的重建保真度所致。
| 原始图像 | FLUX.2 VAE | Qwen-Image VAE | VA-VAE |
![]() | ![]() | ![]() | ![]() |
![]() | ![]() | ![]() | ![]() |
从量化角度来看,先前的工作(wu2025qwen; yao2025reconstruction)报告称,在 256×256 分辨率的 ImageNet 验证集上,VA-VAE 的重建性能(PSNR 27.96,SSIM 0.79)低于 FLUX.2 VAE(31.46,0.90)和 Qwen-Image VAE(33.42,0.92)。在图 38 中,我们进一步提供了在富含文本图像上的定性重建示例。虽然 FLUX.2 VAE 和 Qwen-Image VAE 能够忠实地重建,但 VA-VAE 在渲染的字符中引入了可见的损坏和失真。造成这一局限的一个可能原因是,VA-VAE 是在缺乏富含文本图像的 ImageNet(deng2009imagenet)上训练的。
C.3 在替代设置下比较文本编码器
更大的适配器
在 4.2 节中,我们在默认基线模型上使用小型 MLP 适配器对文本编码器候选方案进行了比较。然而,正如我们在 4.1 节中后续展示的,适配器的大小对模型性能有显著影响。因此,在此我们额外探究:当我们使用由两个 Transformer 模块组成的更大文本编码器适配器(与最终 i1 方案一致)时,文本编码器之间的比较结果是否仍然成立。如图 39 所示,我们的观察结果基本保持不变:T5Gemma 和 T5Gemma2 系列的编码器-解码器模型表现最强,而 FG-CLIP 2 表现最弱。
移除 AdaLN
在我们的基线设置(第 3 节)中,池化后的文本嵌入与时间步嵌入相结合,并通过 AdaLN 传入主干网络。由于这为注入文本信息提供了额外路径,因此在最终 i1 方案中移除 AdaLN 可能会影响不同文本编码器的相对性能。我们在图 40 中展示了每个文本编码器的基准测试结果。我们观察到,整体趋势与使用 AdaLN 的默认设置(图 8)中的趋势高度相似。
C.4 对文本编码器应用系统提示词
在默认设置(见附录 A.3)中,我们直接使用每个文本编码器处理原始提示词,并将最后的隐藏状态作为文本特征。虽然这种设置很常见(cai2025z),但先前的工作也设计了专门的提示词策略,用于将仅解码器语言模型用作文本编码器(ma2024exploring; xie2025sana)。在此,我们遵循 Qwen-Image(wu2025qwen)为 Qwen2.5-VL 文本编码器使用的策略,并将其应用于 Qwen3-VL-2B 和 Qwen3-VL-4B 文本编码器。
具体来说,我们将文生图提示词包裹在一条系统消息中(“通过详细描述物体和背景的颜色、形状、大小、纹理、数量、文字以及空间关系来描述图像:”),并将完整序列输入文本编码器。在前向传播之后,我们丢弃与系统前缀对应的隐藏状态,只保留与文生图提示词对应的部分。由此得到的性能如表 14 所示。我们观察到,使用系统提示词为 Qwen3-VL-2B 带来了小幅提升,但 Qwen3-VL 模型的性能仍不及 T5Gemma 和 T5Gemma2 模型。
| 文本编码器 | 系统提示词 | DPG | PRISM | LongText |
| Qwen3-VL-2B | ✗ | 82.29 | 52.2 | 0.076 |
| ✓ | 82.82 | 52.8 | 0.093 | |
| Qwen3-VL-4B | ✗ | 82.07 | 52.9 | 0.071 |
| ✓ | 82.41 | 52.4 | 0.065 |
C.5 基于训练 FLOPs 比较骨干网络家族
在图 12 中,我们通过训练宽度为 1152、1296、1440、1584 和 1728 的交叉注意力、单流和双流模型,并绘制性能与模型大小的关系图,比较了不同的骨干网络家族。然而,根据实际的训练和推理设置,基于 FLOPs 比较性能可能更具参考价值。因此,在图 41 中,我们绘制了性能与可训练模型 FLOPs 的关系图。这些 FLOPs 是使用 JAX/XLA 的“cost_analysis”计算的,针对扩散模型中所有可训练模块(例如包括文本编码器适配器)的一次前向和反向传播。我们使用训练张量形状,并将结果按全局批次大小和训练步数进行缩放。双流骨干网络仍然实现了最佳的权衡。
C.6 在更大模型上验证建模设计
性能与模型大小的关系
在第4节和附录C.1中,除了我们在不同模型规模上验证过的骨干网络族对比和长跳跃连接外,我们主要使用XL/2尺寸的基线模型来识别建模设计发现。在图43中,我们通过在{1152, 1296, 1440, 1584, 1728}中每个宽度训练一个模型,进一步在多个模型规模的双流MMDiT模型上验证了这些发现。
在不同模型规模下,更大的文本编码器适配器(第4.1节)始终能提供更好的性能-参数权衡。当使用MLP文本编码器适配器时,移除AdaLN(第4.1节)也带来了明显的优势。然而,当使用更大的Transformer文本编码器适配器时,带和不带AdaLN的模型在可比较的参数数量下实现了相似的性能。我们注意到,这仍然表明在文生图扩散模型中,噪声条件化可能并非必要。
最后,虽然附录C.1提供了初步结果,表明结合正弦位置编码和RoPE位置编码、使用三明治归一化以及跨图像和文本流共享归一化可以提升性能,但我们并未在不同模型规模上一致地观察到这些趋势。
性能与训练FLOPs对比
除了在图11和图43中比较不同模型规模的性能外,我们在图44中将性能与可训练模块的估计训练FLOPs进行了对比。我们按照附录C.5中的相同流程计算这些FLOPs。
在基于FLOPs的对比下,大多数趋势保持不变。一个例外是AdaLN在使用更大的基于Transformer的文本编码器适配器时的效果:在这种设置下,带有AdaLN的模型比不带AdaLN的模型具有更好的性能-FLOPs权衡。这是因为AdaLN贡献了模型参数中不可忽视的一部分(例如,在双流基线模型中占参数的18.9%),但仅极小程度地增加了训练FLOPs,因为其投影是对每个样本计算一次,而非对每个token计算一次。
C.7 在其他骨干网络上验证长跳跃连接结果
在 4.2 节中,我们基于双流 MMDiT 主干网络的实验表明,长跳跃连接能够持续提升不同规模模型的性能。在此,我们通过训练基线模型(XL/2 尺寸)的其他变体(分别包含和不包含长跳跃连接),进一步在其他主干网络上验证了这一设计。如图 42 所示,移除长跳跃连接会显著降低大多数主干网络和评测基准上的性能,尤其是在 DPG 和 LongText 上。这进一步表明,长跳跃连接可以广泛地提升模型性能。
C.8 文本特征适配器与图像特征适配器
在 4.1 节中,我们发现,尽管只增加了少量参数,但将文本编码器的小型 MLP 适配器替换为更大的 Transformer 适配器,能显著提升各评测基准的性能。我们推测,这是因为来自预训练语言模型的文本特征需要针对文本到图像生成等下游任务进行适配。在附录 C.6 中,我们展示了更大的 Transformer 适配器比小型 MLP 适配器实现了更好的性能-参数权衡。
| 主干网络 | 适配器 | DPG | PRISM | LongText |
| 交叉注意力 | 默认 | 84.66 | 56.4 | 0.211 |
| + 文本特征的 Transformer 适配器 | 86.33 | 58.7 | 0.414 | |
| + 图像特征的 Transformer 适配器 | 85.27 | 55.4 | 0.250 | |
| 单流 | 默认 | 85.89 | 55.6 | 0.293 |
| + 文本特征的 Transformer 适配器 | 87.64 | 60.0 | 0.472 | |
| + 图像特征的 Transformer 适配器 | 86.10 | 59.7 | 0.345 | |
| 双流 | 默认 | 86.82 | 58.3 | 0.439 |
| + 文本特征的 Transformer 适配器 | 87.67 | 60.7 | 0.576 | |
| + 图像特征的 Transformer 适配器 | 86.23 | 57.0 | 0.378 |
为了进一步验证这种性能提升并非仅仅源于参数数量的增加,我们在图像特征上同样添加了一个 Transformer 模块,具体位置是在图像分块之后、应用位置嵌入之前(详见附录 A.2)。如表 15 所示,与在文本特征上使用该适配器相比,在图像特征上添加此适配器所带来的性能提升要小得多。这表明,更大适配器带来的增益是专门针对适配预训练文本特征的,而不仅仅是简单地增加模型容量。
C.9 探索长跳跃连接的变体
长跳跃连接由 U-Net(ronneberger2015u)推广开来,其作用是为来自早期层的低级特征提供快捷路径,从而简化像素级预测任务的训练。后来,U-ViT(bao2023all)沿用了这一设计,并将其应用于基于 Transformer 架构的扩散模型。然而,尽管这些连接完全对称的结构(即,从左数第 i 层连接到从右数第 i 层)对于 U-Net 的多分辨率编码器-解码器结构来说是自然而然的,但对于基于 Transformer 的模型来说却未必是最优的,因为后者的各模块通常具有相同的特征维度。因此,我们在此探索了 U-ViT 中原始长跳跃连接的不同变体。
层范围
来自较早层的长跳跃连接会跨越更多模块,而靠近中间层的跳跃连接则跨越较少模块。因此,靠近中间层的特征在到达目标层时可能变化较小,这使得这些跳跃连接可能不那么必要。移除它们或许能保持甚至提升性能。为验证这一假设,我们在图45中探索了长跳跃连接可以起始的若干层范围:1-3、1-7、1-11、4-7、4-11、4-14、8-11、8-14 和 12-14。然而,这些变体均未持续优于默认设置。
| 跳跃类型 | DPG | PRISM | LongText |
| 默认 | 84.66 | 56.4 | 0.211 |
| 83.70 | 55.2 | 0.134 | |
| 84.78 | 55.8 | 0.180 |
连接模式
我们进一步探索了长跳跃连接是否应连接表征差异更大的层。我们没有使用默认的对称模式(即将最左侧的第 层连接到最右侧的第 层),而是测试了将第 层连接到第 层或第 层的变体。如表16所示,在交叉注意力主干上,该变体略微提升了DPG,但默认模式整体表现仍然最佳,取得了最高的PRISM和LongText分数。
附录D 数据设计方面的额外结果
在第5节中,我们展示了控制实验,以论证我们在合成字幕生成、提示词改写和数据混合方面的设计动机。我们在附录D.1中提供了关于合成字幕生成设计的更多细节及相应的控制实验,并在附录D.2中提供了支持我们关于数据混合结论的额外结果。
D.1 合成字幕生成中的额外设计
在 5.1 节中,我们通过在 ImageNet-22K 图像上使用不同描述集训练一个基线交叉注意力模型,研究了合成描述生成。在此,我们提供额外的结果,这些结果促使我们在描述生成流程中做出两个选择:在生成描述前对图像进行中心裁剪,以及为每张图像生成多条描述。
| 描述生成器 | DPG | PRISM | LongText |
| Qwen3-VL-30B-A3B | 83.72 | 50.8 | 0.007 |
| + 无中心裁剪 | 83.16 | 51.3 | 0.006 |
| + 5 条描述/图像 | 83.56 | 51.9 | 0.010 |
图像裁剪
。在我们的实验设置中,我们通过将较长边中心裁剪至与较短边等长,从而在正方形图像上进行训练。如果合成描述生成器接收的是未经裁剪的完整图像,生成的描述可能会描述那些后续被裁剪掉的对象,从而在训练图像和描述之间产生语义不匹配。podell2023sdxl 指出,这可能导致文生图模型生成不完整对象的失败模式。此外,默认情况下,在中心裁剪之后,我们会将尺寸大于 512x512 的图像缩小至 512x512,以避免在较大图像上生成描述时速度过慢。
为了理解我们预处理操作的影响,我们使用由 Qwen3-VL-30B-A3B 生成的两组合成描述来训练模型:一组来自完整图像,另一组来自经过裁剪和缩放的方形图像。如表 17 的前两行所示,由裁剪和缩放后的图像生成的描述,其下游性能与由完整图像生成的描述相似。由于中心裁剪和缩放提高了描述生成速度,且未对下游性能产生显著影响,我们在生成所有合成描述之前都应用了这两种操作。
增加每张图像的描述数量,提供了除增加图像数量之外的另一个数据扩展维度。为探索这一点,我们使用 Qwen3-VL-30B-A3B 为每张图像生成了五条描述。如表 17 所示,这在 PRISM 和 LongText 上带来了小幅改进。此外,如图 20 所示,当独特训练图像数量有限时,这种收益变得更加明显。
D.2 等数据集权重下的实验
在第 5.2 节中,在确定所有数据集等权重可以带来强劲性能后,我们进一步探索了在保持其他数据集等权重的同时,对单个数据集进行上采样是否能进一步提升结果。在图 19 中,我们展示了任意一个数据集的上采样并不能在所有基准测试上持续提升性能。在此,我们进一步在图 46 中展示了每个数据集上采样 5 倍的结果,并发现这些结果与图 19 中的观察结果一致。
附录 E 数据集的补充信息
我们在第 3 节简要介绍了用于模型训练的图像数据集,并在第 5 节探讨了合成描述和数据混合策略。在此,我们提供关于训练期间所用图像和描述的更多细节。
E.1 图像可视化
在我们的研究中,我们探索了数据混合策略,并在 12 个公开可用的图像数据集上训练了模型,其中包括 7 个真实图像数据集(ImageNet-22K、YFCC100M、RedCaps、Megalith、Pexels、iNaturalist 2024、Places365-Challenge 2016)、3 个合成数据集(GPT-Image-Edit-1.5M、FLUX-Reason-6M 和 Midjourney v6)以及 2 个文本渲染数据集(RenderedText 和 TextAtlas)。为了定性地了解它们的分布情况,我们从每个数据集中随机抽取 9 张图像,将每张图像调整大小使其短边为 256 像素,中心裁剪为 256x256 的正方形,并将得到的图像可视化在图 47 中。
E.2 图像分辨率统计
如第 6.2 节所述,我们使用除 iNaturalist 之外的所有图像数据集进行 i1 的 256 分辨率训练。对于 512/1024 分辨率的高分辨率训练,我们过滤掉短边小于 512/1024 像素的图像。我们还移除了过滤后图像数量少于 30 万张的任何数据集。
在此,我们在表 18 中报告了在每个分辨率阈值下剩余的图像数量。对于 512 分辨率训练,我们使用除 YFCC 和 iNaturalist 之外的所有数据集。对于 1024 分辨率训练,我们仅使用 FLUX-Reason、TextAtlas、RedCaps、GPT-Edit 和 Midjourney v6(尽管 RenderedText 在基于分辨率过滤后满足图像数量要求,但由于其质量较低,我们将其排除在外)。
我们使用的 RedCaps 数据集包含 500 万张图像,这与原始论文中报告的 1200 万张图像不同。这是因为 RedCaps 图像以 URL 形式提供,其中许多 URL 已无法访问。
| 图像类型 | 数据集 | 图像数量 | 短边 ≥ 512 的图像数量 | 短边 ≥ 1024 的图像数量 |
| 真实 | YFCC | 98,121,424 | 0 | 0 |
| ImageNet-22K | 13,673,551 | 719,427 | 0 | |
| Megalith | 9,393,971 | 9,202,294 | 61,701 | |
| Places | 8,026,628 | 7,345,764 | 0 | |
| RedCaps | 4,817,431 | 4,705,536 | 4,134,303 | |
| iNaturalist | 4,813,543 | 0 | 0 | |
| Pexels | 2,810,634 | 2,810,621 | 0 | |
| 合成 | FLUX-Reason | 5,890,279 | 5,890,279 | 5,890,279 |
| GPT-Edit | 1,553,575 | 1,552,821 | 357,482 | |
| Midjourney v6 | 1,240,185 | 1,240,185 | 1,240,185 | |
| 文本渲染 | RenderedText | 11,977,824 | 11,977,824 | 11,977,824 |
| TextAtlas | 5,397,762 | 4,036,973 | 919,913 |
E.3 不同 VLM 描述器生成的描述文本长度分布
在第 5.1 节中,我们使用相同的元提示词来提示 Qwen2-VL 2B (wang2024qwen2)、Qwen2.5-VL 3B (qwen2.5-vl)、Qwen3-VL-2B、Qwen3-VL-4B 和 Qwen3-VL-30B-A3B (bai2025qwen3),为 ImageNet-22K 生成合成描述文本。然后,我们在每个生成的图像-描述文本数据集上训练一个扩散模型。我们发现,用于合成描述文本的 VLM 选择对下游文生图性能有显著影响。在图 48 中,我们进一步绘制了每个 VLM 生成的描述文本的序列长度分布。我们观察到,能带来更强性能的描述文本集通常也更长。
E.4 不同数据集的描述文本长度分布
在此,我们从每个数据集中随机抽取 10K 张图像,并在图 49 中绘制其合成描述文本的序列长度分布。尽管不同数据集的描述文本长度分布有所差异,但没有哪个数据集与其他数据集存在显著不同。
E.5 用于合成描述文本的元提示词
对于大多数图像数据集,我们使用以下最简提示词进行合成描述文本生成:
然而,对于文本渲染数据集,可以获取到真实文本标注。为了减少 VLM 生成描述文本中的模型幻觉,我们将真实文本包含在描述文本生成的提示词中。对于 TextAtlas,我们使用:
RenderedText 包含使用 Blender 在数字 3D 纸张上渲染的手写文本图像。文本的字体大小、颜色和旋转角度各不相同,纸张在随机光照条件下进行渲染。每个真实标注都包含一个字段“text”,它是一个字符串列表,其中 是图像中的文本行数。列表中的第 个元素对应第 行文本。因此,对于 RenderedText,我们使用:
E.6 数据泄露的完整性检查
除 FLUX-Reason 外,所有数据集的描述文本均由视觉语言模型(VLM)合成生成,因此我们的评估基准中使用的提示词集与用于训练文生图模型的描述文本之间不太可能存在重叠。然而,对于 FLUX-Reason,我们使用了五组描述文本:其中一组是数据集中提供的原始“caption_detail”字段(因其质量已足够高),其余四组由我们合成生成。根据原始论文(fang2026flux),PRISM-Bench 中一半的提示词直接选自 FLUX-Reason,随后已从数据集中移除。为确保 FLUX-Reason 与 PRISM-Bench 之间不存在重叠,我们对每个 FLUX-Reason 的“caption_detail”描述文本与每个 PRISM-Bench 提示词进行了精确字符串匹配。未发现任何重叠。
-
Diffusion models have consistently driven progress in text-to-image generation. However, it is challenging to attribute recent progress to specific modeling and data choices: state-of-the-art open-weight models provide limited ablations, and do not disclose their training data and full training details. The research community needs fully open (weights, data, and code) models as a foundation for further research; yet existing fully open models still fall significantly short of leading models in performance. In this project, we conduct a systematic investigation of the modeling and data design choices in text-to-image diffusion training and inference with 300+ controlled experiments totaling 700K+ TPU v6e hours. Our experiments highlight several empirical findings (e.g., equal weighting is a strong default for mixing curated datasets) and simple design decisions (e.g., larger text encoder adapters improve performance with minimal added parameters) for training strong models. Guided by these insights, we train i1, a 3B-parameter text-to-image diffusion model using only publicly available datasets. i1 is competitive with leading models on five representative benchmarks (GenEval, DPG, PRISM, CVTG-2K, and LongText), and outperforms the best existing fully open model by 29.5 absolute percentage points on average. We provide the i1 checkpoints, training and inference code, and the data processing pipeline. Together, our findings and the i1 recipe establish a practical foundation for future open research in text-to-image diffusion models.
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
![]() | ![]() | ![]() |
Contents
Section 1 Introduction
Since early models like DALL-E 2 (ramesh2022hierarchical), Imagen (saharia2022photorealistic), and Stable Diffusion (rombach2022high), diffusion-based models have driven major advances in text-to-image generation (wu2025qwen; labs2025flux; cai2025z; gao2025seedream) due to their strong capability for generating photorealistic images with fine-grained details. However, despite the superior capabilities of today’s state-of-the-art models, it is often difficult to disentangle which modeling and data choices are truly driving performance. This lack of clarity stems from two factors.
First, leading models often do not release their training data and full training recipe (wu2025qwen; cai2025z; qin2025lumina; cai2025hidream), even when they publicly release model checkpoints. This limits reproducibility and hinders controlled analysis and follow-up work that builds on their designs. While fully open (weights, data, and code) models exist (chen2025blip3; chen2025blip3o; ma2026deco; wang2026pixnerd), they fall substantially short of leading models in performance.
Second, leading models often do not provide thorough ablations of their design choices (wu2025qwen; cai2025z; cai2025hidream; gao2025seedream; ryu2025flite; fang2026flux). In practice, many models bundle numerous architectural, training, and data decisions into a single recipe, making it difficult to attribute improvements to any specific factor. As a result, modern text-to-image diffusion models still lack consensus on many important design choices (e.g., using a single text encoder (qin2025lumina; wu2025qwen) vs. multiple text encoders (esser2024scaling; cai2025hidream)).
To obtain a better understanding of the impact of existing and new architectural and data designs in text-to-image diffusion models, we conduct a series of controlled experiments, primarily on the 256256 low-resolution pre-training stage. Starting from a simple baseline model (yao2025reconstruction), we first explore strategies for incorporating text conditioning from text encoders, as well as noise/timestep conditioning (sun2025noise). Then, we identify backbone architecture designs that lead to stronger performance (bao2023all; esser2024scaling). Finally, we compare design choices in the curation of high-quality image-caption datasets and inference-time prompt enhancements, along with strategies for mixing image datasets.
On the modeling side, we find that (1) using a single strong text encoder with a larger adapter can be more effective than combining multiple text encoders, (2) timestep/noise conditioning and Adaptive Layer Normalization (peebles2023scalable) provide little benefit for text-to-image in our setting, and (3) a dual-stream DiT (esser2024scaling) with long skip connections (bao2023all) is a strong backbone design.
On the data side, we find that (1) training on long captions yields stronger models than training on short captions, but causes them to underperform on short prompts, which can be mitigated by inference-time prompt rewrite, (2) the choice of synthetic captioner is important for downstream performance, (3) training the model on equal numbers of images from each dataset, counting repetitions (called “equal weighting across datasets” hereafter), is a strong default for mixing curated datasets, (4) with a diverse mix of datasets, repeating training data incurs only marginal performance degradation, and (5) broad high-resolution data coverage is not needed to obtain strong high-resolution generation capability from a low-resolution model.
To provide a strong baseline for future open research, we leverage the insights from the controlled experiments to train i1, a text-to-image diffusion model with 3B parameters, on publicly available datasets. At 1024-resolution, i1 achieves state-of-the-art performance among fully open models and outperforms several leading open-weight-only models with much larger parameter counts (e.g., 17B HiDream-I1 (cai2025hidream) and 12B FLUX.1 [Dev] (labs2025flux)) across a diverse set of representative benchmarks.
i1 shows that strong performance can be achieved using only moderately scaled, publicly available image datasets, and highlights the value of carefully exploring the design space: as Figure 4 shows, i1 introduces no significantly new network modules, but instead identifies existing yet underused designs from prior work (e.g., long skip connections) and introduces simple modifications to standard components (e.g., using a larger text encoder adapter). We provide model weights, code, datasets, and detailed recipes for model training and evaluation. Together, our findings and the i1 recipe establish a practical foundation for open text-to-image research, offering both a strong fully open baseline and design insights for building more capable models.
Section 2 Preliminaries
In this section, we provide the terminology and context needed to understand and motivate our controlled experiment setup (Section 3) and later modeling and data design experiments (Sections 4 and 5), with a focus on backbone architectures, text and noise conditioning mechanisms, and existing open training data recipes.
Text-to-image backbone architectures
. Despite alternative paradigms (chang2023muse; sun2024autoregressive; zhou2024transfusion), most leading text-to-image systems use diffusion transformers (DiTs) (peebles2023scalable) trained with flow matching (lipman2022flow). Depending on how text features are incorporated, recent diffusion models generally fall into three categories: cross-attention models (chen2024pixart; xie2025sana; ryu2025flite), single-stream models (qin2025lumina; cai2025z; chen2025dit), and dual-stream MMDiT models (esser2024scaling; wu2025qwen). Cross-attention models inject text embeddings via cross-attention layers, whereas single- and dual-stream models concatenate image and text token sequences. Dual-stream models use modality-specific attention and MLP parameters for image and text tokens, while single-stream models use shared attention and MLP parameters across modalities. Long skip connections are an architectural modification that adds shortcuts between early and later layers. They were explored in earlier work (bao2023all) but are not widely used in modern text-to-image models.
Text and noise conditioning mechanisms
. In recent models, input prompts are encoded by one (cai2025z; qin2025lumina; wu2025qwen) or more (esser2024scaling; cai2025hidream) text encoders. The resulting text features are often passed through a linear (labs2025flux; cai2025hidream; wu2025qwen; cai2025z) or MLP (xie2025sana) adapter that maps them to the hidden dimension of the diffusion model. Across backbone architectures, Adaptive Layer Normalization (AdaLN) (peebles2023scalable) is commonly used to inject timestep information. AdaLN learns a linear projection from timestep embeddings to scaling and shifting factors for attention and MLP inputs, and gating factors for their outputs. In some models, the timestep embedding is combined with a pooled text embedding through element-wise addition before being used for AdaLN conditioning (esser2024scaling; cai2025hidream; labs2025flux).
Open text-to-image data recipes
. Many leading models (wu2025qwen; cai2025z; qin2025lumina; cai2025hidream; labs2025flux) release their weights publicly but do not disclose their training data recipes. Aside from a few models (qin2025lumina; ryu2025flite), even the sources and scale of the training datasets remain undisclosed, limiting the open research community’s understanding of how to construct strong text-to-image training data. Fully open models (chen2025blip3; chen2025blip3o; sehwag2025stretching; ma2026deco; tong2026scaling; wang2026pixnerd) still generally underperform leading systems, and their datasets are often from a similar and limited set of sources (e.g., JourneyDB (sun2023journeydb), SA-1B (kirillov2023segment), and CC12M (changpinyo2021conceptual)). Moreover, fully open recipes scarcely explore data balancing techniques.
Section 3 A Baseline for Controlled Experiments
Expanding beyond the existing designs introduced in Section 2, we study the text-to-image diffusion model design space through controlled experiments at the 256-resolution pre-training stage in Sections 4 and 5. For each set of experiments, we start from the same strong baseline and independently vary a single design choice (i.e., modifications are not accumulated across experiments). Designs that improve performance are later combined in Section 6 to construct our final model, i1 (see Figure 21). We describe the baseline setup below, provide a high-level illustration in Figure 5, and plot detailed architectures in Appendix A.2.
Model
. As illustrated in Figure 5, our backbone architecture is based on LightningDiT-XL/2 (yao2025reconstruction). LightningDiT is a modern DiT architecture (peebles2023scalable) that incorporates common designs for improving performance (e.g., RoPE (su2024roformer), RMS Norm (zhang2019root), SwiGLU FFN (shazeer2020glu)). We add QK-norm (dehghani2023scaling) to stabilize training. To ensure later ablations compare against a strong baseline, we also apply long skip connections (bao2023all) (see Section 2), a less commonly used design that we revisit in Section 4.2 and find helpful for performance.
By default, we use cross-attention to inject text embeddings. For some experiments, we additionally validate on single- and dual-stream variants (see Section 2) of our backbone to ensure generality of our findings. We use AdaLN to condition the model on the sum of the timestep embedding and a pooled text embedding, computed by averaging over text embedding tokens. For the single-stream architecture, we follow Lumina-Image 2.0 (qin2025lumina) and prepend two modality-specific refiner blocks to the backbone. For both single- and dual-stream backbones, we adopt Multimodal-RoPE (wang2024qwen2). By default, we use the encoder part of T5Gemma-2B as our text encoder and use FLUX.2 VAE.
ImageNet-22K | YFCC | RedCaps | Megalith |
Places | Pexels | iNaturalist | FLUX-Reason |
Midjourney v6 | GPT-Edit | TextAtlas | RenderedText |
Data
. We exclusively use publicly available image datasets, including 7 real-image datasets (ImageNet-22K (deng2009imagenet), YFCC100M (thomee2016yfcc100m), RedCaps (desai2021redcaps), Megalith (BoerBohan2024Megalith10m), Pexels (Narugo2024PexelsTaggerV0), iNaturalist 2024 (vendrow2024inquire), Places365-Challenge 2016 (zhou2017places)), 3 synthetic datasets (GPT-Image-Edit-1.5M (wang2025gpt), FLUX-Reason-6M (fang2026flux), and Midjourney v6 (CortexLM2024MidjourneyV6)), and 2 text-rendering datasets (RenderedText (Wendler2024RenderedText) and TextAtlas (wang2025textatlas5m)). By default, we naively combine the 168M images in these datasets without weighting (a design choice that we revisit later in Section 5.2). In pre-training, all images are center-cropped to squares and resized to 256256. We present example images from each dataset in Figure 6. We generate one long synthetic caption per image using the prompt “Describe the image in detail using one paragraph.” with Qwen3-VL-30B-A3B (bai2025qwen3) in FP8 precision (the VLM receives images that are center-cropped to squares, and resized to 512512 if larger than 512512). Further information is provided in Appendix E.
| baseline variant | DPG | PRISM | LongText |
| cross-attention | 84.66 | 56.4 | 0.211 |
| single-stream | 85.89 | 55.6 | 0.293 |
| dual-stream | 86.82 | 58.3 | 0.439 |
Training and inference
. We train the model using the flow matching (lipman2022flow) objective for 500K iterations (i.e., 25% of the 2M-step 256-resolution pre-training stage of our final i1 model) with a batch size of 512 and a learning rate of 1e-4. We use a 250-step Euler integrator with a classifier-free guidance (ho2022classifier) scale of 12 for sampling. More details are in Appendix A.1.
Prompt: On a reflective metallic table, there is a brightly colored handbag featuring a floral pattern next to a freshly sliced avocado… with silverware and a clear glass water bottle positioned neatly beside the avocado… (77 words)

Prompt: Lindsey Wixson stands confidently in a golden wheat field, donning a wide-brimmed straw hat, bold red sunglasses, and a vibrant red fur-trimmed top, accessorized with a sparkling diamond necklace, embodying summer elegance.

Prompt: An elegant, professional-looking mobile interface for a productivity and habit-tracking app named "DailyFlow". Positioned prominently at the top is the app name in bold, rounded typography colored soothing teal… (170 words)

Evaluation
. We use three widely used benchmarks to provide signals for our controlled experiments: DPG-Bench (hu2024ella), PRISM-Bench (fang2026flux), and LongText-Bench (geng2025x). All three benchmarks use VLMs as evaluators. DPG and PRISM measure fine-grained prompt-following capabilities across diverse prompts, where PRISM additionally evaluates image aesthetics. LongText specifically evaluates text-rendering capability. We present example prompts from each benchmark in Figure 7. We use original prompts for DPG and LongText and rewrite PRISM prompts using Qwen3-4B (yang2025qwen3) with the simple meta-prompt in Section 5.1. Table 1 reports the performance of our baselines on these benchmarks.
Section 4 Modeling
We study modeling design modifications to the baseline introduced in Section 3. We first revisit text and noise conditioning mechanisms, including multiple text encoders and AdaLN, and identify stronger alternative designs. We then explore backbone architecture choices. More results are provided in Appendix C.
4.1 Text and Noise Conditioning
Existing methods have explored various ways of incorporating text and noise conditioning into the backbone diffusion model. Some use a single text encoder (cai2025z; qin2025lumina; wu2025qwen), while others concatenate features from multiple text encoders (esser2024scaling; cai2025hidream). In addition, embeddings of the noise level are often injected into the model through AdaLN (peebles2023scalable), sometimes together with a pooled text embedding (esser2024scaling; cai2025hidream; labs2025flux). We investigate this broad design space and show that, rather than combining multiple encoders, it is more beneficial to use a single strong text encoder with a larger adapter. We also find that AdaLN-based conditioning on noise level and pooled text embeddings may not be necessary for text-to-image models.
Text encoder
. Early models (e.g., SD 1.5 (rombach2022high)) primarily used CLIP-style text encoders. Later models adopted the encoder–decoder model T5 (raffel2020exploring; esser2024scaling; cai2025hidream). Most recently, models often use decoder-only LLMs or VLMs (wu2025qwen; qin2025lumina), a trend often attributed to their powerful reasoning and complex instruction following capabilities (xie2025sana).
Here, we compare (1) a modern CLIP-based encoder (FG-CLIP 2 (xie2025fg)), (2) two families of modern encoder–decoder models, T5Gemma (zhang2025encoder) and T5Gemma2 (zhang2025t5gemma), (3) a family of modern decoder-only LLMs (Qwen3 (yang2025qwen3)), and (4) a family of modern decoder-only VLMs (Qwen3-VL (bai2025qwen3)). Unless otherwise specified, we use instruction-tuned checkpoints when available and otherwise use the corresponding base checkpoints. Figure 8 reports the text-to-image performance using each model as the text encoder (comparisons under alternative settings are in Appendices C.3 and C.4).
We observe that instruction tuning has minimal impact (e.g., T5Gemma-2B vs. T5Gemma-2B (base)) and larger models do not necessarily perform better (e.g., T5Gemma-2B vs. T5Gemma-9B). Most importantly, encoder-decoder models (T5Gemma and T5Gemma2) achieve the best overall performance, outperforming FG-CLIP 2 and the decoder-only LLMs and VLMs. This leads to the first finding that affects our design:
Combining text encoders
. Many recent models (esser2024scaling; cai2025hidream) combine text features from multiple encoders (e.g., CLIP, T5, and LLMs). Here, we experiment with different combinations of the text encoders evaluated in Figure 8. Previous work uses different strategies to combine embeddings from different encoders (e.g., embedding- vs. sequence-dimension concatenation). In this work, we concatenate text features along the sequence dimension and use a separate adapter for each encoder to accommodate their different embedding dimensions. This avoids the need to pad text features to a common sequence length, which would often be required when concatenating features along the embedding dimension.
We combine one of the strongest text encoders from Figure 8, T5Gemma-2B, with one additional encoder. As Table 2 shows, combining it with T5Gemma2-1B or FG-CLIP 2 yields the best performance. However, combining all three (T5Gemma-2B, T5Gemma2-1B, and FG-CLIP 2) provides no substantial further gains.
| type | T5G-2B | T5G2-1B | T5G2-4B | Qwen3-VL-2B | FG-CLIP 2 | DPG | PRISM | LongText |
| baseline | ✓ | 84.66 | 56.4 | 0.211 | ||||
| +1 encoder | ✓ | ✓ | 85.62 | 58.4 | 0.303 | |||
| ✓ | ✓ | 84.86 | 55.8 | 0.270 | ||||
| ✓ | ✓ | 84.80 | 57.4 | 0.264 | ||||
| ✓ | ✓ | 85.72 | 57.7 | 0.285 | ||||
| +2 encoder | ✓ | ✓ | ✓ | 85.37 | 58.8 | 0.272 | ||
| ✓ | ✓ | ✓ | 85.28 | 58.1 | 0.351 |
Although combining text encoders improves performance, does the improvement arise from the diverse representations provided by different encoders or simply from the increased sequence length and additional parameters introduced by the adapters? To investigate this, we construct two baselines that repeat the T5Gemma-2B text embeddings: the first uses two separate adapters for the two identical copies of embeddings (thus increasing both sequence length and adapter parameters), while the second uses a shared adapter (thus increasing only sequence length). As shown in Table 3, repeating the embeddings with two separate adapters brings a noticeable improvement, whereas using a shared adapter produces results similar to the baseline without repetition. This suggests that the gains from combining multiple text encoders may largely stem from the additional adapters rather than from diverse text encoder features or longer sequences.
| text encoder | DPG | PRISM | LongText |
| T5Gemma-2B | 84.66 | 56.4 | 0.211 |
| repeat w/1 MLP | 84.93 | 55.8 | 0.225 |
| repeat w/2 MLP | 85.09 | 56.5 | 0.309 |
We sanity check that the two MLP adapters learn distinct features by comparing embeddings for DPG-Bench, PRISM, and LongText prompts from each adapter. We measure cosine similarity between the two embeddings for each token and average it across tokens for each prompt. The resulting distribution, shown in Figure 9, is largely negative, indicating that the two adapters indeed capture different representations.
Larger text encoder adapter
. To further investigate the hypothesis that performance improvements from combining multiple text encoders largely stem from the additional adapter parameters instead of diverse text encoder features, we replace the small MLP adapter (2.6M parameters) used in the default setup (Section 3) with larger transformer adapters (17.2M parameters/block) with the same width as the backbone blocks. As shown in Figure 10, despite the marginal increase in parameter count, increasing the adapter capacity consistently improves performance across all backbone architectures. Nonetheless, expanding the adapters beyond two transformer blocks yields only marginal additional gains.
Furthermore, as the “default” and “+2 encoders” rows of Table 4 show, when using a larger adapter, combining multiple text encoders yields much smaller gains across all backbones, especially on DPG and PRISM. This further suggests that the benefit of multiple text encoders can be captured by increasing the adapter capacity for a single text encoder. Importantly, using multiple encoders increases the text sequence length, substantially raising memory and computational cost, whereas using a larger adapter does not.
| MLP adapter (default) | transformer adapter (1x block) | ||||||||
| #params | DPG | PRISM | LongText | #params | DPG | PRISM | LongText | ||
| cross-attn | |||||||||
| default | 0.89B | 84.66 | 56.4 | 0.211 | 0.91B | 86.33 | 58.7 | 0.414 | |
| +2 encoders | 0.90B | 85.37 0.71 | 58.8 2.4 | 0.272 0.061 | 0.94B | 86.47 0.14 | 59.5 0.8 | 0.491 0.077 | |
| no pooled emb | 0.89B | 85.98 1.32 | 57.5 1.1 | 0.391 0.180 | 0.91B | 86.37 0.04 | 59.4 0.7 | 0.446 0.032 | |
| no timestep | 0.89B | 82.58 2.08 | 54.7 1.7 | 0.185 0.026 | 0.91B | 84.71 1.62 | 58.9 0.2 | 0.418 0.004 | |
| no AdaLN | 0.66B | 84.99 0.33 | 57.4 1.0 | 0.351 0.140 | 0.67B | 85.13 1.20 | 59.7 1.0 | 0.413 0.001 | |
| single-stream | |||||||||
| default | 0.82B | 85.89 | 55.6 | 0.293 | 0.83B | 87.64 | 60.0 | 0.472 | |
| +2 encoders | 0.83B | 84.89 1.00 | 56.3 0.7 | 0.439 0.146 | 0.87B | 87.29 0.35 | 59.0 1.0 | 0.428 0.044 | |
| no AdaLN | 0.57B | 87.38 1.49 | 59.0 3.4 | 0.390 0.097 | 0.58B | 87.39 0.25 | 59.5 0.5 | 0.410 0.062 | |
| dual-stream | |||||||||
| default | 1.24B | 86.82 | 58.3 | 0.439 | 1.25B | 87.67 | 60.7 | 0.576 | |
| +2 encoders | 1.25B | 87.34 0.52 | 59.6 1.3 | 0.514 0.075 | 1.29B | 87.76 0.09 | 60.8 0.1 | 0.588 0.012 | |
| no AdaLN | 1.01B | 87.82 1.00 | 60.3 2.0 | 0.508 0.069 | 1.02B | 87.38 0.29 | 60.7 0.0 | 0.554 0.022 | |
Removing AdaLN conditioning
. Adaptive Layer Normalization (AdaLN) (peebles2023scalable) is a standard component in modern text-to-image diffusion models (labs2025flux; wu2025qwen; qin2025lumina; cai2025hidream). It is typically used to inject timestep embeddings and pooled text embeddings into the backbone. A recent study (sun2025noise) showed that in class-conditional image generation, removing noise conditioning only minimally affects performance, especially for flow matching models. If AdaLN can be removed without harming performance, the model could become more parameter-efficient.
Interestingly, as shown in Table 4, removing AdaLN from the default setup (Section 3) consistently improves performance when the text encoder adapter is a small MLP. However, the effect becomes much smaller when using a larger transformer adapter. To better understand this behavior, we perform additional ablations on the cross-attention backbone, where AdaLN conditions only on pooled text embeddings or only on timestep embeddings, rather than their sum. We evaluate these variants with both small and large adapters.
We find that AdaLN conditioning on pooled text embeddings reduces performance when the adapter is small (84.99 82.58 on DPG), but has a much smaller effect when the adapter is large (85.13 84.71 on DPG). This suggests that the performance gains from removing AdaLN may primarily result from poorly learned features when using the small MLP adapter. However, even when a larger adapter is used, conditioning on the pooled text and timestep embeddings through AdaLN still provides marginal additional benefit.
4.2 Backbone Architecture
In this subsection, we revisit the long skip connection design and provide a controlled comparison of popular backbone families based on the baseline setup in Section 3. We include additional analyses of positional embeddings, normalization, and VAEs in Appendices C.1 and C.2.
Long skip connections
add shortcuts between early and later layers. They were first popularized by U-Net (ronneberger2015u) and later applied to diffusion models in U-ViT (bao2023all). While they were shown to improve performance (bao2023all; li2024hunyuan; liu2024playground), they have not been widely applied to modern text-to-image models. In Figure 11, we revisit this design by training dual-stream variants of the baseline (Section 3) with and without long skip connections at multiple model widths (1152, 1296, 1440, 1584, and 1728), while keeping all other configurations fixed. We find that long skip connections consistently improve performance across model sizes, potentially due to enhanced model expressivity. Additional FLOPs-based analysis is in Figure 44, and results on other backbones are in Appendix C.7.
Backbone family
. Today’s leading models differ in their choice of backbone: some use cross-attention, some use single-stream architectures, and others use dual-stream architectures (see Section 2 for details). We measure model performance for cross-attention, single-stream, and dual-stream backbones at multiple model widths (1152, 1296, 1440, 1584, and 1728 for all three backbone families) while keeping all other model configurations fixed. Figure 12 plots model performance against parameter count. We observe that the dual-stream backbone achieves the best performance-parameter trade-off.
Section 5 Data
Besides modeling architectures, high-quality image-caption data is important for text-to-image training. In this section, we first study synthetic captioning designs, and show that training on long captions yields stronger models but can lead to poor performance on short prompts, which we mitigate via prompt rewriting at inference. We then explore dataset mixing and find that equal weighting across datasets is a strong default.
5.1 Synthetic Captions and Prompt Rewrite
Prompt-following capability in text-to-image models fundamentally relies on high-quality image-caption pairs in the training data. Earlier work, such as Parti (yu2022scaling) and DALL-E 3 (betker2023improving), showed that training on highly descriptive synthetic captions generated by vision-language models can substantially improve performance. Since then, the majority of text-to-image models (chen2024pixart; esser2024scaling; qin2025lumina; xie2025sana) have leveraged synthetic captions during training.
Here, we explore several design choices in synthetic caption generation and their impact on model performance (more results in Appendix D.1). In particular, we find that training on long synthetic captions yields stronger models, but these models can underperform on short prompts, necessitating inference-time prompt rewrite. To reduce computational cost for caption generation, all experiments in this section are conducted on the ImageNet-22K dataset rather than the full training set used in the default baseline setting (Section 3).
. To explore how caption quality impacts downstream text-to-image performance, we generate captions using five VLMs: Qwen2-VL 2B, Qwen2.5-VL 3B, Qwen3-VL-2B, Qwen3-VL-4B, and Qwen3-VL-30B-A3B. As reported in Figure 13, the choice of synthetic captioner has a substantial impact on downstream text-to-image performance. We note that the small differences on LongText are primarily due to the overall poor text-rendering performance of models trained on ImageNet-22K, which contains few text-rich images. This result therefore does not imply that captioner quality is unimportant for text rendering.
. By default, we train our models using only long synthetic captions (see Section 3). While our models can achieve strong performance on the original DPG and LongText prompts, they perform poorly on original GenEval prompts. We find that this may be explained by the much shorter prompts in GenEval compared to DPG and LongText (see Figure 35): simply repeating the GenEval prompts 12 times leads to a large improvement in performance (0.17 0.49). This observation suggests that the poor performance on original, short GenEval prompts may stem from training exclusively on long captions.
To further understand this, we generate an additional set of short captions using the prompt “Describe the image using one short sentence.” The distributions of prompt lengths are shown in Figure 14. We mix these short captions with the original long captions using different sampling weights and report the resulting GenEval scores in Table 5. We observe that (1) training primarily on short captions (e.g., 0% or 20% long captions) improves performance on the original short GenEval prompts and (2) models trained with higher proportions of long captions perform better when the GenEval prompts are repeated.
| % of long captions in training captions | performance on GenEval prompts | ||||
| original prompts (short) | repeated prompts | rewritten prompts (long) | |||
| 4 | 12 | 20 | |||
| 0% | 0.47 | 0.55 | 0.34 | 0.24 | 0.60 |
| 20% | 0.47 | 0.54 | 0.53 | 0.50 | 0.67 |
| 40% | 0.35 | 0.59 | 0.55 | 0.54 | 0.70 |
| 60% | 0.37 | 0.60 | 0.57 | 0.54 | 0.73 |
| 80% | 0.26 | 0.57 | 0.54 | 0.47 | 0.73 |
| 100% | 0.17 | 0.48 | 0.49 | 0.46 | 0.73 |
While repeating the short prompts can recover the performance, it introduces unnatural prompt structures. To address this issue, we instead use an LLM (Qwen3-4B) to rewrite the GenEval prompts using the following meta-prompt:
“I have a short text-to-image prompt {prompt}. Please expand it into a descriptive paragraph, while making sure the generated image still clearly includes all the items mentioned in the original prompt. Please only output the rewritten prompt and nothing else.”
As shown in the rightmost column of Table 5 and Figure 15, rewriting the GenEval prompts substantially improves model performance. Notably, training on long captions and evaluating on rewritten prompts (0.73) significantly outperforms training on short captions and evaluating on original, repeated, or rewritten prompts. This suggests that, even when inference prompts are originally short (e.g., GenEval), it is preferable to train on long captions and increase the inference prompt length to match the training distribution (e.g., via prompt rewriting), rather than training on short captions to match the original inference prompt length.
| Prompt: a photo of a wine glass and a bear | ![]() | ![]() | ![]() | ![]() |
| Prompt: a photo of a zebra right of a parking meter | ![]() | ![]() | ![]() | ![]() |
| train: short, test: original (short) | train: long, test: original (short) | train: long, test: repeated 12 (long) | train: long, test: rewritten (long) |
5.2 Data Mixing
All experiments up to this point naively combine all datasets without explicit dataset-level weighting. Because our training corpus (see Section 3) is highly imbalanced (e.g., YFCC contributes 98M of 168M images), we implicitly assign much larger weights to a few large datasets, which can dominate the training signal. In this subsection, we study how dataset composition and dataset-level reweighting affect performance.
Contributions of dataset components
. To understand how each dataset contributes to performance, we train a separate model on each dataset. Evaluation results are shown in Figure 16(a). Among real-image datasets, ImageNet-22K and YFCC achieve the best overall performance, while iNaturalist performs substantially worse, likely due to its narrow domain. FLUX-Reason and GPT-Edit perform particularly well on PRISM. LongText scores are low for every dataset except TextAtlas, consistent with the scarcity of text-containing images in real and synthetic datasets. This suggests that text rendering capability relies on specialized text-rich datasets. To control for dataset size, we further train models on random 1M subsets of each dataset. As shown in Figure 16(b), the relative performance trends remain largely the same as in the full dataset case.
Further, we test whether real, synthetic, or text-rendering data can be removed from the baseline (Section 3) without harming performance. As shown in Figure 17, removing real images hurts DPG, whereas removing synthetic images hurts PRISM. Further, LongText performance is directly correlated with the proportion of text rendering data (10.4% for “full”, 66.7% for “remove real”, 10.9% for “remove synthetic”, and 0% for “remove text”). These results indicate that the three groups of images provide complementary benefits.
Equal dataset weighting
. By default (Section 3), we naively combine all datasets without explicit dataset-level weighting, so each dataset’s effective sampling weight is simply its number of images. Inspired by the data balancing strategy of capping the number of data points from a single source in VLM training (tong2024cambrian), we cap each dataset’s sampling weight using four hand-picked thresholds. Results in Figure 18 show that a threshold of 1.2M, which gives equal weight to all datasets, achieves strong overall performance.
Given the effectiveness of equal dataset weighting (see Figure 18), we further explore two simple variants. First, we remove low-quality real datasets one at a time while keeping the remaining datasets equally weighted. Table 6 shows that removing iNaturalist provides a clear gain across all benchmarks, while further removing additional real datasets offers no substantial improvement. Second, after removing iNaturalist, we test whether any single dataset should be emphasized by upweighting one dataset by 3 or 5 while keeping the remaining datasets equally weighted. As shown in Figure 19 and Figure 46, upweighting any single dataset does not surpass the performance of the fully balanced dataset.
| datasets | DPG | PRISM | LongText |
| full | 85.14 | 58.2 | 0.335 |
| remove iNaturalist | 85.56 | 58.7 | 0.384 |
| remove iNaturalist + Megalith | 85.13 | 59.0 | 0.438 |
| remove iNaturalist + Megalith + Places | 85.18 | 57.9 | 0.453 |
Data magnitude
. Figure 16 provides preliminary evidence that subsampling datasets often has marginal impact on model performance. To probe how much performance depends on the unique number of images in the training set, we also train on random subsets of ImageNet-22K. As shown in Figure 20, especially when using 5 captions per image, subsampling from 13.7M to 0.4M images only causes marginal degradation. Only when shrinking to 0.1M do we see a substantial drop. Since our 500K-step recipe already repeats the full ImageNet-22K set 18.7 times, these results suggest that using fewer unique images and repeating them more often may not substantially degrade performance for text-to-image diffusion models.
| subset size for each dataset | unique #imgs seen | DPG | PRISM | LongText |
| full | 88.1M∗ | 85.56 | 58.7 | 0.384 |
| 1.0M | 11.0M | 85.34 | 57.7 | 0.384 |
| 0.4M | 4.4M | 84.67 | 57.7 | 0.382 |
| 0.1M | 1.1M | 84.71 | 57.4 | 0.349 |
We further extend the dataset subsampling experiments on a single dataset to mixtures of datasets. Specifically, we begin with a data mixture that assigns equal weight to the 11 datasets, excluding iNaturalist. For each dataset in the mixture, we randomly subsample it to contain exactly 1.0M, 0.4M, or 0.1M images while maintaining equal sampling weights across datasets. The resulting model performance is shown in Table 7. Even when each dataset is reduced to 0.4M images (resulting in 4.4M unique images seen instead of 88.1M), the performance decrease across benchmarks is minimal. This suggests that, with a diverse mix of datasets, repeating training data incurs only marginal performance degradation in text-to-image diffusion training.
Section 6 i1-3B: State-of-the-Art Performance Among Fully Open Models
In the previous sections, we explored the modeling and data designs that can improve text-to-image performance. Building on these insights, we train i1, a model with 3B parameters that performs competitively with leading models across several representative benchmarks.111We are additionally training a 1B model and will release it soon. In this section, we describe the final pre-training, high-resolution training, and inference setups and experiments, and present the evaluation results.
6.1 Low-Resolution Pre-training
Model
. The architecture of i1 is illustrated in Figure 21. It uses a dual-stream MMDiT backbone with long skip connections, the FLUX.2 VAE, and T5Gemma-2B as the text encoder, along with a large adapter composed of two transformer blocks. i1 removes all AdaLN parameters and thus does not use noise conditioning. Additionally, we use both sinusoidal and RoPE positional embeddings, and share sandwich normalizations across text and image streams (see Appendix C.1 for corresponding controlled experiments).
Data
. We use the best data mixing recipe identified in Section 5.2, where we assign equal weights to 6 real image datasets, 3 synthetic datasets, and 2 text-rendering datasets. We use Qwen3-VL-30B-A3B to generate multiple long synthetic captions for each image. Due to resource constraints, we generate five synthetic captions per image for ImageNet-22K, Pexels, RenderedText, GPT-Edit, RedCaps, FLUX-Reason, TextAtlas, and Midjourney v6, two per image for YFCC, and one per image for Places and Megalith.
| Prompt: Argentinian soccer star Lionel Messi in the heat of the 2022 FIFA World Cup Final against France. He is… about to strike the ball with his left foot… (240 words) | ![]() | ![]() | ![]() | ![]() |
| Prompt: An appealing poster… announcing a folk music concert event… At the top-center, the inviting phrase "Let Acoustic Melodies Inspire Your Soul"… (109 words) | ![]() | ![]() | ![]() | ![]() |
| 100K iterations | 200K iterations | 500K iterations | 2M iterations |
Training.
We extend the number of training iterations in the default recipe (Section 3) to 2M steps while keeping all other hyperparameters unchanged. We train i1 at 256-resolution until performance plateaus around 2M steps, as shown in Figure 22. We additionally show example generated images in Figure 23 and observe that the benchmark improvements are accompanied by improved image quality. Details of the training setup and compute resources are in Appendix A.1.
6.2 High-Resolution Training
Data and modeling
. To construct the 512- and 1024-resolution training sets, we retain only images whose shorter edge is at least 512 or 1024 pixels, respectively. We remove any dataset entirely if the filtered set contains fewer than 0.3M images (see the resolution statistics for each dataset in Appendix E.2). Based on our findings in Section 5.2, we further subsample each dataset with more than 1M images to 1M images and assign equal sampling weight to every dataset. At 1024-resolution, we discard RenderedText due to its low quality (see Figure 16). Following esser2024scaling, we perform positional index interpolation and timestep schedule shifting during 512- and 1024-resolution training (details in Appendix A.1).
Results
. We train the model for 0.5M steps at 512-resolution and 0.3M steps at 1024-resolution. The benchmark performance trends during training are in Appendix A.1, and the final 1024-resolution checkpoint is evaluated in Section 6.3. As shown in Figure 24, 512-resolution training substantially improves the LongText score (0.75 0.92). We further illustrate the improvements in text rendering with qualitative examples in Figure 25.
We also study how different dataset components contribute to 512-resolution training. Starting from the 256-resolution checkpoint, we train separate models using only real image datasets, only synthetic image datasets, or only text-rendering datasets at 512-resolution. As shown in Figure 24, training on either real or synthetic image datasets yields LongText improvements comparable to training on the full dataset, despite both subsets containing limited text-rich images. This suggests that strong high-resolution generation capability does not require high-resolution training data to match the full breadth of the low-resolution pre-training data.
![]() | ![]() |
![]() | ![]() |
6.3 Inference and Evaluation
Inference setup
. During inference, we use a CFG scale of 12 and apply the Rescale CFG technique (lin2024common) with a rescale strength of 1. Unlike previous methods (wang2024emu3; deng2025emerging; pan2025transfer) that apply prompt rewriting to particular benchmarks, we use a single meta-prompt (details in Appendix B.3) for rewriting all input prompts to match training prompt lengths, as motivated in Section 5.1.
Benchmarks
. We evaluate our model on five representative benchmarks commonly used in the technical reports of recent image generation models (cai2025z; qin2025lumina; cai2025hidream; cui2025emu3): GenEval (ghosh2023geneval), DPG-Bench (hu2024ella), PRISM-Bench (fang2026flux), CVTG-2K (du2025textcrafter), and LongText-Bench (geng2025x). GenEval focuses on object-centric image generation and evaluates a fixed set of object attributes and relationships. DPG-Bench and PRISM-Bench provide fine-grained evaluation of general prompt-following capabilities, with PRISM-Bench additionally assessing image aesthetics. CVTG-2K and LongText-Bench evaluate a model’s ability to generate images containing detectable text that matches the description in the input prompt.
We note that prior work has suggested that GenEval may be misaligned with human judgment (kamath2025geneval) and poorly correlated with human-perceived model capability (cao2025hunyuanimage). Additionally, it is a common practice in current models (chen2025blip3; ma2026deco; wang2026pixnerd) to fine-tune on BLIP3o-60K (chen2025blip3), which can inflate GenEval scores, as BLIP3o-60K fine-tuning was found to significantly improve GenEval scores but not other benchmarks (wu2025openuni). Therefore, we report GenEval results only for completeness and note that they may not accurately reflect model capability.
model #params GenEval DPG-Bench PRISM CVTG-2K LongText-Bench API call only GPT Image 1 [High] (gptimage1) - 0.84* 85.15* - 0.8569* 0.956* Seedream 3.0 (gao2025seedream) - 0.84* 88.27* - 0.5924* 0.896* Open weights only FLUX.1 [Dev] (labs2025flux) 12B 0.66* 83.84* 65.1 0.4965* 0.607* SD3 Medium (esser2024scaling) 2B 0.62* 84.08* 61.9 0.4037 0.322 Janus-Pro-7B (chen2025janus) 7B 0.80* 84.19* 60.0 0.0667 0.019* BAGEL (deng2025emerging) 14B 0.88* 85.44 61.8 0.3642 0.373* HiDream-I1-Full (cai2025hidream) 17B 0.83* 85.89* 66.1 0.7738 0.543* Lumina-Image 2.0 (qin2025lumina) 3B 0.73* 87.20* 63.5 0.1577 0.088 Z-Image (cai2025z) 6B 0.84* 88.14* 74.2 0.8671* 0.935* Qwen-Image (wu2025qwen) 20B 0.87* 88.32* 73.9 0.8288* 0.943* Open weights + data + training code BLIP3o-4B (chen2025blip3) 4B 0.77 79.73 53.2 0.0353 0.023 PixNerd (wang2026pixnerd) 1B 0.73* 80.9* 53.3 0.0006 0.020 DeCo (ma2026deco) 1B 0.86* 81.4* 53.1 0.0014 0.003 BLIP3o-N-S (chen2025blip3o) 3B 0.87 81.98 56.8 0.2493 0.110 BLIP3o-N-G-G (chen2025blip3o) 3B 0.90 81.93 57.5 0.2442 0.114 BLIP3o-N-G-T (chen2025blip3o) 3B 0.86 79.77 56.8 0.3330 0.153 i1 (Ours) 3B 0.84 86.73 70.1 0.8531 0.922
Results
. We compare the i1 model with leading image generation systems in Table 8. i1 achieves state-of-the-art performance among fully open models on all five benchmarks except GenEval. It also outperforms several leading weight-only models, including Lumina-Image 2.0, HiDream-I1, and FLUX.1 [Dev]. i1’s strong performance reflects the combined effect of the modeling and data choices identified throughout our study.
Section 7 Discussion and Conclusion
Fully open recipes support cumulative research
in text-to-image modeling. A challenge in current text-to-image research is that strong models are often released as opaque endpoints rather than as inspectable scientific artifacts. As a result, progress can be difficult to attribute across various (potentially undisclosed) design factors. Our study advocates for fully open recipes that seek to understand which design choices reliably matter. By releasing the model, code, data recipe, and ablations behind i1, we aim to provide not only a strong baseline, but also a reference point for more cumulative and reproducible research.
Strong performance does not require sophisticated designs
. The strong performance of recent text-to-image models can create the impression that frontier capability requires increasingly specialized architectures, proprietary data, or heavily engineered recipes. Our study provides a counterpoint: strong performance can be achieved with moderately scaled (e.g., 4.4M, see Section 5.2) and publicly available datasets and a careful exploration of the current modeling design space. We believe that this is encouraging for open research, as competitive text-to-image models need not begin from inaccessible data or undisclosed training procedures.
Limitations and future work
. This work has several limitations. First, our evaluation relies primarily on automated benchmarks, which emphasize prompt following, rather than human preference. Thus, although i1 approaches leading weight-only models (e.g., Qwen-Image) on these benchmarks, its generated images remain noticeably inferior in overall visual quality (we present failure cases in Appendix B.5). Second, due to resource constraints, all experiments are conducted with models of roughly 3B parameters or smaller. Further experiments are needed to determine whether our findings continue to hold at substantially larger scales. Third, our exploration only covers a subset of the text-to-image diffusion model design space: designs such as multi-aspect ratio training222We are working on a multi-aspect ratio model and will release it soon., data filtering (startsev2026alchemist), deep fusion of decoder-only LLMs and diffusion transformer for text encoding (liu2024playground; shi2026lmfusion), and reinforcement learning (wallace2024diffusion; liu2026flow) are omitted. Future work could extend our recipe to larger models and further explore the design space while preserving the simplicity and openness of the overall pipeline.
Acknowledgements
We gratefully thank the Google TPU Research Cloud (TRC) program for providing the primary computing resources for this project. Additional support was provided by the Princeton Research Computing resources at Princeton University, which are managed by a consortium of groups led by the Princeton Institute for Computational Science and Engineering (PICSciE) and Research Computing. We would like to thank Liang-Chieh Chen, Ishan Misra, Kaiming He, Yida Yin, Haozhe Chen, Wenhao Chai, Linrong Cai, Linzhan Mou, and Xingyu Fu for valuable discussions and feedback. We also thank Yufeng Xu, Shengbang Tong, Yiyang Lu, and Hanhong Zhao for helpful discussions on TPU. We are grateful to Cihang Xie’s research group for sharing their JAX DiT codebase, which served as the launching point for our research.
References
Appendix
Appendix A Implementation Details
In this section, we provide further details on our modeling and training configurations.
A.1 Configuration
Hardware
. Our model training and inference are conducted on TPU v4, v5p, and v6e with JAX (jax2018github). Benchmark evaluations are performed on NVIDIA A100, H100, and H200 GPUs.
General configuration
. Our default baseline models largely follow the XL/2 model configurations used in previous diffusion models (peebles2023scalable; yao2025reconstruction), which use a hidden size of 1152, 16 attention heads, an MLP ratio of 4.0, and a patch size of 2. However, unlike those models, we use 29 layers instead of 28. By default, during both training and inference, we maintain the text encoder in bf16 while keeping all other parameters in fp32. To ensure the models fit into memory, we shard model parameters and optimizer states across devices using JAX pjit/GSPMD (xu2021gspmd), following ZeRO-style fully sharded data parallelism (rajbhandari2020zero).
| config | value |
| optimizer | Adam |
| learning rate | 1e-4 |
| weight decay | 0 |
| optimizer momentum | |
| batch size | 512 |
| learning rate schedule | constant |
| gradient clipping | 1 |
| training objective | flow matching |
| training steps | 500K |
| training timestep distribution | lognorm(0, 1) |
| inference timestep shift value (esser2024scaling) | 0.3 |
| inference steps | 250 |
| CFG scale (ho2022classifier) | 12 |
| CFG rescale strength (lin2024common) | 0 |
| CFG interval (kynkaanniemi2024applying) | [0, 1] |
256-resolution controlled experiments
. Table 9 summarizes the training configuration for all controlled experiments in Sections 4 and 5. All models are trained for 500K iterations, but training time varies because different experiments use different model components. For the cross-attention baseline, 500K steps take 31.0 hours on a TPU v6e-64 machine.
| training stage | #images | training steps | batch size | training timestep shift value (esser2024scaling) | TPU v5p-128 hours |
| 256-resolution | 162.9M | 2.0M | 512 | N/A | 383.0 |
| 512-resolution | 9.7M | 0.5M | 512 | N/A | 174.4 |
| 1024-resolution | 4.3M | 0.3M | 128 | 3.33 | 150.9 |
i1 training
. In Table 10, we detail the training configurations and compute resources for the final i1 model at each training stage. All unspecified configurations are kept the same as in Table 9. The benchmark performance trends across iterations for the high-resolution training stages are shown in Figure 26. We observe that 512-resolution training substantially improves performance on PRISM and LongText, whereas 1024-resolution training has a smaller effect, with performance remaining close to that of the 512-resolution checkpoint from which it is initialized. For 1024-resolution training, we additionally compare models trained with a timestep shift value of 3.33 against models trained without a timestep shift, and find that applying the training timestep shift consistently improves performance.

(a) 512-resolution training

(b) 1024-resolution training
A.2 Baseline Architectures
As described in Section 3, our controlled experiments in Sections 4 and 5 are all based on a fixed baseline architecture. We vary one design choice at a time while keeping all other configurations identical to the baseline. Although we use the cross-attention backbone as the default in our baseline, we additionally validate some design choices on single-stream and dual-stream backbones. In this section, we provide illustrations of the three backbone architectures.
Cross-attention backbone
passes text conditioning information to the backbone through cross-attention layers inserted between the self-attention and feed-forward network layers. The architecture is illustrated in Figure 27.
Single-stream backbone
concatenates the text features and noisy image features along the sequence dimension and processes the entire sequence using a single set of backbone weights. The architecture is illustrated in Figure 28.
Dual-stream backbone
concatenates the text features and noisy image features along the sequence dimension, but uses separate backbone parameters for the text tokens and image tokens. The architecture is illustrated in Figure 29.
A.3 Details on Text Encoders
Model version
. For all T5Gemma models, we use the UL2 (tay2023ul) variant, as it has better encoder representations (zhang2025encoder). For the T5Gemma-9B model, we use the variant with a 2B decoder instead of the one with a 9B decoder. For the FG-CLIP 2 model, we use the “long” mode.
Truncation
. Following mainstream implementations (esser2024scaling; blackforestlabs_flux2_2025; cai2025z; wu2025qwen), we use right truncation for the text tokenizers. We truncate to 256 tokens for all text encoders except FG-CLIP 2, which we truncate to 196 tokens because it is trained on up to 196 tokens.
Hidden states
. For encoder-decoder models (i.e., the T5Gemma and T5Gemma2 families), we use the encoder’s final-layer hidden states as text token features. For decoder-only models (i.e., the Qwen3 and Qwen3-VL families), we use the last hidden states from the final transformer layer as text token features. By default, we input the text-to-image prompt directly into the text encoder to obtain features. Some previous work (ma2024exploring; xie2025sana; wu2025qwen) applied system prompts to LLM/VLM text encoders; we ablate the effect of system prompts in Appendix C.4.
Appendix B Additional Information on Inference and Evaluation
B.1 Qualitative Comparison with Stable Diffusion 3 Medium
In Figures 2 and 3, we presented selected example images generated by our i1 model. In Figures 30 and 31, we provide four additional curated examples of our model’s generations, and compare them with images generated by Stable Diffusion 3 Medium using the same prompts.
Prompt: Veronica Lake (1922-1973), the US actress, is depicted sitting in an armchair, dressed in a red blouse under a white apron dress, engrossed in reading a book, evoking a scene from around 1955.
Prompt: A bright, welcoming bakery interior, captured in warm, soft morning light, showcasing an appealing, rustic wooden-framed chalkboard menu placed prominently against a white brick wall. At the center of the chalkboard, in large, elegant hand-drawn lettering, reads clearly "Today’s Specials: Sourdough Bread & Cinnamon Rolls". Just below this central message, smaller text neatly notes "Freshly baked every morning". At the top right corner of the chalkboard, subtly written in a playful cursive handwriting, are the words "Homemade with Passion". Around the borders of the menu, slightly faded and vintage-inspired illustrations of wheat stalks and pastries subtly frame the text, enhancing the artisanal bakery atmosphere. Beside the main chalkboard stands a smaller wooden sign, on which handwritten text reads "Free samples available!", accompanied by a decorative arrow directing customers toward the display counter. The textual elements are distinct, stylish, and naturally handwritten, evoking a genuine, artisanal feel, perfectly complementing the inviting bakery ambiance.
Prompt: A serene oil painting titled "Midnight Moon" by David Forks captures a solitary figure standing on rocky shores under a luminous full moon, rendered with dramatic lighting and a deep blue color palette, evoking a contemplative and atmospheric mood.
Prompt: A contemporary, artistic movie poster with minimalistic, impactful textual layout. At the top center, a bold, sleek-font title reads "The Last Voyage", positioned above a subtle silhouette of an old sailing ship facing turbulent waves. Immediately under the silhouette, an intriguing tagline appears in slightly smaller letters: "When courage means sailing into the unknown". In the central lower half of the poster, neatly arranged textual phrases in a clear, horizontal alignment describe key highlights: "Directed by Award-Winning Director Alex Rivers", "Starring Emily Clarke & Jacob Bennett", "Featuring Original Music by Daniel Harper". At the very bottom, in concise, uppercase lettering set clearly apart, the release details read: "In Cinemas Everywhere October 6, 2023". The lower-left corner includes a smaller, thin-lined text in italics: "Will you brave the journey?"
B.2 Evaluating Other Models under Different Inference Settings
The inference settings of i1 (see Section 6.3) use a CFG scale of 12, which is higher than the default values used by many existing text-to-image diffusion models. For example, PixArt- (chen2024pixart) uses a default CFG scale of 4.5, Lumina-Image 2.0 (qin2025lumina) uses 4, SANA (xie2025sana) uses 4.5, and Stable Diffusion 3 (esser2024scaling) uses 7. In addition, we use a custom meta-prompt for inference-time prompt rewriting. These choices may raise the question of whether our inference settings give our method an unfair advantage over baseline models. To examine this, we evaluate Lumina-Image 2.0 and Stable Diffusion 3 Medium under alternative inference settings, including larger CFG scales and prompts rewritten with our meta-prompt (see Appendix B.3). The results are shown in Table 11. We find that neither increasing the CFG scale nor using rewritten prompts substantially improves the performance of either model.
| prompt | CFG scale | DPG | PRISM | LongText |
| original | 4 | 87.20 | 63.5 | 0.088 |
| 8 | 87.39 | 60.7 | 0.100 | |
| 12 | 87.56 | 60.9 | 0.101 | |
| 16 | 87.84 | 58.6 | 0.107 | |
| rewritten | 4 | 85.37 | 63.1 | 0.092 |
(a) Lumina-Image 2.0
| prompt | CFG scale | DPG | PRISM | LongText |
| original | 7 | 84.08 | 61.9 | 0.322 |
| 8 | 85.49 | 61.2 | 0.341 | |
| 12 | 84.94 | 56.2 | 0.361 | |
| 16 | 82.82 | 50.1 | 0.368 | |
| rewritten | 7 | 84.43 | 61.0 | 0.313 |
(b) Stable Diffusion 3 Medium
B.3 Meta-Prompt for Prompt Rewrite
In Section 5.1, we found that training on short captions leads to weaker overall models, whereas training on long captions yields stronger models but leads to poor performance on short prompts. Prompt rewriting can mitigate this training-inference prompt-length mismatch by expanding short inference prompts, making training on long captions preferable to training on short captions, even when the original inference prompts are short. For the experiments in Table 5, we used a simple, minimal meta-prompt for expanding short GenEval prompts into longer prompts.
However, always instructing the prompt-rewriting LLM to expand the input prompts may not be optimal, since inference prompts are of variable length and can even be longer than training captions. Therefore, we design a more comprehensive meta-prompt that instructs the model to follow two different sets of guidelines depending on the complexity of the input prompt. At the end of the meta-prompt, we additionally include 20 hand-crafted pairs of original and rewritten prompts as in-context examples to guide the LLM.
The meta-prompt, including the 20 in-context examples, is provided below.
B.4 Ablation of Inference Steps
In both the controlled experiments in Sections 4 and 5 and the evaluation of i1 in Table 8, we fix the number of inference steps to 250. However, we note that 250 steps are not necessary for strong model performance. In Figure 32, we evaluate the performance of our i1 model using 5, 10, 20, and 50 inference steps, and show a qualitative example in Figure 33. We observe that we can reduce the number of inference steps to as low as 20 without substantially hurting generation quality.
| 10 | 20 | 250 |
![]() | ![]() | ![]() |
B.5 Failure Cases of i1
Despite i1’s strong performance, we note that i1 still exhibits several important failure cases, as illustrated in Figure 34. As Figure 34(a) shows, especially when tasked with generating multiple small human figures in a group setting, i1 sometimes generates human faces with poor fidelity and unnatural facial expressions, as well as malformed hands or limbs. Moreover, i1 does not always respect physical properties and can sometimes generate physically implausible images. Figure 34(b) provides one such example: i1 fails to capture the physical behavior of a mirror, as the reflection suggests that the mirror is parallel to the car window, which is physically inconsistent.
B.6 Prompt Length Distributions of Benchmarks
In Figure 35, we visualize the prompt length distributions of the benchmarks used in the final evaluation of i1 (see Table 8). We observe that GenEval has much shorter prompts than the other benchmarks. This motivated our focus on GenEval when analyzing the poor short-prompt performance of models trained exclusively on long captions, as well as the corresponding mitigation techniques (see Section 5.1).
Appendix C Additional Results on Modeling Designs
In Section 4, we drew several conclusions about modeling designs through controlled experiments. Here, in Appendices C.1 and C.2, we provide additional experimental results that motivate the positional embedding, normalization, and VAE used in the final i1 model. In the remaining subsections, we provide additional results and analyses that validate our findings from Section 4 under alternative settings.
C.1 Positional Embedding and Normalization
In our final i1 model, we use both sinusoidal and RoPE positional embeddings, adopt sandwich normalization, and share normalization layers across the text and image streams in MMDiT. We describe below the experiments that motivated these design choices.
Positional embeddings
. Current text-to-image diffusion models often use only one type of positional embedding (e.g., sinusoidal (esser2024scaling) or RoPE (cai2025z; wu2025qwen; qin2025lumina)). Inspired by the design in LightningDiT (yao2025reconstruction), we explore whether combining sinusoidal and RoPE embeddings improves the text-to-image performance of a diffusion transformer. As shown in Figure 36, combining the two meaningfully improves benchmark performance for cross-attention and dual-stream models. As such, we use both sinusoidal and RoPE positional embeddings in our final i1 model.
Normalization
. While pre-norm (xiong2020layer) (i.e., normalizing the input to the attention and feed-forward network modules) has been the dominant choice for diffusion transformers, an earlier work (ding2021cogview) introduced sandwich norm (i.e., normalizing both the input and the output of the attention and feed-forward network modules) to stabilize training, which has recently been adopted by Z-Image (cai2025z). We introduce sandwich normalization to our baseline models in Figure 37 and find that it can stably improve performance across backbone architectures and benchmarks.
| model | DPG | PRISM | LongText |
| shared norms | 86.82 | 58.3 | 0.439 |
| separate norms | 86.16 | 57.6 | 0.415 |
Further, while using separate normalization layers for text and image modalities has been the default for existing MMDiT models (esser2024scaling; cai2025hidream; wu2025qwen), we explore whether a more unified feature distribution from normalization layers shared across the modalities would benefit model performance. As shown in Table 12, sharing the normalizations indeed consistently improves performance.
C.2 VAEs
| VAE | DPG | PRISM | LongText |
| FLUX.2 | 84.66 | 56.4 | 0.211 |
| Qwen-Image | 83.29 | 54.3 | 0.266 |
| VA-VAE | 85.07 | 55.5 | 0.126 |
We compare VA-VAE (yao2025reconstruction) with the VAEs used in frontier models FLUX.2 (blackforestlabs_flux2_2025) and Qwen-Image (wu2025qwen). Overall, FLUX.2 achieves the most balanced performance across all benchmarks. Likely due to its alignment with pre-trained semantic features during training, VA-VAE achieves the strongest performance on DPG-Bench, which emphasizes semantic alignment of generated images with prompts. However, it performs much worse than the others on LongText, which evaluates fine-grained text rendering. This may be due to its lower reconstruction fidelity.
| original | FLUX.2 VAE | Qwen-Image VAE | VA-VAE |
![]() | ![]() | ![]() | ![]() |
![]() | ![]() | ![]() | ![]() |
Quantitatively, prior work (wu2025qwen; yao2025reconstruction) reports that on the ImageNet validation set at 256256 resolution, VA-VAE has lower reconstruction performance (PSNR 27.96, SSIM 0.79) than FLUX.2 VAE (31.46, 0.90) and Qwen-Image VAE (33.42, 0.92). In Figure 38, we further provide qualitative reconstruction examples on text-rich images. While FLUX.2 VAE and Qwen-Image VAE reconstruct faithfully, VA-VAE introduces visible corruption and distortion in the rendered characters. One possible reason for this limitation is that VA-VAE was trained on ImageNet (deng2009imagenet), which lacks text-rich images.
C.3 Comparing Text Encoders under Alternative Settings
Larger adapter
. In Section 4.2, we compared the text encoder candidates on our default baseline model with a small MLP adapter. However, as we later showed in Section 4.1, the size of the adapter has a substantial impact on model performance. Thus, here we additionally explore whether our comparisons between text encoders still hold when we use a larger text encoder adapter consisting of two transformer blocks as in our final i1 recipe. As shown in Figure 39, our observations still largely hold: the T5Gemma and T5Gemma2 families of encoder-decoder models achieve the strongest performance, while FG-CLIP 2 is the weakest.
AdaLN removed
. In our baseline setup (Section 3), a pooled text embedding is combined with the timestep embedding and passed into the backbone through AdaLN. Since this provides an additional path for injecting text information, removing AdaLN, as in our final i1 recipe, may affect the relative performance of different text encoders. We present the benchmark results for each text encoder in Figure 40. We observe that the overall trends are highly similar to the trends in the default setting, where AdaLN is used (Figure 8).
C.4 Applying System Prompts to Text Encoders
In the default setup (see Appendix A.3), we directly process the raw prompt with each text encoder and use the last hidden states as text features. While this setup is common (cai2025z), prior work has also designed specialized prompting strategies when using decoder-only language models as text encoders (ma2024exploring; xie2025sana). Here, we follow the strategy used by Qwen-Image (wu2025qwen) for the Qwen2.5-VL text encoder and apply it to the Qwen3-VL-2B and Qwen3-VL-4B text encoders.
Concretely, we wrap the text-to-image prompt in a system message (“Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:”), and feed the full sequence into the text encoder. After the forward pass, we discard the hidden states corresponding to the system prefix and keep only those corresponding to the text-to-image prompt. The resulting performance is shown in Table 14. We observe that using the system prompt brings minor improvements for Qwen3-VL-2B, but the Qwen3-VL models still underperform the T5Gemma and T5Gemma2 models.
| text encoder | system prompt | DPG | PRISM | LongText |
| Qwen3-VL-2B | ✗ | 82.29 | 52.2 | 0.076 |
| ✓ | 82.82 | 52.8 | 0.093 | |
| Qwen3-VL-4B | ✗ | 82.07 | 52.9 | 0.071 |
| ✓ | 82.41 | 52.4 | 0.065 |
C.5 Comparing Backbone Families with Training FLOPs
In Figure 12, we compared different backbone families by training cross-attention, single-stream, and dual-stream models with widths of 1152, 1296, 1440, 1584, and 1728, and plotting performance against model size. However, depending on the practical training and inference setting, comparing performance across FLOPs may be more informative. In Figure 41, we therefore plot performance against trainable model FLOPs. These FLOPs are computed using JAX/XLA’s “cost_analysis” for one forward and backward pass through all trainable modules in the diffusion model (including e.g. the text encoder adapter). We use the training tensor shapes and scale the result by the global batch size and the number of training steps. The dual-stream backbone still achieves the best trade-off.
C.6 Validating Modeling Designs on Larger Models
Performance vs. model size
. In Section 4 and Appendix C.1, except for the backbone-family comparison and long skip connections, which we validated across model sizes, we mainly identified modeling design findings using an XL/2-sized baseline. In Figure 43, we further validate these findings on dual-stream MMDiT models across multiple model sizes by training one model for each width in {1152, 1296, 1440, 1584, and 1728}.
Across model sizes, larger text encoder adapters (Section 4.1) consistently provide better performance-parameter trade-offs. Removing AdaLN (Section 4.1) also yields a clear advantage when using an MLP text encoder adapter. However, when using a larger transformer text encoder adapter, models with and without AdaLN achieve similar performance at comparable parameter counts. We note that this still suggests that noise conditioning may not be necessary in text-to-image diffusion models.
Finally, while Appendix C.1 provided preliminary results suggesting that combining Sinusoidal and RoPE positional embeddings, using sandwich normalization, and sharing normalizations across image and text streams can improve performance, we do not observe these trends consistently across model scales.
Performance vs. training FLOPs
. In addition to comparing performance across model sizes in Figures 11 and 43, we compare performance against estimated training FLOPs for trainable modules in Figure 44. We compute these FLOPs following the same procedure as in Appendix C.5.
Most trends remain unchanged under the FLOPs-based comparison. One exception is the effect of AdaLN when using a larger transformer-based text encoder adapter: in this setting, models with AdaLN have a better performance-FLOPs trade-off than models without AdaLN. This is because AdaLN contributes a non-trivial fraction of the model parameters (e.g., 18.9% of parameters for the dual-stream baseline) but only minimally increases training FLOPs, since its projection is computed once per sample rather than once per token.
C.7 Validating Long Skip Connection Results on Other Backbones
In Section 4.2, we showed that long skip connections consistently improve model performance across model sizes, based on experiments with a dual-stream MMDiT backbone. Here, we further validate this design on other backbones by training other variants of the baseline model (which is XL/2-sized) with and without long skip connections. As Figure 42 shows, removing long skip connections noticeably reduces performance across most backbones and benchmarks, especially on DPG and LongText. This further suggests that long skip connections can broadly benefit model performance.
C.8 Text Feature Adapter vs. Image Feature Adapter
In Section 4.1, we found that replacing a small MLP adapter with a larger transformer adapter for the text encoder substantially improves performance across benchmarks, despite adding few parameters. We hypothesize that this is because text features from pre-trained language models need to be adapted for downstream tasks such as text-to-image generation. In Appendix C.6, we showed that the larger transformer adapter achieves a better performance-parameter trade-off than the smaller MLP adapter.
| backbone | adapter | DPG | PRISM | LongText |
| cross-attention | default | 84.66 | 56.4 | 0.211 |
| + transformer adapter for text features | 86.33 | 58.7 | 0.414 | |
| + transformer adapter for image features | 85.27 | 55.4 | 0.250 | |
| single-stream | default | 85.89 | 55.6 | 0.293 |
| + transformer adapter for text features | 87.64 | 60.0 | 0.472 | |
| + transformer adapter for image features | 86.10 | 59.7 | 0.345 | |
| dual-stream | default | 86.82 | 58.3 | 0.439 |
| + transformer adapter for text features | 87.67 | 60.7 | 0.576 | |
| + transformer adapter for image features | 86.23 | 57.0 | 0.378 |
To further verify that this improvement is not merely due to increased parameter count, we analogously add a transformer block to the image features, after patchification and before applying positional embeddings (see Appendix A.2). As shown in Table 15, adding this adapter to the image features results in much smaller performance improvement compared to using it on the text features. This suggests that the gains from larger adapters are specific to adapting pre-trained text features rather than simply increasing model capacity.
C.9 Exploring Variants of Long Skip Connections
Long skip connections were popularized by U-Net (ronneberger2015u) to provide shortcuts for low-level features from earlier layers, thereby easing training for pixel-level prediction tasks. Later, U-ViT (bao2023all) followed this design and applied it to transformer-based diffusion models. However, while the exactly symmetric structure of these connections (i.e., the -th leftmost layer is connected to the -th rightmost layer) is natural for the multi-resolution encoder-decoder structure of U-Net, it is not necessarily optimal for transformer-based models, whose blocks often have the same feature dimensionality. Therefore, here, we explore different variants of the original long skip connections in U-ViT.
Layer range
. Long skip connections from earlier layers skip across more blocks, whereas those closer to the middle skip across fewer blocks. As a result, features from layers closer to the middle may have changed less by the time they reach their destination layers, making these skip connections potentially less necessary. Removing them could therefore preserve performance or even improve it. To test this hypothesis, in Figure 45, we explore several ranges of layers from which long skip connections can start: 1-3, 1-7, 1-11, 4-7, 4-11, 4-14, 8-11, 8-14, and 12-14. However, none of these variants consistently outperforms the default setting.
| skip type | DPG | PRISM | LongText |
| default | 84.66 | 56.4 | 0.211 |
| 83.70 | 55.2 | 0.134 | |
| 84.78 | 55.8 | 0.180 |
Connection pattern
. We further explore whether long skip connections should connect layers with larger representational differences. Instead of using the default symmetric pattern, which connects the -th leftmost layer to the -th rightmost layer, we test variants that connect the -th layer to the -th or -th layer. As shown in Table 16, on the cross-attention backbone, the variant slightly improves DPG, but the default pattern still performs best overall, achieving the highest PRISM and LongText scores.
Appendix D Additional Results on Data Designs
In Section 5, we presented controlled experiments motivating our designs for synthetic captioning, prompt rewrite, and data mixing. We provide further details on our synthetic captioning designs, along with corresponding controlled experiments, in Appendix D.1, and additional results supporting our conclusion on data mixing in Appendix D.2.
D.1 Additional Designs in Synthetic Captioning
In Section 5.1, we studied synthetic caption generation by training a baseline cross-attention model on ImageNet-22K images with different caption sets. Here, we provide additional results that motivated two choices in our caption-generation pipeline: center-cropping images before captioning and generating multiple captions per image.
| captioner | DPG | PRISM | LongText |
| Qwen3-VL-30B-A3B | 83.72 | 50.8 | 0.007 |
| + no center-crop | 83.16 | 51.3 | 0.006 |
| + 5 captions/image | 83.56 | 51.9 | 0.010 |
Image cropping
. In our experimental setup, we train on square images obtained by center-cropping the longer edge to match the shorter edge. If the synthetic captioner receives the full uncropped images, the generated captions may describe objects that are later cropped out, creating a semantic mismatch between the training images and captions. podell2023sdxl suggested that this may contribute to the failure mode of text-to-image models generating partial objects. Further, by default, after center-cropping, we resize images larger than 512512 down to 512512 to avoid slow captioning on larger images.
To understand the impact of our pre-processing operations, we train our model on two sets of synthetic captions generated by Qwen3-VL-30B-A3B: one produced from the full images and the other from the cropped and resized square images. As shown in the first two rows of Table 17, captions generated from the cropped and resized images lead to similar downstream performance as captions from full images. Since center-cropping and resizing improve captioning speed without meaningfully affecting downstream performance, we apply both operations before generating all synthetic captions.
. Increasing the number of captions per image provides another axis of data scaling, beyond increasing the number of images. To explore this, we generate five captions per image using Qwen3-VL-30B-A3B. As shown in Table 17, this leads to modest improvements on PRISM and LongText. Moreover, as shown in Figure 20, this benefit becomes more pronounced when the number of unique training images is limited.
D.2 Experiments under Equal Dataset Weights
In Section 5.2, after identifying that equal weighting for all datasets can result in strong performance, we further explored whether upsampling a single dataset, while keeping the others equally weighted, could further improve results. In Figure 19, we demonstrated that upsampling any one dataset by does not consistently improve performance across all benchmarks. Here, we further show in Figure 46 the results for upsampling each dataset by 5, and find that the results are consistent with the observations in Figure 19.
Appendix E Additional Information on Datasets
We briefly introduced the image datasets used for model training in Section 3 and explored synthetic captioning and data mixing strategies in Section 5. Here, we provide additional details about the images and captions used during training.
E.1 Visualizations of Images
In our study, we explored data mixing strategies and trained models on 12 publicly available image datasets, including 7 real-image datasets (ImageNet-22K (deng2009imagenet), YFCC100M (thomee2016yfcc100m), RedCaps (desai2021redcaps), Megalith (BoerBohan2024Megalith10m), Pexels (Narugo2024PexelsTaggerV0), iNaturalist 2024 (vendrow2024inquire), Places365-Challenge 2016 (zhou2017places)), 3 synthetic datasets (GPT-Image-Edit-1.5M (wang2025gpt), FLUX-Reason-6M (fang2026flux), and Midjourney v6 (CortexLM2024MidjourneyV6)), and 2 text-rendering datasets (RenderedText (Wendler2024RenderedText) and TextAtlas (wang2025textatlas5m)). To provide a qualitative understanding of their distributions, we randomly sample 9 images from each dataset, resize each image so that its shorter edge is 256 pixels, center-crop it to a 256256 square, and visualize the resulting images in Figure 47.
E.2 Image Resolution Statistics
As discussed in Section 6.2, we use all image datasets, excluding iNaturalist, for the 256-resolution training of i1. For high-resolution training at 512/1024 resolution, we filter out images whose shorter edge is smaller than 512/1024 pixels. We also remove any dataset that contains fewer than 0.3M images after filtering.
Here, we report the number of images that remain under each resolution threshold in Table 18. For 512-resolution training, we use all datasets except YFCC and iNaturalist. For 1024-resolution training, we use only FLUX-Reason, TextAtlas, RedCaps, GPT-Edit, and Midjourney v6 (although RenderedText satisfies the image-count requirement after resolution-based filtering, we exclude it due to its low quality).
The RedCaps dataset we use contains 5M images, which differs from the 12M images reported in the original paper. This is because RedCaps images are provided as URLs, many of which are no longer accessible.
| image type | dataset | #imgs | #imgs w/ shorter edge512 | #imgs w/ shorter edge1024 |
| real | YFCC | 98,121,424 | 0 | 0 |
| ImageNet-22K | 13,673,551 | 719,427 | 0 | |
| Megalith | 9,393,971 | 9,202,294 | 61,701 | |
| Places | 8,026,628 | 7,345,764 | 0 | |
| RedCaps | 4,817,431 | 4,705,536 | 4,134,303 | |
| iNaturalist | 4,813,543 | 0 | 0 | |
| Pexels | 2,810,634 | 2,810,621 | 0 | |
| synthetic | FLUX-Reason | 5,890,279 | 5,890,279 | 5,890,279 |
| GPT-Edit | 1,553,575 | 1,552,821 | 357,482 | |
| Midjourney v6 | 1,240,185 | 1,240,185 | 1,240,185 | |
| text-rendering | RenderedText | 11,977,824 | 11,977,824 | 11,977,824 |
| TextAtlas | 5,397,762 | 4,036,973 | 919,913 |
E.3 Caption Length Distribution for Different VLM Captioners
In Section 5.1, we used the same meta-prompt to prompt Qwen2-VL 2B (wang2024qwen2), Qwen2.5-VL 3B (qwen2.5-vl), Qwen3-VL-2B, Qwen3-VL-4B, and Qwen3-VL-30B-A3B (bai2025qwen3) to generate synthetic captions for ImageNet-22K. We then train a diffusion model on each resulting image-caption dataset. We find that the choice of VLM used for synthetic captioning has a substantial impact on downstream text-to-image performance. In Figure 48, we further plot the sequence length distribution of the captions generated by each VLM. We observe that caption sets that lead to stronger performance also tend to be longer.
E.4 Caption Length Distribution for Different Datasets
Here, we randomly sample 10K images from each dataset and plot the sequence length distribution of their synthetic captions in Figure 49. Although caption length distributions vary across datasets, no dataset differs substantially from the others.
E.5 Meta-Prompt for Synthetic Captioning
For most image datasets, we use the following minimal prompt for synthetic captioning:
For text-rendering datasets, however, ground-truth text annotations are available. To reduce hallucinations in the VLM-generated captions, we include the ground-truth text in the captioning prompt. For TextAtlas, we use:
RenderedText consists of images of handwritten text rendered on digital 3D sheets of paper using Blender. The text varies in font size, color, and rotation, and the paper is rendered under random lighting conditions. Each ground-truth annotation contains a field, “text”, which is a list of strings, where is the number of text lines in the image. The -th element of the list corresponds to the -th line of text. Therefore, for RenderedText, we use:
E.6 Sanity Check for Data Leakage
For all datasets except FLUX-Reason, all captions are synthetically generated by a VLM, so overlap between the prompt sets used in our evaluation benchmarks and the captions used to train our text-to-image models is unlikely. However, for FLUX-Reason, we use five sets of captions: one set is the original “caption_detail” field provided in the dataset (since it is already high-quality), and the remaining four are synthetically generated by us. According to the original paper (fang2026flux), half of PRISM-Bench’s prompts were directly selected from FLUX-Reason and then removed from the dataset. To ensure that there is no overlap between FLUX-Reason and PRISM-Bench, we perform exact string matching between each FLUX-Reason “caption_detail” caption and each PRISM-Bench prompt. We find no overlap.

































































