{hanjian.thu123,liujinlai.licio}@bytedance.com
{wangjiahuan.123,bingyue.peng,yuanzehuan}@bytedance.com
代码与模型:
yuanzehuan@bytedance.com
摘要
尽管扩散模型在视觉生成领域占据主导地位,但其计算效率仍然低下,因为它们对复杂度不同的样本分配了相同的计算量。相比之下,自回归(AR)模型天生具有复杂度感知能力,其变动的似然值即可证明这一点,但这类模型常常受限于有损离散化 token 化以及误差累积。在本工作中,我们提出了生成式精炼网络(GRN),这是一种旨在解决上述问题的下一代视觉合成范式。其核心在于,GRN 通过一种理论上近乎无损的分层二进制量化(HBQ)技术,解决了离散化 token 化的瓶颈问题,实现了与连续式方法相媲美的重建质量。基于 HBQ 的潜在空间,GRN 从根本上对 AR 生成进行了升级,引入了一种全局精炼机制,能够像人类艺术家作画一样,逐步完善和修正生成的作品。此外,GRN 还整合了一种基于熵的采样策略,能够在无需牺牲视觉质量的前提下,实现复杂度感知的自适应步长生成。在 ImageNet 基准测试中,GRN 在图像重建(0.56 rFID)和类别条件图像生成(1.81 gFID)任务上均创下了新纪录。我们还将 GRN 扩展到了更具挑战性的文生图和文生视频任务中,在同等规模下实现了更优的性能。我们公开了所有模型和代码,以促进对 GRN 的进一步研究。
1 引言
视觉生成领域发展迅猛,其主要驱动力是扩散 Transformer 架构的规模化应用 [dit, sora, hunyuanvideo, waver, alive]。这些模型通过沿着一个学习到的速度场逐步整合轨迹——该速度场将简单的噪声先验传输到经验数据分布——展现出合成高质量视觉内容的强大能力。然而,这种连续流范式天生缺乏自适应步长能力。由于通过均方误差(MSE)进行优化,且不依赖显式似然,这些模型被限制在固定的步数内,无论样本的复杂程度如何变化,都僵化地为所有样本分配相同的计算资源。
与此同时,受大语言模型中 token 级似然估计成功经验的启发 [gpt3.5, gpt4],自回归(AR)模型在视觉合成领域也引起了广泛的研究兴趣 [videogpt, keyuVAR, hanjInfinity, wang2024emu3]。然而,当前的自回归方法受限于两个关键缺陷。首先,当使用离散 token 而非连续表示时,它们本质上存在重建质量较差的问题。其次,其严格的因果预测机制——无论是逐 token 操作还是逐尺度操作——都会在多步生成过程中不可避免地导致严重的误差累积。这暴露了其纠错能力的严重缺失,因为模型无法追溯性地修正之前的错误。此外,即使在掩码自回归模型 [maskgit, bert] 中采用并行预测,高置信度的 token 也会变得不可更改,后续无法修正。因此,这类模型本质上仍然缺乏全局性的精修机制。
这些观察结果启发了一种简单而直观的、基于精炼的自回归框架,并具备自适应计算能力。为此,我们提出了生成式精炼网络(GRN),这是一种新一代视觉合成范式,旨在克服扩散模型计算成本固定僵化的局限以及标准自回归模型固有的缺陷。具体来说,为了解决离散模型 token 重建质量不佳的问题,我们首先提出了分层二值量化(HBQ)。通过在不增加潜在通道的情况下确保重建误差呈指数级衰减,HBQ 使得离散图像和视频 tokenizer 能够实现近乎无损的重建,在更高压缩率下达到与连续 tokenizer 相当的性能。在这些稳健的表示基础上,GRN 采用熵引导的采样机制,执行一种复杂度感知的自适应步长生成过程。它根据视觉内容的不同难度动态分配计算负载,同时利用全局精炼机制来追溯性地减轻累积误差。
在多种视觉任务上进行的大量实验验证了我们框架的优越性。在用于类别条件图像合成的 ImageNet 256×256 基准测试上,GRN 在图像重建和生成质量两方面均创下了新纪录。此外,为了展示其卓越的任务泛化能力和可扩展性,我们还成功地将 GRN 扩展到了高分辨率文生图(T2I)和文生视频(T2V)场景。在扩展后,GRN 展现出生成逼真的 1024×1024 图像以及动态、高保真、时长 2 到 10 秒的 480p 视频的能力。总之,我们的主要贡献如下:
-
我们提出了 GRN,这是一种新一代视觉合成框架。其特点是具备全局精炼机制和复杂度感知生成能力,实现了稳健且高效的视觉生成。
-
我们引入了分层二值量化,并贡献了一系列离散图像/视频 tokenizer。这是首次,离散视觉 tokenizer 在相同潜在维度下与连续 tokenizer 达到了同等水平。
-
大量实验表明,GRN 在标准 C2I 基准测试上取得了最先进的结果,rFID 达到 0.56,gFID 达到 1.81。当扩展到更具挑战性的 T2I 和 T2V 任务时,与同等规模的方法相比,它展现出了更优越的性能。
2 相关工作
2.1 视觉分词器
视觉分词器 [ldm, vqvae, vqgan] 通过压缩视觉内容以实现高效生成。早期的向量量化方法 [vqvae, vqgan] 将连续特征映射到离散码本,但受限于可扩展性不足,这促使了免查找方法 [BSQ, fsq] 的出现,以实现更大的词汇表。尽管如此,与连续表示之间仍存在性能差距。近期的工作 [hanjInfinity, bitdance] 旨在通过大幅扩展词汇表来缩小这一差距,其性能超越了连续 VAE。然而,这种提升是以收敛速度变慢和生成模型规模变大作为代价的,这促使人们探索更高效的量化方案。
2.2 自回归模型
受大语言模型的启发,[vqgan, llamagen, videopoet, wang2025editinfinity] 通过下一个 token 预测来探索视觉生成。MaskGIT [maskgit] 使用并行解码来加速生成,它首先生成高置信度的 token,然后迭代地填充剩余部分。VAR [keyuVAR] 将自回归转变为下一尺度预测,从而提升了质量并实现了更快的推理速度。然而,AR 模型仍然受限于有损的离散分词和误差累积,并且仍落后于扩散方法。尽管 Infinity [hanjInfinity] 通过随机翻转比特级 token 引入了自我修正机制,但其假设扩散误差低于 30% 的条件仅覆盖了有限的模式。
2.3 自适应步长生成
扩散模型主导了视觉生成领域 [FLUX, SDXL, Stable Diffusion 3, Sora, Wan],但通常需要数十步推理步骤。蒸馏方法 [DMD, DMD2] 大幅减少了采样步数,却仍然依赖于预定义的固定步数调度。这种“一刀切”的策略在简单提示词上浪费了计算资源。最近,AdaDiff [AdaDiff] 采用一个外部网络来确定实例特定的步数,并使用策略梯度方法来最大化奖励。这种复杂的流程需要一个额外的网络和奖励信号。
3 方法
3.1 视觉分词器
视觉分词器在学习紧凑的潜在空间以压缩高维真实数据方面起着至关重要的作用。我们采用了 Wan 2.1 [Wan] 中提出的 3D 因果 VAE 设计,以便在统一的框架中对图像和视频进行分词。具体来说,给定一张图像或一段视频,分词器将其时空信息编码到特定维度,同时将通道数扩展到特定数量。由于我们的目标是实现无损离散压缩,而 VAE 特征是连续信号,因此我们将特征量化视为一个信号变换问题。受信号处理中哈尔小波 [Haar1910Theorie] 的启发,我们引入了层次化二进制量化,将 VAE 特征转换为离散特征。
层次化二进制量化。我们首先在 VAE 编码器之后附加一个非线性激活函数,将特征表示从无界范围映射到闭区间。如图 2 所示,中的每个元素都基于一个具有中心定义的桶的二叉树,按照公式 1 和公式 2 进行多轮二进制量化。
| (1) |
| (2) |
其中,当 时, 是一个值为 -1 的狄拉克δ函数,否则为 1。由此我们得到量化后的二值标签 ,其中 。这里的 是分层二值量化的总轮数。通过这种方式,我们从粗到细地进行量化,以表示不同频率的信息,并且第 轮的量化误差小于 。量化误差的上界随轮数呈指数级衰减,理论上可以快速实现无损量化。图 3 展示了从量化中间结果重建的图像,揭示了从粗到细的特性。
| (3) |
随后,可以根据公式 3 推导出量化后的特征 。HBQ 的详细算法见附录 A。在视觉分词器的训练阶段,量化后的特征 作为输入被送入解码器,以重建原始图像或视频 。遵循训练离散视觉分词器的常见做法,我们采用直通估计器(STE)将梯度反向传播至编码器。训练损失是重建损失()、LPIPS 感知损失()和来自 PatchGAN 判别器的 GAN 损失()的加权组合。
分词后,我们得到大小为 的二值输出。然而,由于这相当于引入一个大小为 的巨大码本,几乎不可能通过合并 和 维度将输出编码为 INT 标量来进行生成。受按位 token [hanjInfinity] 的启发,我们提出了 GRN 的两种变体,即 GRNind 和 GRNbit,以支持生成。对于 GRNind,我们简单地将 维度编码为 INT 标量,得到 。对于 GRNbit,我们拼接最后两个维度并预测 。对于这两种变体,我们展平时空维度,并通过多头 token 预测 [deepseek_v3] 并行预测每个 token 的整个通道维度。
3.2 生成式精炼网络
受人类绘画直觉的启发,我们提出了一种极其简洁的自回归精炼框架用于视觉生成,该框架从随机 token 映射开始。令 表示第 步中 token 映射的状态。目标是基于当前状态 预测绘制映射 。为了显式地形式化这一过程,我们将 定义为三个组成部分的组合:随机映射 、绘制映射 以及二元选择映射 。该关系在公式 (4) 中正式表达,其中 通过根据 中的值从 或 中选择来构建。
| (4) |
直观上, 代表当前的绘制内容,而 则对应于没有任何信息的空白区域,模拟了人类绘画过程中的中间步骤。 的设计目的是使累积统计量 (即 中 1 的比例)在精炼步骤中从 0% 单调递增至 100%。因此, 逐渐收敛到理想的 token 映射。为了获得 ,我们采用了一个 Transformer 架构,并将 近似为
| (5) |
通过拟合真实的 token 映射来建模下一步的绘制,其中 表示生成条件,例如类别嵌入或文本。 根据 按照公式 (6) 构建。我们还研究了一种基于预测置信度的替代方法来构建 。然而,如附录 E.3 所述,该方法产生了较差的结果。
| (6) |
我们引入了一种复杂度感知的采样策略,以控制其不仅保持单调性,同时考虑其不确定性。下文将详细讨论。基于预测结果和选择映射,状态按照公式 4 所述进行相应更新。由此,形成了一个渐进式生成与优化的连贯循环。自回归机制使我们不仅能够以高确定性改进越来越多的模型 token,还能在绘制过程中随着更多上下文的纳入而消除明显错误。理想情况下,随着信息不断积累,该过程会收敛到最佳结果。该简化过程如图 4 所示。
训练。在训练阶段的每次迭代中,我们从均匀分布中随机采样随机 token,用于 GRNind 和 GRNbit。二值映射也以不同的选择比例进行均匀采样,该比例控制有多少真实 token 被用作输入。因此,Transformer 的输入由从真实 token 中采样的 token 和从随机 token 中采样的 token 组成。请注意,token 采样是沿着空间、时间和通道维度随机进行的,不附加任何先验知识(公式 6)。表示 token 总数,等于。以公式 4 构建的包含部分信息的作为输入,我们的目标是预测真实 token,类似于扩散设置中的 x 预测,并使用公式 7 所示的简单交叉熵损失。这里表示真实 token。
| (7) |
GRN 的详细训练和推理过程见附录 B。GRN 与其他自回归模型之间的其他比较见附录 C。
复杂度感知采样。我们提出了一种基于熵的调度函数来确定,其中表示优化步骤的索引。具体来说,我们在生成过程中计算第步的平均熵,如下所示:
| (8) |
在公式 8 中,我们将 表示为模型 token 索引,将 表示为类别索引,其中 是类别总数。注意,对于 GRNind 和 GRNbit 分别有相应定义。生成复杂度通过熵 来衡量,其取值范围在 0 到 1 之间。由于较小的 表示更高的预测置信度,我们分配较少的细化步骤,同时让 更陡峭地增加,从而从 中保留更多信息。相反,当高熵表明复杂度较高时,我们应用更多的细化步骤以及 的渐进式增长。具体来说,我们将公式定义为
| (9) |
这里, 表示从特定步骤计算出的平均熵。我们设置了一个预热阶段,其中 和 ,因为我们观察到在初始步骤中熵值不稳定。超参数 控制自适应步骤的动态范围,而 是偏置项。我们还会对 的值进行裁剪,以确保总推理步骤数保持在 范围内。
| 方法 | 分词器 | 空间 | 潜变量 | 通道 | 压缩 | rFID | LPIPS | SSIM | PSNR |
|---|---|---|---|---|---|---|---|---|---|
| 类型 | 比率 | 通道 | 比特 | 比率 | |||||
| SD-VAE [ldm] | 连续型 | 16 | 16 | 16 | 24 | 0.87 | - | 0.68 | 24.08 |
| RAE [RAE] | 连续型 | 16 | 768 | 16 | 0.5 | 0.62 | 0.25 | 0.44 | 19.20 |
| VAR† [keyuVAR] | 离散型 | 16 | 不适用 | 不适用 | 193 | 0.85 | 0.15 | 0.64 | 22.47 |
| LlamaGen [llamagen] | 离散型 | 16 | 1 | 14 | 439 | 2.19 | - | 0.68 | 20.79 |
| Open-MAGVIT2 [open_magvit2] | 离散型 | 16 | 1 | 18 | 341 | 1.17 | - | - | 22.64 |
| HBQ (M=4) | 离散型 | 16 | 16 | 4 | 96 | 0.56 | 0.13 | 0.71 | 23.01 |
| 方法 | 分词器 | 潜变量 | 空间 | 时间 | 通道 | 压缩 | rFVD | LPIPS | SSIM | PSNR |
| 类型 | 通道 | 步长 | 步长 | 比特 | 比率 | |||||
| Wan 2.1 | 连续型 | 16 | 8 | 4 | 16 | 24 | 19.5 | 0.058 | 0.929 | 34.10 |
| Wan 2.2 | 连续型 | 48 | 16 | 4 | 16 | 32 | 22.6 | 0.052 | 0.932 | 34.54 |
| HBQ (无量化) | 连续型 | 16 | 16 | 4 | 16 | 96 | 144.6 | 0.141 | 0.879 | 31.14 |
| HBQ (M=4) | 离散型 | 16 | 16 | 4 | 4 | 384 | 163.6 | 0.148 | 0.872 | 30.40 |
| HBQ (M=6) | 离散型 | 16 | 16 | 4 | 6 | 256 | 148.8 | 0.142 | 0.878 | 30.98 |
| HBQ (M=8) | 离散 | 16 | 16 | 4 | 8 | 192 | 144.9 | 0.141 | 0.879 | 31.10 |
| HBQ(无量化) | 连续 | 64 | 16 | 4 | 16 | 24 | 43.2 | 0.078 | 0.935 | 34.79 |
| HBQ (M=4) | 离散 | 64 | 16 | 4 | 4 | 96 | 50.6 | 0.084 | 0.930 | 33.97 |
| HBQ* (M=4) | 离散 | 64 | 16 | 4 | 4 | 96 | 26.3 | 0.068 | 0.938 | 34.73 |
4 实验
4.1 视觉分词器
实现。我们引入了两种视觉分词器:一种专为类别条件图像生成设计的纯图像分词器,以及一种为文生图和文生视频任务设计的联合图像-视频分词器。两种分词器均采用 Wan 2.1 [Wan] 中的 3D 因果编码器和解码器架构,并从零开始训练。具体来说,纯图像分词器在 OpenImages 数据集 [openimages] 上训练,而联合分词器则在公开可用的图像和视频数据集的组合上训练。训练过程中,整体目标包含重建损失、感知损失和对抗(GAN)损失。纯图像分词器的各项损失权重分别设为 1.0、1.0 和 0.3,联合图像-视频分词器的权重则设为 1.0、0.2 和 0.005。
结果。如表 1 所示,我们的分词器在 256×256 ImageNet 基准测试上达到了最先进的重建性能。利用四轮 HBQ,它实现了显著的 0.56 rFID。这一结果不仅大幅超越了在 4 倍更高压缩率下运行的连续 SD-VAE(0.87 rFID),也显著优于其他领先方法,包括 RAE(0.62)、VAR(0.85)、LlamaGen(2.19)和 Open-MAGVIT-v2(1.17)。这些结果凸显了我们的方法在严格压缩条件下实现高保真重建的卓越能力。
在表 2 中,我们展示了一系列联合图像-视频模型 tokenizer,探讨了不同 HBQ 轮次和潜在通道维度的影响。我们首先分析了 HBQ 轮次的效果,并观察到一个明显趋势:随着 HBQ 轮次增加,rFVD 和 PSNR 等重建指标持续改善。虽然四到六轮已能取得强劲性能,但八轮配置的重建质量几乎与连续基线持平,如表 2 和图 5 所示。这表明,我们的 HBQ tokenizer 能够在实现更高压缩率的同时,匹配其连续对应版本的保真度。关键在于,这一成果是在不增加潜在通道的情况下实现的。虽然其他方法(例如 Infinity [hanjInfinity]、BitDance [bitdance])也能缩小与连续模型的差距,但它们通常依赖于扩展潜在维度。正如近期研究 [stable-diffusion3、hanjInfinity、dcae1p5] 所指出的,这种方法往往会减慢收敛速度,并需要更大的模型。相比之下,我们的主要成果是在不增加潜在维度的情况下取得的。
我们还尝试将潜在通道从 16 扩展到 64。仅此一项改动就将 PSNR 从 30.40 提升到了令人瞩目的 33.97。值得注意的是,这一性能与最先进的 Wan 2.1 tokenizer 相当,但压缩率却高出 4 倍。通过仔细调整 GAN 损失权重并增加训练迭代次数,HBQ tokenizer 在 SSIM 和 PSNR 指标上均超越了 Wan 2.1 和 Wan 2.2。有关调优的更多细节,请参见附录 E.1。
4.2 类别到图像的结果
实现。遵循 JiT [jitpaper],我们在原始 Transformer 架构中引入了 SwiGLU、RMSNorm、RoPE、qk-norm 以及上下文内类别条件控制。我们以四种不同模型规模训练 GRNind:130M、458M、952M 和 2B,分别记为 GRN-B、GRN-L、GRN-H、GRN-G。我们在 ImageNet [imagenet] 数据集上对它们进行了 600 个 epoch 的训练。学习率设置为 2e-4,并在训练过程中保持不变。我们随机丢弃 10% 的条件用于无分类器引导。在推理阶段,我们通过网格搜索寻找最佳解码超参数。其他实现细节见附录 4。
| 类型 | 模型 | 分词器 | 损失 | 参数量 | Gflops | FID | IS |
| 扩散 | DiT-L/2 [dit] | C | MSE | 458M | - | 5.02 | 167.2 |
| 扩散 | DiT-XL/2 [dit] | C | MSE | 675M | 119 | 2.27 | 278.2 |
| 流 | SiT-XL/2 [sit] | C | MSE | 675M | 119 | 2.06 | 277.5 |
| 流 | REPA [repa], SiT-XL/2 | C | MSE | 675M | 119 | 1.42 | 305.7 |
| 流 | RAE [RAE], DiT-XL/2 | C | MSE | 839M | 146 | 1.13 | 262.6 |
| 流 | JiT-B/16 [jitpaper] | C | MSE | 131M | 25 | 3.66 | 275.1 |
| 流 | JiT-L/16 [jitpaper] | C | MSE | 459M | 88 | 2.36 | 298.5 |
| 流 | JiT-H/16 [jitpaper] | C | MSE | 953M | 182 | 1.86 | 303.4 |
| 流 | JiT-G/16 [jitpaper] | C | MSE | 2B | 383 | 1.82 | 292.6 |
| 混合 | MAR [MAR] | C | MSE | 943M | - | 1.55 | 303.7 |
| 混合 | BitDance-H-1x [bitdance] | D | MSE | 1B | - | 1.24 | 304.4 |
| AR | LlamaGen-L [llamagen] | D | CE | 343M | - | 3.07 | 256.1 |
| AR | LlamaGen-XL [llamagen] | D | CE | 775M | - | 2.62 | 244.1 |
| AR | LlamaGen-XXL [llamagen] | D | CE | 1.4B | - | 2.34 | 253.9 |
| AR | MaskGIT [maskgit] | D | CE | 227M | - | 6.18 | 182.1 |
| AR | VAR-d20 [keyuVAR] | D | CE | 600M | - | 2.57 | 302.6 |
| AR | VAR-d24 [keyuVAR] | D | CE | 1B | - | 2.09 | 312.9 |
| AR | VAR-d30 [keyuVAR] | D | CE | 2B | - | 1.92 | 323.1 |
| AR | RandAR-XXL [pang2025randar] | D | CE | 1.4B | - | 2.15 | 322.0 |
| AR | GRN-B | D | CE | 130M | 25 | 3.56 | 280.3 |
| AR | GRN-L | D | CE | 458M | 88 | 2.64 | 314.8 |
| AR | GRN-H | D | CE | 952M | 182 | 2.06 | 316.1 |
| AR | GRN-G | D | CE | 2B | 383 | 1.81 | 299.0 |
结果。如表 3 所示,GRN 在 ImageNet 256×256 类别条件生成任务上与最先进的扩散模型、混合模型和自回归模型进行了基准对比。我们的 GRN-B 模型参数量仅为 MaskGIT 的近一半,却实现了更优的 FID 分数 3.56(对比 6.18),展现出卓越的效率。我们最大的变体 GRN-G 取得了最先进的 FID 分数 1.81,可与顶尖的扩散模型和混合模型相媲美。值得注意的是,GRN-G 在 FID 和 Inception Score 两项指标上均超越了 DiT 和 SiT 等基础模型。这一点意义重大,因为它们是当前许多工业级文生图和文生视频模型的骨干架构。此外,GRN-G 还优于自回归模型 LlamaGen 和 VAR。我们将这一优势归功于我们提出的全局细化生成方法,该方法有效缓解了误差传播问题。图 1 中展示的 GRN-G 生成的高质量视觉样本也证实了其能力。更多未经筛选的定性结果请参阅附录中的图 4。这些强劲的结果确立了 GRN 作为高保真视觉生成领域一个强大且可扩展的基线模型,激励着将其应用于更复杂的文生图和文生视频任务。
4.3 文生图结果
实现。我们从头开始训练 GRNbit 用于文生图任务,参数量为 2B。与类别到图像模型不同,文生图模型利用上下文自注意力机制而非 adaln-zero 来注入条件。该模型先在大型公开数据集上进行预训练,随后在一个小型高质量专有数据集上进行微调。我们首先在预训练数据集上以 256×256 分辨率训练 GRN,共 150K 次迭代,批次大小约为 15,400,学习率为 2e-4。然后,我们使用一个更小的高质量数据集,在 1024 分辨率下对 GRN 进行微调。在此阶段,我们以 2048 的批次大小和 2e-5 的学习率训练 GRN 共 60K 次迭代。更多实现细节见附录第 4 节。此外,我们观察到 GRN 在生成小脸等精细细节方面显著优于 Infinity。
结果。如表 4 所示,我们的模型在配备重写器后,在 GenEval 基准测试 [ghosh2024geneval] 上取得了 0.76 的总体得分。虽然我们的模型在性能上不及 Z-Image-Turbo [cai2025z]、HiDream [cai2025hidream]、Qwen-Image [qwenimage2025report] 和 BitDance [ai2026bitdance] 等更大规模的方法,但必须指出模型规模上的显著差异:这些模型使用了 6B 到 20B 的参数,而 GRN 则是一个紧凑得多的 2B 模型。在与同等规模的模型比较时,GRN 展现出更优的性能,显著优于类似 2B 大小的模型,如 SD3 Medium [stable-diffusion3](0.62)和 Infinity [hanjInfinity](0.71)。附录中图 5 的定性结果展示了 GRN 在生成高保真度、多样化且准确遵循用户提示词的图像方面的强大能力。
| 模型 | 参数量 | 数据量 | 单个物体 | 两个物体 | 计数 | 颜色 | 位置 | 颜色属性 | 总体 |
|---|---|---|---|---|---|---|---|---|---|
| 专有模型 | |||||||||
| GPT Image 1 [openai2025gpt4oimage] | 无 | 无 | 0.99 | 0.92 | 0.85 | 0.92 | 0.75 | 0.61 | 0.84 |
| Seedream 3.0 [gao2025seedream] | 无 | 无 | 0.99 | 0.96 | 0.91 | 0.93 | 0.47 | 0.80 | 0.84 |
| 扩散模型 | |||||||||
| PixArt-[chen2023pixart] | 0.6B | 无 | 0.98 | 0.50 | 0.44 | 0.80 | 0.08 | 0.07 | 0.48 |
| SD3 Medium [stable-diffusion3] | 2B | 无 | 0.98 | 0.74 | 0.63 | 0.67 | 0.34 | 0.36 | 0.62 |
| JanusFlow [janusflow] | 1.3B | 无 | 0.97 | 0.59 | 0.45 | 0.83 | 0.53 | 0.42 | 0.63 |
| FLUX.1-Dev [FLUX] | 12B | 无 | 0.98 | 0.81 | 0.74 | 0.79 | 0.22 | 0.45 | 0.66 |
| SD3.5-Large [stable-diffusion3] | 8B | 无 | 0.98 | 0.89 | 0.73 | 0.83 | 0.34 | 0.47 | 0.71 |
| Lumina-Image-2.0 [qin2025lumina] | 2.6B | 111M | - | 0.87 | 0.67 | - | - | 0.62 | 0.73 |
| Show-o2 [xie2025show] | 无 | 无 | 1.00 | 0.87 | 0.58 | 0.92 | 0.52 | 0.62 | 0.76 |
| Z-Image-Turbo [cai2025z] | 6B | 无 | 1.00 | 0.95 | 0.77 | 0.89 | 0.65 | 0.68 | 0.82 |
| HiDream-I1-Full [cai2025hidream] | 17B | 无 | 1.00 | 0.98 | 0.79 | 0.91 | 0.60 | 0.72 | 0.83 |
| Z-Image [cai2025z] | 6B | 无 | 1.00 | 0.94 | 0.78 | 0.93 | 0.62 | 0.77 | 0.84 |
| Qwen-Image [qwenimage2025report] | 20B | 无 | 0.99 | 0.92 | 0.89 | 0.88 | 0.76 | 0.77 | 0.87 |
| BAGEL [deng2025emerging] | 14B | 无 | 0.98 | 0.95 | 0.84 | 0.95 | 0.78 | 0.77 | 0.88 |
| 自回归模型 | |||||||||
| Emu3-Gen [wang2024emu3] | 8B | 无 | 0.98 | 0.71 | 0.34 | 0.81 | 0.17 | 0.21 | 0.54 |
| Infinity†[hanjInfinity] | 2B | 160M | - | 0.85 | - | - | 0.49 | 0.57 | 0.73 |
| Janus-Pro [chen2025janus] | 无 | 无 | 0.99 | 0.89 | 0.59 | 0.90 | 0.79 | 0.66 | 0.80 |
| Tar [han2025vision] | 无 | 无 | 0.98 | 0.92 | 0.83 | 0.85 | 0.80 | 0.65 | 0.84 |
| NextStep-1 [team2025nextstep] | 14B | 无 | - | - | - | - | - | - | 0.73 |
| BitDance [ai2026bitdance] | 14B | 无 | 1.00 | 0.96 | 0.71 | 0.95 | 0.72 | 0.83 | 0.86 |
| GRN † | 2B | 80M | 0.99 | 0.90 | 0.72 | 0.84 | 0.52 | 0.60 | 0.76 |
| 模型 | 参数量 | 人类 | 场景 | 多重 | 外观 | 质量 | 语义 | 总体 |
|---|---|---|---|---|---|---|---|---|
| 动作 | 物体 | 风格 | 得分 | 得分 | ||||
| 扩散 / 流模型 | ||||||||
| AnimateDiff-V2 [animatediff] | 1.5B | 92.60 | 50.19 | 36.88 | 22.42 | 82.90 | 69.75 | 80.27 |
| VideoCrafter-2.0 [videocrafter] | 1.5B | 95.00 | 55.29 | 40.66 | 25.13 | 82.20 | 73.42 | 80.44 |
| OpenSora V1.2 [opensora] | 1.1B | 85.80 | 42.47 | 58.41 | 23.89 | 80.71 | 73.30 | 79.23 |
| Show-1 [show-1] | 6B | 95.60 | 47.03 | 45.47 | 23.06 | 80.42 | 72.98 | 78.93 |
| URSA [ursa] | 1.7B | - | 52.30 | 70.60 | - | 83.40 | 78.50 | 82.40 |
| CogVideoX-5B [cogvideox] | 5B | 99.40 | 53.20 | 62.11 | 24.91 | 82.75 | 77.04 | 81.61 |
| HunyuanVideo [hunyuanvideo] | 13B | 94.40 | 53.88 | 68.55 | 19.80 | 85.09 | 75.82 | 83.24 |
| Wan 2.1 [Wan] | 14B | 98.80 | 53.67 | 81.44 | 21.13 | 85.64 | 80.95 | 84.70 |
| 自回归模型 | ||||||||
| Nova† [nova] | 0.6B | 95.20 | 54.06 | 77.52 | 20.92 | 80.39 | 79.05 | 80.12 |
| Emu3 [wang2024emu3] | 8B | 77.71 | 37.11 | 44.64 | 20.92 | 84.09 | 68.43 | 80.96 |
| Lumos-1 [yuanlumos] | 3.6B | - | - | - | - | 79.50 | 73.50 | 78.30 |
| InfinityStar† [infinitystar] | 8B | 96.43 | 52.08 | 78.66 | 21.81 | 84.73 | 79.78 | 83.74 |
| GRN† | 2B | 93.75 | 50.44 | 70.83 | 21.30 | 84.41 | 77.35 | 82.99 |
4.4 文生视频结果
实现细节。除了类别到图像和文生图生成任务之外,我们将 GRN 扩展到了最具挑战性的文生视频合成任务。GRN 的 T2V 变体与其 T2I 变体共享相同的架构,但仅在视频数据上进行训练。为此,我们整理了一个包含约 4000 万个视频片段的训练数据集,每个片段的分辨率至少为 256×256,时长为 2 到 10 秒。训练过程包括两个阶段。首先,我们在 192p 分辨率下以 4096 的批次大小和 2e-4 的学习率训练 GRN 150K 次迭代。随后,我们切换到 480p 分辨率进行微调,以 1350 的较小批次大小和 2e-5 的学习率额外训练 9K 次迭代。更多实现细节见附录 4。
实验结果。如表5所示,GRN在根据文本提示生成视频方面展现出卓越性能。与当前基于扩散和流的模型(包括AnimateDiff-V2、VideoCrafter-2.0、OpenSora V1.2、Show-1和CogVideoX-5B)进行基准对比时,GRN在质量分数、语义分数和总体分数上均取得了显著更高的得分。值得注意的是,尽管GRN仅有2B参数,却超越了规模大得多的CogVideoX-5B模型,凸显了其卓越的参数效率。此外,我们的方法也优于同等规模的离散扩散模型URSA。与自回归模型(如Nova、Emu3和Lumos-1)相比,GRN的性能优势更为明显。虽然8B参数模型InfinityStar目前以83.74的总体分数保持领先,但我们相信通过扩大GRN的规模可以缩小这一性能差距。我们在图6中展示了定性结果。更多结果请参见附录中的图6和图7。生成的视频不仅准确捕捉了用户提示的语义细节,还保持了较高的美学和视觉质量。
4.5 消融实验
4.5.1 预测索引 vs. 预测比特
| CFG | 区间 | FID | IS | ||
|---|---|---|---|---|---|
| GRNind-B | 2.4 | [0.4,1] | 1.33 | 3.56 | 280.3 |
| GRNbit-B | 2.4 | [0.44,1] | 1.23 | 3.63 | 285.5 |
| GRNind-L | 2.0 | [0.40,1] | 1.30 | 2.64 | 314.8 |
| GRNbit-L | 1.9 | [0.45,1] | 1.20 | 2.47 | 287.0 |
如第 3 节所述,GRN 支持预测离散索引(GRNind)或其二进制表示(GRNbit)。我们在 256×256 类别条件图像生成任务上,使用两种模型规模(GRN-B(130M)和 GRN-L(458M))对这两种预测目标进行了比较。对于每种变体,我们进行了网格搜索以确定最优解码参数,包括 CFG、CFG 区间和温度。结果如表 6 所示,表明两种方法均取得了相当的性能。具体而言,对于较小的 GRN-B 模型,预测索引获得了略优的 FID 分数。相反,对于较大的 GRN-L 模型,预测比特被证明更优,其 FID 达到 2.47,优于 2.64。这表明 GRN 在类别到图像生成任务上非常适合这两种预测格式。
我们进一步将此比较扩展到更具挑战性的 T2V 生成任务。如图 7 所示,我们观察到比特预测方法生成的视频质量更好,伪影更少。我们推测这是因为预测比特提供了更明确的监督信号,并减轻了索引预测中固有的 token 混叠效应,从而在复杂生成任务上展现出更优的性能。虽然一些先前的工作认为比特预测假设比特之间相互独立,会导致次优结果,但我们的全局精炼机制有效解决了这一问题。
4.5.2 全局精炼机制
| CFG | 区间 | FID | IS | ||
|---|---|---|---|---|---|
| 精炼 | 2.4 | [0.44,1] | 1.23 | 3.63 | 285.5 |
| 掩码 | 2.4 | [0.44,1] | 1.23 | 185.62 | 4.3 |
| 掩码 | 8.0 | [0.00,1] | 0.50 | 18.13 | 220.2 |
在消融研究中,我们验证了全局精化机制(称为 Refine)的有效性。我们将其性能与传统的基于掩码的生成流程(如 MaskGIT 或 BERT)进行了对比,后者会固定已生成的 token。结果令人瞩目:使用相同的解码超参数时,基于掩码的方法会崩溃并生成无意义的输出(FID=185.62),详见 Tab. 7。即使通过网格搜索找到最优解码参数(更高的 CFG、更低的温度),基于掩码的方法(FID=18.13)仍显著落后于我们的方法(FID=3.63)。该实验清楚地表明,我们的精化范式有效缓解了标准自回归模型中的关键弱点——错误传播,从而实现了更优的性能。
4.5.3 复杂度感知采样
我们在 GRNbit-B 上评估了复杂度感知采样的有效性,超参数设置为 。遵循扩散模型的标准设置,我们将最大精炼步数设为 。为了在性能与效率之间取得平衡,我们根据经验将最小步数设为 。我们合成了 63,000 张图像,并将其分配到的生成步数分布绘制成直方图(图 9)。结果表明,我们提出的方法使 GRN 能够根据不同的复杂度水平动态分配计算资源。如图所示,不同样本被分配了不同的精炼步数,范围从 20 到 50 步。超过 97.9% 的样本所需的精炼步数少于 50 步。有趣的是,大约 6.3% 的图像仅用最少 20 步生成,这表明模型对这些特定预测具有很高的置信度。与固定步数基线(所有样本均为 50 步)相比,我们的复杂度感知采样将 FID 从 3.56 降低至 3.47,同时实现了显著的加速(最高 2.5 倍,平均 1.25 倍)。此外,我们将复杂度感知采样应用于文生图任务,并设置 。图 8 中的定性结果直观地证实了我们的方法有效地实现了复杂度感知的自适应步数生成。未来工作方面,我们计划探索精炼步数蒸馏,该方法与复杂度感知采样天然兼容,并能实现更高效的视觉生成。
4.5.4 训练与推理成本
如表 3 所示,我们报告了 C2I 中一次前向传播的 Gflops。训练成本与一次前向传播相当,而推理成本随步数增加而增加。对于 T2I,GRN-2B 生成一张 1024×1024 图像需要 2-8 秒,平均 4 秒(对比 Infinity 1 秒、FluxDev 12 秒、Qwen-Image 33 秒)。对于 T2V,GRN 生成一段 480p、81 帧的视频需要 125 秒(对比 Wan2.1 480 秒、Wan2.2 437 秒)。请注意,这些基线模型使用了 Diffusers 的优化,而 GRN 在纯 PyTorch eager 模式下运行。我们使用相同的设备测试了这些结果。
我们还对 GRNbit 的比特预测目标和解码参数进行了消融研究。为简洁起见,更多细节请参考附录 E.4 和附录 E.5。
5 结论
在本文中,我们提出了 GRN,这是一个以全局精化机制和复杂度感知生成为特征的下一代视觉合成框架。我们提出了分层二值量化(HBQ),用于开发一系列离散图像和视频分词器,这些分词器在使用相同数量潜在通道并提供显著更高压缩率的同时,性能与连续分词器相当。在生成阶段,GRN 在图像重建和类别条件图像生成方面均取得了新的最先进成果。大量实验表明,在同等规模下,GRN 在文生图和文生视频任务中均超越了现有的自回归和基于扩散的方法。
此外,作为一个完全构建在离散 token 之上的自回归框架,我们相信 GRN 能够更自然地集成到现有的大语言模型中。对离散文本和视觉 token 的统一学习可以极大地促进多模态理解与生成。从根本上说,GRN 解决了长期限制以往视觉自回归生成模型的量化损失和误差累积问题。我们相信它有潜力成为当前占主导地位的 Transfusion [transfusion] 架构的有力竞争者。
6 局限性
这项工作也存在若干局限性。由于计算资源有限,我们尚未将训练计算量或模型规模扩展到领先视觉生成模型的水平。此外,对于文生视频任务,我们观察到 GRN 在与人类相关的场景中表现更好。生成的视频有时可能缺乏丰富的视觉细节,并出现失真。我们相信,通过平衡数据分布和扩大模型规模,这些局限性可以得到缓解。
7 致谢
我们感谢陆瑞彪在数据收集和视频演示方面的贡献,以及吴辉在基础设施方面提供的宝贵建议。
参考文献
附录 A 分层二值量化算法
我们在算法1中概述了所提出的分层二元量化(HBQ)的流程。HBQ 二元 token 按从粗到细的顺序排列。序列中靠前的 token 对应核心语义概念,而靠后的 token 则引入高频细节,逐步丰富表征。此外,我们的多轮量化过程会顺序生成二元 token,每一轮对应一个比特位。这种内在结构使得 HBQ 特别适合直接的逐比特预测任务。与 FSQ [fsq] 等整体量化向量的方法不同,我们的方法为逐比特生成(即 GRNbit)提供了更自然的框架。
附录 B 生成式精炼网络算法
算法2 概述了单次训练步骤的伪代码。模型输入(记为 )是一个混合特征图,由一部分真实 token 和一部分随机 token 组成。以 为输入,GRN 被训练来预测完整的真实 token 集合。尽管方法简单,但这种训练策略隐式地教会了模型区分可靠(真实)和不可靠(随机)的 token。因此,模型学会了在保留可靠 token 的同时,对不可靠的 token 进行精炼。
GRN 的采样过程在算法 3 中详细说明(为清晰起见,我们省略了复杂度感知采样)。该过程类似于人类绘画,即对状态进行迭代优化。在每一步,当前状态由已绘制的内容(由 `表示`)和填充了随机 token 的剩余空白区域(由 `表示`)组成。然后,模型基于当前状态预测一组完整的 token。这些新预测中随机选择的一个子集随后被用于更新状态,以进行下一步优化。这种直接的随机选择机制巧妙地将三个基本操作统一到一个框架中:
-
填充:将预测的 token 引入先前空白的区域。
-
优化:提高先前预测 token 的质量。
-
擦除:用随机 token 替换先前预测的 token。
为了更好地说明这一过程,论文中的图 4 以及附录中的图 10 的第四列都可视化了 GRN 的采样过程。我们特别强调了上述填充、优化和擦除 token 的具体示例。
附录 C 与其他自回归模型的区别
在图 10 中,我们将 GRN 与传统的自回归模型进行了比较,包括 GPT 风格的自回归模型(下一个 token 预测)[llamagen]、VAR(下一尺度预测)[keyuVAR] 以及掩码自回归模型(BERT 风格)[maskgit]。传统自回归模型受限于固定的生成顺序,先前生成的 token 不可更改,这常常导致错误传播等问题。与此形成鲜明对比的是,GRN 采用了灵活的全局优化策略,利用其独特的填充、优化和擦除机制。这使得我们的模型能够迭代地修正和增强整个视觉表征,有效缓解了传统自回归模型中固有的错误传播问题。
附录 D 实现细节
模型架构。表 8 总结了我们提出的 C2I、T2I 和 T2V 模型的架构细节。遵循 JiT [jitpaper] 的方法,我们为 C2I 任务实现了四种变体,模型规模从 130M 参数扩展到 2B 参数。对于 T2I 和 T2V 生成,我们引入了一种新的 2B 参数架构,该架构专门设计用于满足 FlexAttention 的要求,即注意力头维度必须是 128 的倍数。此外,我们的模型支持序列打包以加速训练,并采用 NaViT [navit] 来处理任意宽高比和分辨率。
视觉分词器。对于C2I生成任务,我们采用一个仅在OpenImages数据集[openimages]上训练的纯图像视觉分词器。该分词器具有16维潜在空间,并使用了4轮HBQ。它将256×256的图像压缩为16×16×16×4的二进制token,空间步长为16,在ImageNet基准测试上实现了0.56的最先进重建FID。相比之下,我们的T2I和T2V模型共享一个统一的视觉分词器,该分词器在图像和视频的混合数据上联合优化。这个统一分词器配置了64维潜在空间、空间步长16、时间步长4,同样经过4轮HBQ,在我们的视频重建基准测试上达到了26.3的FVD和34.73的PSNR。已发布分词器的更多结果见论文中的表1和表2,以及附录中的表9。
训练。对于我们的C2I模型(GRN-B/L/H/G),我们使用256×256分辨率的图像在ImageNet数据集上进行训练。模型训练了600个epoch,相当于75万次迭代,批大小为1024。整个训练过程中我们采用恒定的2e-4学习率。我们应用了10%的条件丢弃率以实现无分类器引导。对于我们的GRN-T2I和GRN-T2V模型,我们采用了从粗到精的训练策略。例如,GRN-T2I模型在256×256分辨率下训练了15万次迭代,在1024×1024分辨率下训练了6万次迭代,批大小分别为15400和2048,并使用相应的2e-4和2e-5学习率。其他超参数,如恒定学习率调度、零权重衰减和10%条件丢弃,在这些模型中保持一致。
采样。在推理阶段,我们采用无分类器引导来提升样本质量并增强对条件信息的遵循。对于在 ImageNet 上训练的模型,我们根据在 [0, 0.5] 范围内找到的最优阈值动态启动 CFG。CFG 强度在 [1.0, 3.0] 范围内扫描以寻找最佳结果。对于文本条件模型(GRN-T2I 和 GRN-T2V),我们在整个采样过程中都应用 CFG。我们在 [1.0, 4.0] 范围内搜索最优 CFG 强度,并在 [0.5, 1.5] 的温度范围内生成多样且高保真的结果。对于 C2I、T2I 和 T2V 任务的基准实验,我们使用 50 个固定的细化步骤。在第 4.5.3 节中,我们将复杂度感知采样应用于 C2I 和 T2I 任务。所提出的复杂度感知采样方法由参数 和 控制,其中 在 [300, 1200] 范围内优化,而 则相应设置以确定总采样步数。
| GRN-B | GRN-L | GRN-H | GRN-G | GRN-T2I | GRN-T2V | |
| 架构 | ||||||
| 深度 | 12 | 24 | 32 | 40 | 28 | 28 |
| 隐藏维度 | 768 | 1024 | 1280 | 1664 | 2304 | 2304 |
| 注意力头数 | 12 | 16 | 16 | 16 | 18 | 18 |
| 参数量 | 130M | 458M | 952M | 2B | 2B | 2B |
| 数据 | ImageNet 256×256 图像 | 图像 | 视频 | |||
| 分词器 | ||||||
| 分词器步长 | 16×16 | 16×16×4 | ||||
| 分词器维度 | 16 | 64 | ||||
| HBQ 轮次 | 4 | 4 | ||||
| 训练 | ||||||
| 轮数 | 600 | 不适用 | ||||
| 迭代次数 | 750K | 150K/60K | 150K/9K | |||
| 批量大小 | 1024 | 15400/2048 | 4096/1350 | |||
| 学习率 | 2e-4 | 2e-4/2e-5 | 2e-4/2e-5 | |||
| 学习率调度 | 常数 | 常数 | ||||
| 权重衰减 | 0 | 0 | ||||
| 条件丢弃率 | 10% | 10% | ||||
| 采样 | ||||||
| , | 在 [300,1200] 内最优, 相应设置 | |||||
| 扫描范围 | [1.0,1.5] | [0.5,1.5] | ||||
| CFG 扫描范围 | [1.0, 3.0] | [1.0,4.0] | ||||
| CFG 间隔 | 在 [0, 0.5] 内最优以启动 CFG | 未使用 | ||||
附录 E 更多消融实验
E.1 分词器中的 GAN 损失
图 5 展示了一个清晰的趋势:随着引入更多 HBQ 轮次,重建质量稳步提升。尽管四到六轮就足以获得不错的性能,但八轮模型缩小了差距,取得了与未量化基线几乎相同的结果。
我们进一步研究了改变 GAN 损失权重的影响。为此,我们使用不同的 值微调了一个基线模型(一个具有 64 个潜在通道和四轮量化的 HBQ 分词器)。如表 9 所示,将 GAN 损失权重从 0.001 增加到 0.02 显著提升了感知质量,将 rFVD 从 48.6 降低到 28.6。然而,这是以重建保真度下降为代价的,PSNR 从 34.05 下降到 33.73。根据经验,我们发现权重为 0.005 时,这两个指标之间达到了有效的平衡。因此,我们将此设置应用于我们的 GRN-T2I 和 GRN-T2V 模型。
| 方法 | 通道 | 步长 | 压缩率 | rFVD | LPIPS | SSIM | PSNR | |
| Wan 2.1 (patchify) | 不适用 | 64 | 16164 | 24 | 19.5 | 0.058 | 0.929 | 34.10 |
| HBQ (基线) | 0.001 | 64 | 16164 | 96 | 50.6 | 0.084 | 0.930 | 33.97 |
| HBQ () | 0.001 | 64 | 16164 | 96 | 48.6 | 0.083 | 0.930 | 34.05 |
| HBQ () | 0.005 | 64 | 16164 | 96 | 30.0 | 0.078 | 0.928 | 33.98 |
| HBQ () | 0.02 | 64 | 16164 | 96 | 28.6 | 0.081 | 0.925 | 33.73 |
| HBQ () | 0.1 | 64 | 16164 | 96 | 29.5 | 0.084 | 0.923 | 33.55 |
| 重建 | GRN-B(130M) | GRN-L(458M) | ||||
|---|---|---|---|---|---|---|
| rFID | PSNR | gFID | IS | gFID | IS | |
| HBQ-2 | 1.23 | 21.24 | 3.22 | 304.1 | 2.71 | 309.4 |
| HBQ-4 | 0.54 | 22.81 | 3.54 | 293.0 | 2.63 | 290.1 |
| HBQ-6 | 0.49 | 23.06 | 4.74 | 235.2 | 3.10 | 253.6 |
| HBQ-8 | 0.48 | 23.09 | 5.29 | 208.8 | 3.48 | 235.2 |
E.2 HBQ 轮数对生成性能的影响
如表 10 所示,尽管 HBQ (M=2) 对于具有 130M 参数的 GRN-B 表现最佳,但当扩展到 GRN-L(458M 参数)时,HBQ (M=4) 超越了 HBQ (M=2)。我们得出结论,大模型受益于更多的 HBQ 轮数。另一个有趣的观察是,尽管从 6 轮到 8 轮,重建指标几乎保持不变,但生成指标却存在显著差异。这表明当信息分布在更多不同的比特上时,确实需要更大的模型容量。此外,我们的实验揭示,MSB 的准确率远高于 LSB。未来在 HBQ (16dim, M=8) 和 HBQ (32dim, M=4) 之间的比较研究将构成一条引人注目的实验路线。
E.3 二值选择映射:随机 vs. 置信度
| 采样 | 指标 | |||
|---|---|---|---|---|
| 随机 | 置信度 | FID | IS | |
| GRNbit-B | ✓ | 3.63 | 285.5 | |
| GRNbit-B | ✓ | 10.64 | 246.8 | |
如公式 6 所述,二值选择图在生成过程中没有先验约束的情况下构建,这意味着我们随机选择当前预测来更新状态,以进行下一步的细化。为了理解这种随机采样的重要性,我们使用基于置信度的采样替代方案进行了一项实验。在这种设置中,待更新的 token 是根据其预测置信度来选择的。虽然这优先考虑了被认为“更正确”的 token,但结果却是性能严重下降(FID:3.63 对 10.64),详见表 11。这种反直觉结果的原因在于训练与推理模式之间的差异。我们的模型被训练为在真实标签与随机 token 均匀分布的状态下运行。基于置信度的方法通过选择高置信度的 token 打破了这一假设,这些 token 并非均匀分布,而是聚集在一起。这种分布偏移使得输入远离了训练期间学到的流形,导致生成过程灾难性失败。
E.4 比特预测目标:绝对值与相对值
我们研究了 GRNbit 中二值标签的两种不同预测目标:绝对比特与相对比特。绝对比特预测直接针对真实比特值,而相对比特预测则针对输入比特是否应该被翻转。这可以表述为预测一个残差,其中“1”表示需要翻转,“0”表示保持不变。如图 11 所示,我们的实验表明,预测绝对比特能产生更优的结果,与相对比特预测生成的图像相比,其生成的图像具有显著更好的结构稳定性。
E.5 解码超参数
| 方法 | FID | IS | ||||||
| 线性 | / | / | 50 | 50 | 50 | 0.0 | 3.56 | 290.3 |
| 余弦 | / | / | 50 | 50 | 50 | 0.0 | 3.94 | 272.5 |
| 复杂度 | 700 | -620 | 20 | 50 | 40 | 7.4 | 3.47 | 287.3 |
| 线性 | / | / | 30 | 30 | 30 | 0.0 | 5.06 | 262.3 |
| 余弦 | / | / | 30 | 30 | 30 | 0.0 | 7.87 | 224.5 |
| 复杂度 | 700 | -630 | 20 | 40 | 24 | 4.0 | 3.79 | 270.9 |
| 复杂度 | 600 | -527 | 20 | 50 | 39 | 6.8 | 3.50 | 286.6 |
| 600 | -527 | 10 | 50 | 39 | 8.2 | 3.53 | 282.1 | |
| 700 | -620 | 20 | 50 | 40 | 7.4 | 3.47 | 287.3 | |
| 700 | -630 | 20 | 40 | 24 | 4.0 | 3.79 | 270.9 | |
| 800 | -714 | 20 | 50 | 40 | 7.9 | 3.47 | 285.9 |
关于解码参数,我们以 GRNbit-B 作为代表性示例。最优参数经发现为 、 以及间隔为 。我们通过实验观察到,许多参数组合都能达到类似的结果。直观上,增大或减小 会鼓励更多样化的生成,但也可能导致更大的不稳定性。引入 CFG 间隔是为了在应用较高 CFG 强度时恢复多样性,因为它在初始解码步骤(这些步骤对于确定整体语义至关重要)中禁用了 CFG。在保持其他参数固定的情况下,改变每个参数的效果详见图 12。如表 12 所示,我们的复杂度感知采样在更少的步数内优于线性调度和余弦调度。在我们的复杂度感知采样中,主要参数是 和 。我们调整 以控制动态范围,调整 以控制平均步数。 和 将步数裁剪到合理范围内,需要较少的调整。表 12 显示,当 处于 [600, 800] 的合适范围内时,结果良好。
附录 F 更多定性结果
F.1 C2I 定性结果
与 JiT [jitpaper] 类似,我们在图 13 中展示了由 GRN-G 生成的未经筛选的 256x256 样本。为确保对我们模型能力的公平且具有代表性的可视化,这些图像使用了与实现最佳 FID 1.81 时相同的 CFG 尺度 1.7 和 CFG 间隔 = [0.3, 1.0] 生成。这与在定性示例中使用更高 CFG 尺度的常见做法形成对比,因为后者可能无法反映模型由 FID 衡量的真实性能。
F.2 T2I 定性结果
在图 14 中,我们展示了由 GRN-T2I 生成的 1024x1024 图像。
F.3 T2V 定性结果
在图 15 和图 16 中,我们展示了更多 GRN-T2V 的文本到视频生成结果。
{hanjian.thu123,liujinlai.licio}@bytedance.com
{wangjiahuan.123,bingyue.peng,yuanzehuan}@bytedance.com
Code and models:
yuanzehuan@bytedance.com
Abstract
While diffusion models dominate the field of visual generation, they remain computationally inefficient, as they allocate uniform computational effort to samples with varying levels of complexity. In contrast, autoregressive (AR) models are inherently complexity-aware, as evidenced by their variable likelihoods, but are often hindered by lossy discrete tokenization and error accumulation. In this work, we introduce Generative Refinement Networks (GRN), a next-generation visual synthesis paradigm that addresses these issues. At its core, GRN addresses the discrete tokenization bottleneck through a theoretically near-lossless Hierarchical Binary Quantization (HBQ), achieving a reconstruction quality comparable to continuous counterparts. Built upon HBQ’s latent space, GRN fundamentally upgrades AR generation with a global refinement mechanism that progressively perfects and corrects artworks — like a human artist painting. Besides, GRN integrates an entropy-guided sampling strategy, enabling complexity-aware, adaptive-step generation without compromising visual quality. On the ImageNet benchmark, GRN establishes new records in image reconstruction (0.56 rFID) and class-conditional image generation (1.81 gFID). We also scale GRN to more challenging text-to-image and text-to-video generation, delivering superior performance on an equivalent scale. We release all models and code to foster further research on GRN.
1 Introduction
The field of visual generation has advanced rapidly, driven primarily by scaling diffusion transformers [dit, sora, hunyuanvideo, waver, alive]. By progressively integrating trajectories along a learned velocity field that transports simple noise prior to the empirical data distribution, these models demonstrate strong capabilities in synthesizing high-quality visual content. However, this continuous flow paradigm inherently lacks adaptive-step capacity. Optimized via mean squared error (MSE) without explicit likelihoods, these models are restricted to a fixed number of steps, rigidly allocating identical computational resources to all samples regardless of varying levels of complexity.
Meanwhile, inspired by the success of token-level likelihood estimation in large language models[gpt3.5, gpt4], autoregressive (AR) models have also garnered extensive research interest in visual synthesis[videogpt, keyuVAR, hanjInfinity, wang2024emu3]. Nevertheless, current AR approaches are bottlenecked by two critical shortcomings. First, they intrinsically suffer from inferior reconstruction quality when utilizing discrete tokens as opposed to continuous representations. Second, their strictly causal prediction mechanism, whether they operate token-by-token or scale-by-scale, inevitably causes severe error accumulation over multi-step generation. This exposes a critical lack of error-correction capability, as the model cannot retroactively refine previous mistakes. Furthermore, even with parallel prediction in masked AR models [maskgit, bert], high-confidence tokens become immutable and cannot be revised later. Consequently, such models still inherently lack a holistic refinement mechanism.
These observations motivate a simple yet intuitive refinement-based AR framework with adaptive computation. To this end, we introduce Generative Refinement Networks (GRN), a next-generation visual synthesis paradigm designed to overcome the rigidly fixed computational costs of diffusion models and the inherent shortcomings of standard autoregressive models. Specifically, to address the inferior reconstruction of discrete tokens, we first propose Hierarchical Binary Quantization (HBQ). By ensuring an exponential decay of reconstruction error without increasing latent channels, HBQ empowers discrete image and video tokenizers to achieve near-lossless reconstruction, matching the performance of continuous tokenizers at a higher compression rate. Building upon these robust representations, GRN executes a complexity-aware adaptive-step generation process by employing an entropy-guided sampling mechanism. It dynamically distributes computational loads based on the varying difficulty of visual content, while employing a global refinement mechanism to retroactively mitigate accumulated errors.
Extensive experiments on diverse visual tasks validate the superiority of our framework. On the ImageNet 256256 benchmark [imagenet] for class-conditional image synthesis, GRN sets a new record for both image reconstruction and generation quality. Furthermore, demonstrating exceptional task generalization and scalability, we also successfully scale GRN to high-resolution text-to-image (T2I) and text-to-video (T2V) scenarios. When scaled up, GRN demonstrates the capability to generate photorealistic 10241024 images alongside dynamic, high-fidelity 480p videos ranging from 2 to 10 seconds. In summary, our main contributions are as follows:
-
We propose GRN, the next-generation visual synthesis framework. It is characterized by a global refinement mechanism and complexity-aware generation, achieving robust and efficient visual generation.
-
We introduce Hierarchical Binary Quantization and contribute a series of discrete image/video tokenizers. For the first time, discrete visual tokenizers are on par with continuous ones with the same latent dimensions.
-
Extensive experiments show that GRN achieves state-of-the-art results on standard C2I benchmarks, with an rFID of 0.56 and a gFID of 1.81. When scaled to more challenging T2I and T2V tasks, it demonstrates superior performance compared to methods at an equivalent scale.
2 Related Work
2.1 Visual Tokenizer
Visual tokenizers [ldm, vqvae, vqgan] compress visual content for efficient generation. Early vector quantization methods [vqvae, vqgan] map continuous features to a discrete codebook, but suffer from limited scalability, prompting lookup-free approaches [BSQ, fsq] to enable larger vocabularies. Despite this, a performance gap to continuous representations remains. Recent works [hanjInfinity, bitdance] aim to close this gap by drastically scaling vocabularies, outperforming continuous VAEs. However, this gain comes at the cost of slower convergence and larger generative models, motivating more efficient quantization schemes.
2.2 Autoregressive Models
Inspired by large language models, [vqgan, llamagen, videopoet, wang2025editinfinity] explore visual generation via next-token prediction. MaskGIT [maskgit] accelerates generation using parallel decoding, where it first generates high-confidence tokens and then iteratively fills in the remainder. VAR [keyuVAR] shifts autoregression to next-scale prediction, improving quality and achieving over faster inference. Nevertheless, AR models remain limited by lossy discrete tokenization and error accumulation, and still lag behind diffusion methods. Although Infinity [hanjInfinity] introduces self-correction by randomly flipping bitwise tokens, its assumption of less than 30% diffuse errors covers limited patterns.
2.3 Adaptive-step Generation
Diffusion models dominate visual generation [FLUX, sdxl, stable-diffusion3, sora, Wan], but typically require tens of inference steps. Distillation methods [dmd, dmd2] reduce sampling steps substantially, yet still rely on predefined schedules with fixed steps. This “one-size-fits-all” strategy wastes computational resources on simple prompts. Recently, AdaDiff [adadiff] employs an external network to determine instance-specific steps and uses a policy gradient method to maximize the reward. The sophisticated pipeline requires an additional network and reward signals.
3 Method
3.1 Visual Tokenizer
The visual tokenizer plays a vital role in learning a compact latent space to compress high-dimensional realistic data. We adopt the 3D causal VAE design proposed in Wan 2.1 [Wan] so that images and videos can be tokenized in a unified framework. Specifically, given an image or a video , the tokenizer encodes its spatio-temporal information into dimensions while expanding the number of channels to . Since our goal is lossless discrete compression, and VAE features are continuous signals, we frame feature quantization as a signal transformation problem. Inspired by Harr wavelet [haar1910theorie] in signal processing, we introduce Hierarchical Binary Quantization to transform VAE features into discrete ones.
Hierarchical Binary Quantization. We first append a non-linear activation function after the VAE encoder to map the feature representation from an unbounded range to the closed interval . As illustrated in Fig. 2, each element in undergoes several rounds of binary quantization based on a binary tree of buckets with center as defined in Eq. 1 and Eq. 2.
| (1) |
| (2) |
where is a delta function with -1 when and 1 otherwise. Then we obtain the quantized binary labels , where . Here is the total number of rounds of hierarchical binary quantization. In this way, we perform quantization from coarse to fine to represent information of different frequencies, and the quantization error for the round is less than . The upper bound of the quantization error decays exponentially with the number of rounds, theoretically enabling lossless quantization to be achieved rapidly. Fig. 3 shows the images reconstructed from the quantized intermediate results, revealing the coarse-to-fine property.
| (3) |
Subsequently, the quantized feature can be derived according to Eq. 3. The detailed algorithm for HBQ is provided in Appendix A. During the training phase of the visual tokenizer, the quantized feature is taken as input to the decoder to reconstruct the raw image or video . Following the common practice for training discrete visual tokenizers, we adopt the Straight-Through Estimator (STE) to backpropagate gradients to the encoder. The training loss is a weighted combination of the reconstruction loss (), LPIPS perceptual loss (), and the GAN loss () from a PatchGAN discriminator.
After tokenization, we obtain binary outputs with the size . However, it is nearly impossible to encode outputs into INT scalars to perform generation by merging and dimensions due to its equivalence to introducing a codebook with the large size . Inspired by bitwise tokens [hanjInfinity], we propose two variants of GRN, i.e. GRNind and GRNbit to support generation. For GRNind we simply encode the dimension to INT scalars, resulting in . For GRNbit we concatenate the last two dimensions and predict . For both variants, we flatten the spatiotemporal dimensions and predict the entire channel dimension in parallel for each token via multi-token prediction [deepseek_v3].
3.2 Generative Refinement Network
Inspired by the intuition of human drawing, we propose an elegantly simple autoregressive refinement framework for visual generation, which commences with a random token map. Let represent the state of the token map in step . The objective is to predict the drawing map , based on the current state . To explicitly formulate this process, we define as a composition of three components: a random map , a drawing map , and a binary selection map . The relationship is formally expressed in Eq. 4, where is constructed by selecting from or based on the values in .
| (4) |
Intuitively, represents the current drawing while corresponds to the blank area without any information, mimicking the intermediate step during the human drawing procedure. is designed to make the accumulation statistic , namely the proportion of ones in , increase monotonically from 0% to 100% during the refinement steps. Therefore, gradually converges to ideal token maps. In order to obtain , we employ a transformer and approximate as
| (5) |
to model the next-step drawing by fitting real token maps, where denotes the generation condition, such as class embeddings or texts. is constructed following Eq. 6 based on . An alternative approach based on prediction confidence was also investigated for constructing . However, it produced inferior results, as detailed in Appendix E.3.
| (6) |
We introduce a complexity-aware sampling strategy to control to not only maintain its monotonicity, but also consider the uncertainty of . We will discuss the details below. Based on the prediction and selection map , the state is updated accordingly as mentioned in Eq. 4. Thus, a coherent loop of progressive generation and refinement is formed. The autoregressive mechanism allows us not only to improve an increasing number of tokens with high certainty, but also to erase obvious errors with more context included as the drawing proceeds. Ideally, the process converges to the best result as more and more information accumulates. The toy process is demonstrated in Fig. 4.
Training. For each iteration in the training stage, we randomly sample random tokens from a uniform distribution for GRNind and for GRNbit. The binary map is also uniformly sampled with varying selection ratios that control how many real tokens are used as input. Therefore, the input to the transformer consists of tokens sampled from ground-truth tokens and tokens sampled from random tokens . Note that token sampling is randomly conducted along the spatial, temporal, and channel dimensions, with no additional priors (Eq. 6). refers to the total number of tokens, which equals . Taking constructed by Eq. 4 as input with partial information, our goal is to predict ground-truth tokens similar to x-prediction in diffusion setting using the simple Cross-Entropy loss as illustrated in Eq. 7. Here denotes the ground truth token.
| (7) |
The detailed training and inference process for GRN is provided in Appendix B. Additional comparisons between GRN and other autoregressive models are provided in Appendix C.
Complexity-Aware Sampling. We propose an entropy-guided scheduling function to determine , where refers to the index of the refinement step. In particular, we calculate the average entropy for step during generation as
| (8) |
In Eq. 8, we denote as the token index and as the category index, where is the total number of categories. Note that for GRNind and for GRNbit. Generation complexity is measured by the entropy , bounded between 0 and 1. Given that a smaller denotes greater predictive confidence, we allocate fewer refinement steps alongside a steeper increase in , thereby retaining more information from . Conversely, when high entropy indicates substantial complexity, we apply more refinement steps and a moderate progression of . Specifically, we formulate as
| (9) |
Here, represents the average entropy calculated from a specific step. We set a warm-up period with and , as we observed that the entropy values are unstable during the initial steps. The hyperparameters control the dynamic range of adaptive steps, and is the bias. We also clip the value of to ensure the total number of inference steps remains within the range of .
| Method | Tokenizer | Spatial | Latent | Channel | Compress | rFID | LPIPS | SSIM | PSNR |
|---|---|---|---|---|---|---|---|---|---|
| Type | Ratio | Channel | Bits | Ratio | |||||
| SD-VAE [ldm] | Continuous | 16 | 16 | 16 | 24 | 0.87 | - | 0.68 | 24.08 |
| RAE [RAE] | Continuous | 16 | 768 | 16 | 0.5 | 0.62 | 0.25 | 0.44 | 19.20 |
| VAR† [keyuVAR] | Discrete | 16 | N/A | N/A | 193 | 0.85 | 0.15 | 0.64 | 22.47 |
| LlamaGen [llamagen] | Discrete | 16 | 1 | 14 | 439 | 2.19 | - | 0.68 | 20.79 |
| Open-MAGVIT2 [open_magvit2] | Discrete | 16 | 1 | 18 | 341 | 1.17 | - | - | 22.64 |
| HBQ (M=4) | Discrete | 16 | 16 | 4 | 96 | 0.56 | 0.13 | 0.71 | 23.01 |
| Method | Tokenizer | Latent | Spatial | Temporal | Channel | Compress | rFVD | LPIPS | SSIM | PSNR |
| Type | Channel | Stride | Stride | Bits | Ratio | |||||
| Wan 2.1 | Continuous | 16 | 8 | 4 | 16 | 24 | 19.5 | 0.058 | 0.929 | 34.10 |
| Wan 2.2 | Continuous | 48 | 16 | 4 | 16 | 32 | 22.6 | 0.052 | 0.932 | 34.54 |
| HBQ (w/o quant) | Continuous | 16 | 16 | 4 | 16 | 96 | 144.6 | 0.141 | 0.879 | 31.14 |
| HBQ (M=4) | Discrete | 16 | 16 | 4 | 4 | 384 | 163.6 | 0.148 | 0.872 | 30.40 |
| HBQ (M=6) | Discrete | 16 | 16 | 4 | 6 | 256 | 148.8 | 0.142 | 0.878 | 30.98 |
| HBQ (M=8) | Discrete | 16 | 16 | 4 | 8 | 192 | 144.9 | 0.141 | 0.879 | 31.10 |
| HBQ (w/o quant) | Continuous | 64 | 16 | 4 | 16 | 24 | 43.2 | 0.078 | 0.935 | 34.79 |
| HBQ (M=4) | Discrete | 64 | 16 | 4 | 4 | 96 | 50.6 | 0.084 | 0.930 | 33.97 |
| HBQ* (M=4) | Discrete | 64 | 16 | 4 | 4 | 96 | 26.3 | 0.068 | 0.938 | 34.73 |
4 Experiments
4.1 Visual Tokenizer
Implementation. We introduce two visual tokenizers: an image-only tokenizer tailored for class-conditional image generation, and a joint image-video tokenizer designed for text-to-image and text-to-video generation tasks. Both tokenizers adopt the 3D causal encoder and decoder architecture from Wan 2.1 [Wan] and are trained from scratch. Specifically, the image-only tokenizer is trained on the OpenImages dataset [openimages], while the joint tokenizer is trained on a combination of publicly available image and video datasets. During training, the overall objective comprises reconstruction, perceptual, and adversarial (GAN) losses. The respective loss weights are set to 1.0, 1.0, and 0.3 for the image-only tokenizer, and 1.0, 0.2, and 0.005 for the joint image-video tokenizer.
Results. As shown in Tab. 1, our tokenizer achieves state-of-the-art reconstruction performance on the 256256 ImageNet benchmark. Utilizing four HBQ rounds, it achieves a remarkable rFID of 0.56. This result significantly surpasses not only the continuous SD-VAE (0.87 rFID) while operating at a 4 higher compression rate, but also substantially outperforms other leading methods, including RAE (0.62), VAR (0.85), LlamaGen (2.19), and Open-MAGVIT-v2 (1.17). These results underscore our method’s superior ability to achieve high fidelity reconstruction under stringent compression.
In Tab. 2, we present a series of joint image-video tokenizers, exploring the impact of varying HBQ rounds and latent channel dimensions. We first analyze the effect of HBQ rounds and observe a clear trend: reconstruction metrics such as rFVD and PSNR consistently improve as the number of HBQ rounds increases. While four to six rounds already yield strong performance, an eight-round configuration achieves reconstruction quality nearly identical to that of the continuous baseline as depicted in Tab. 2 and Fig. 5. This demonstrates that our HBQ tokenizer can match its continuous counterpart’s fidelity while operating at a higher compression rate. Crucially, this is achieved without increasing the latent channels. While other methods (e.g., Infinity [hanjInfinity], BitDance [bitdance]) can also bridge the gap to continuous models, they typically rely on expanding the latent dimension. As recent studies [stable-diffusion3, hanjInfinity, dcae1p5] suggest, such an approach often slows convergence and necessitates larger models. In contrast, our primary results are achieved without increasing the latent dimension.
We also experimented with expanding the latent channels from 16 to 64. This single change boosts the PSNR from 30.40 to an impressive 33.97. Notably, this performance is comparable to the state-of-the-art Wan 2.1 tokenizer, but is achieved at a 4 higher compression rate. By carefully tuning the GAN loss weight and training for more iterations, the HBQ tokenizer outperforms Wan 2.1 and Wan 2.2 in terms of SSIM and PSNR. More details on tuning are provided in Appendix E.1.
4.2 Class-to-Image Results
Implementation. Following JiT [jitpaper], we incorporate SwiGLU, RMSNorm, RoPE, qk-norm, and in-context class conditioning to the original transformer. We train GRNind with four different model sizes: 130M, 458M, 952M, and 2B, denoted as GRN-B, GRN-L, GRN-H, GRN-G, respectively. We train them for 600 epochs on the ImageNet [imagenet] dataset. The learning rate is set to 2e-4 and is constant during training. We randomly discard 10% conditions for Classifier-Free Guidance. During the inference stage, we grid search the best decoding hyperparameters. Additional implementation details are provided in Appendix 4.
| Type | Model | Tokenizer | Loss | Param | Gflops | FID | IS |
| Diffusion | DiT-L/2 [dit] | C | MSE | 458M | - | 5.02 | 167.2 |
| Diffusion | DiT-XL/2 [dit] | C | MSE | 675M | 119 | 2.27 | 278.2 |
| Flow | SiT-XL/2 [sit] | C | MSE | 675M | 119 | 2.06 | 277.5 |
| Flow | REPA [repa], SiT-XL/2 | C | MSE | 675M | 119 | 1.42 | 305.7 |
| Flow | RAE [RAE], DiT-XL/2 | C | MSE | 839M | 146 | 1.13 | 262.6 |
| Flow | JiT-B/16 [jitpaper] | C | MSE | 131M | 25 | 3.66 | 275.1 |
| Flow | JiT-L/16 [jitpaper] | C | MSE | 459M | 88 | 2.36 | 298.5 |
| Flow | JiT-H/16 [jitpaper] | C | MSE | 953M | 182 | 1.86 | 303.4 |
| Flow | JiT-G/16 [jitpaper] | C | MSE | 2B | 383 | 1.82 | 292.6 |
| Hybrid | MAR [MAR] | C | MSE | 943M | - | 1.55 | 303.7 |
| Hybrid | BitDance-H-1x [bitdance] | D | MSE | 1B | - | 1.24 | 304.4 |
| AR | LlamaGen-L [llamagen] | D | CE | 343M | - | 3.07 | 256.1 |
| AR | LlamaGen-XL [llamagen] | D | CE | 775M | - | 2.62 | 244.1 |
| AR | LlamaGen-XXL [llamagen] | D | CE | 1.4B | - | 2.34 | 253.9 |
| AR | MaskGIT [maskgit] | D | CE | 227M | - | 6.18 | 182.1 |
| AR | VAR-d20 [keyuVAR] | D | CE | 600M | - | 2.57 | 302.6 |
| AR | VAR-d24 [keyuVAR] | D | CE | 1B | - | 2.09 | 312.9 |
| AR | VAR-d30 [keyuVAR] | D | CE | 2B | - | 1.92 | 323.1 |
| AR | RandAR-XXL [pang2025randar] | D | CE | 1.4B | - | 2.15 | 322.0 |
| AR | GRN-B | D | CE | 130M | 25 | 3.56 | 280.3 |
| AR | GRN-L | D | CE | 458M | 88 | 2.64 | 314.8 |
| AR | GRN-H | D | CE | 952M | 182 | 2.06 | 316.1 |
| AR | GRN-G | D | CE | 2B | 383 | 1.81 | 299.0 |
Results. As shown in Tab. 3, GRN is benchmarked against state-of-the-art diffusion, hybrid, and autoregressive models on the ImageNet 256256 class-conditional generation task. With nearly half the parameters of MaskGIT [maskgit], our GRN-B model achieves a superior FID of 3.56 (vs. 6.18), demonstrating remarkable efficiency. Our largest variant, GRN-G, achieves a state-of-the-art FID of 1.81, rivaling top diffusion and hybrid models. Notably, GRN-G surpasses foundational models like DiT [dit] and SiT [sit] in both FID and Inception Score. This is significant as they form the backbone of many current industrial T2I and T2V models. Furthermore, GRN-G outperforms the autoregressive model LlamaGen [llamagen] and VAR [keyuVAR]. We attribute this advantage to our proposed global refinement generation, which effectively mitigates error propagation. The high-quality visual samples generated by GRN-G, shown in Fig. 1, also confirm its capabilities. Please refer to Fig. 4 in the Appendix for additional uncurated qualitative results. These strong results establish GRN as a powerful and scalable baseline for high-fidelity visual generation, motivating its application to more complex text-to-image and text-to-video tasks.
4.3 Text-to-Image Results
Implementation. We train GRNbit for the text-to-image task with 2B parameters from scratch. In contrast to C2I models, the T2I model leverages in-context self-attention instead of adaln-zero to inject conditions. The model was pre-trained on large-scale public datasets and subsequently fine-tuned on a small, high-quality proprietary dataset. We first train GRN on the pre-training dataset at a resolution of 256256 for 150K iterations using a batch size of around 15,400 and a learning rate of 2e-4. Then we fine-tune GRN at 1024 resolution with a smaller, high-quality dataset. In this stage, we train GRN for 60K iterations using a batch size of 2048 and a learning rate of 2e-5. Additional implementation details are provided in Appendix 4. Furthermore, we observe that GRN significantly outperforms Infinity in generating fine details such as small faces.
Results. As shown in Tab. 4, our model, augmented with a rewriter, achieves an overall score of 0.76 on the GenEval benchmark [ghosh2024geneval]. While our model is outperformed by larger-scale methods such as Z-Image-Turbo [cai2025z], HiDream [cai2025hidream], Qwen-Image [qwenimage2025report], and BitDance [ai2026bitdance], it is crucial to note the significant disparity in model size: these models utilize 6B to 20B parameters, whereas GRN is a far more compact 2B model. When compared at an equivalent scale, GRN demonstrates superior performance, significantly outperforming models of a similar 2B size like SD3 Medium [stable-diffusion3] (0.62) and Infinity [hanjInfinity] (0.71). The qualitative results in Fig. 5 in the Appendix showcase GRN’s strong capability to generate high-fidelity and diverse images that accurately follow user prompts.
| Model | #Param | #Data | Single Obj. | Two Obj. | Count | Colors | Pos. | Color Attri. | Overall |
|---|---|---|---|---|---|---|---|---|---|
| Proprietary Models | |||||||||
| GPT Image 1 [openai2025gpt4oimage] | N/A | N/A | 0.99 | 0.92 | 0.85 | 0.92 | 0.75 | 0.61 | 0.84 |
| Seedream 3.0 [gao2025seedream] | N/A | N/A | 0.99 | 0.96 | 0.91 | 0.93 | 0.47 | 0.80 | 0.84 |
| Diffusion Models | |||||||||
| PixArt-[chen2023pixart] | 0.6B | N/A | 0.98 | 0.50 | 0.44 | 0.80 | 0.08 | 0.07 | 0.48 |
| SD3 Medium [stable-diffusion3] | 2B | N/A | 0.98 | 0.74 | 0.63 | 0.67 | 0.34 | 0.36 | 0.62 |
| JanusFlow [janusflow] | 1.3B | N/A | 0.97 | 0.59 | 0.45 | 0.83 | 0.53 | 0.42 | 0.63 |
| FLUX.1-Dev [FLUX] | 12B | N/A | 0.98 | 0.81 | 0.74 | 0.79 | 0.22 | 0.45 | 0.66 |
| SD3.5-Large [stable-diffusion3] | 8B | N/A | 0.98 | 0.89 | 0.73 | 0.83 | 0.34 | 0.47 | 0.71 |
| Lumina-Image-2.0 [qin2025lumina] | 2.6B | 111M | - | 0.87 | 0.67 | - | - | 0.62 | 0.73 |
| Show-o2 [xie2025show] | N/A | N/A | 1.00 | 0.87 | 0.58 | 0.92 | 0.52 | 0.62 | 0.76 |
| Z-Image-Turbo [cai2025z] | 6B | N/A | 1.00 | 0.95 | 0.77 | 0.89 | 0.65 | 0.68 | 0.82 |
| HiDream-I1-Full [cai2025hidream] | 17B | N/A | 1.00 | 0.98 | 0.79 | 0.91 | 0.60 | 0.72 | 0.83 |
| Z-Image [cai2025z] | 6B | N/A | 1.00 | 0.94 | 0.78 | 0.93 | 0.62 | 0.77 | 0.84 |
| Qwen-Image [qwenimage2025report] | 20B | N/A | 0.99 | 0.92 | 0.89 | 0.88 | 0.76 | 0.77 | 0.87 |
| BAGEL [deng2025emerging] | 14B | N/A | 0.98 | 0.95 | 0.84 | 0.95 | 0.78 | 0.77 | 0.88 |
| Autoregressive Models | |||||||||
| Emu3-Gen [wang2024emu3] | 8B | N/A | 0.98 | 0.71 | 0.34 | 0.81 | 0.17 | 0.21 | 0.54 |
| Infinity†[hanjInfinity] | 2B | 160M | - | 0.85 | - | - | 0.49 | 0.57 | 0.73 |
| Janus-Pro [chen2025janus] | N/A | N/A | 0.99 | 0.89 | 0.59 | 0.90 | 0.79 | 0.66 | 0.80 |
| Tar [han2025vision] | N/A | N/A | 0.98 | 0.92 | 0.83 | 0.85 | 0.80 | 0.65 | 0.84 |
| NextStep-1 [team2025nextstep] | 14B | N/A | - | - | - | - | - | - | 0.73 |
| BitDance [ai2026bitdance] | 14B | N/A | 1.00 | 0.96 | 0.71 | 0.95 | 0.72 | 0.83 | 0.86 |
| GRN † | 2B | 80M | 0.99 | 0.90 | 0.72 | 0.84 | 0.52 | 0.60 | 0.76 |
| Models | # Params | Human | Scene | Multiple | Appear. | Quality | Semantic | Overall |
|---|---|---|---|---|---|---|---|---|
| Action | Objects | Style | Score | Score | ||||
| Diffusion / Flow Models | ||||||||
| AnimateDiff-V2 [animatediff] | 1.5B | 92.60 | 50.19 | 36.88 | 22.42 | 82.90 | 69.75 | 80.27 |
| VideoCrafter-2.0 [videocrafter] | 1.5B | 95.00 | 55.29 | 40.66 | 25.13 | 82.20 | 73.42 | 80.44 |
| OpenSora V1.2 [opensora] | 1.1B | 85.80 | 42.47 | 58.41 | 23.89 | 80.71 | 73.30 | 79.23 |
| Show-1 [show-1] | 6B | 95.60 | 47.03 | 45.47 | 23.06 | 80.42 | 72.98 | 78.93 |
| URSA [ursa] | 1.7B | - | 52.30 | 70.60 | - | 83.40 | 78.50 | 82.40 |
| CogVideoX-5B [cogvideox] | 5B | 99.40 | 53.20 | 62.11 | 24.91 | 82.75 | 77.04 | 81.61 |
| HunyuanVideo [hunyuanvideo] | 13B | 94.40 | 53.88 | 68.55 | 19.80 | 85.09 | 75.82 | 83.24 |
| Wan 2.1 [Wan] | 14B | 98.80 | 53.67 | 81.44 | 21.13 | 85.64 | 80.95 | 84.70 |
| AutoRegressive Models | ||||||||
| Nova† [nova] | 0.6B | 95.20 | 54.06 | 77.52 | 20.92 | 80.39 | 79.05 | 80.12 |
| Emu3 [wang2024emu3] | 8B | 77.71 | 37.11 | 44.64 | 20.92 | 84.09 | 68.43 | 80.96 |
| Lumos-1 [yuanlumos] | 3.6B | - | - | - | - | 79.50 | 73.50 | 78.30 |
| InfinityStar† [infinitystar] | 8B | 96.43 | 52.08 | 78.66 | 21.81 | 84.73 | 79.78 | 83.74 |
| GRN† | 2B | 93.75 | 50.44 | 70.83 | 21.30 | 84.41 | 77.35 | 82.99 |
4.4 Text-to-Video Results
Implementation. Beyond the class-to-image and text-to-image generation tasks, we extend GRN to the most challenging text-to-video synthesis task. The T2V variant of GRN shares the same architecture as its T2I counterpart but is trained exclusively on video data. For this purpose, we curated a training dataset of approximately 40 million video clips, each with a resolution of at least 256 256 and a duration of 2 to 10 seconds. The training process consists of two stages. First, we train GRN at a 192p resolution for 150K iterations with a batch size of 4096 and a learning rate of 2e-4. Subsequently, we switch to a 480p resolution for fine-tuning, training for an additional 9K iterations with a reduced batch size of 1350 and a learning rate of 2e-5. Additional implementation details are provided in Appendix 4.
Results. As demonstrated in Tab. 5, GRN exhibits superior performance in generating videos from textual prompts. When benchmarked against contemporary diffusion and flow-based models—including AnimateDiff-V2 [animatediff], VideoCrafter-2.0 [videocrafter], OpenSora V1.2 [opensora], Show-1 [show-1], and CogVideoX-5B [cogvideox]—GRN achieves significantly higher scores across quality, semantic, and overall scores. Notably, despite having only 2B parameters, GRN surpasses the much larger CogVideoX-5B [cogvideox] model, highlighting its exceptional parameter efficiency. Furthermore, our approach outperforms URSA [ursa], a discrete diffusion model of comparable size. The performance advantage of GRN becomes even more pronounced when compared to autoregressive counterparts such as Nova [nova], Emu3 [wang2024emu3], and Lumos-1 [shenoy2024lumosempoweringmultimodal]. While the 8B parameter model, InfinityStar [infinitystar], currently holds a higher overall score of 83.74, we are confident that the performance gap can be bridged by scaling up the size of GRN. We present qualitative results in Fig. 6. Please refer to Fig. 6 and Fig. 7 in the Appendix for additional results. The generated videos not only accurately capture the semantic details of the user prompts but also maintain a high degree of aesthetic and visual quality.
4.5 Ablation Studies
4.5.1 Predict Indices vs. Predict Bits
| CFG | interval | FID | IS | ||
|---|---|---|---|---|---|
| GRNind-B | 2.4 | [0.4,1] | 1.33 | 3.56 | 280.3 |
| GRNbit-B | 2.4 | [0.44,1] | 1.23 | 3.63 | 285.5 |
| GRNind-L | 2.0 | [0.40,1] | 1.30 | 2.64 | 314.8 |
| GRNbit-L | 1.9 | [0.45,1] | 1.20 | 2.47 | 287.0 |
As detailed in Sec. 3, GRN supports predicting either discrete indices (GRNind) or their binary representations (GRNbit). We compare these two prediction targets on the 256256 class-conditional image generation task using two model scales: GRN-B (130M) and GRN-L (458M). For each variant, we performed a grid search to identify the optimal decoding parameters, including CFG, CFG interval, and the temperature . The results, presented in Tab. 6, indicate that both approaches achieve comparable performance. Specifically, for the smaller GRN-B model, predicting indices yields a slightly better FID score. Conversely, for the larger GRN-L model, predicting bits proves superior, achieving a lower FID of 2.47 compared to 2.64. This suggests that GRN is well-suited for both prediction formats on the class-to-image generation task.
We further extend this comparison to the more challenging T2V generation task. As illustrated in Fig. 7, we observe that the bit prediction approach generates better videos with fewer artifacts. We hypothesize that this is because predicting bits provides a more explicit supervisory signal and mitigates the token aliasing effect inherent in index prediction, thus demonstrating superior performance on complex generation tasks. While some prior works argue that bit prediction assumes independence between bits, leading to suboptimal results, our global refinement mechanism effectively addresses this issue.
4.5.2 Global Refinement Mechanism
| CFG | interval | FID | IS | ||
|---|---|---|---|---|---|
| Refine | 2.4 | [0.44,1] | 1.23 | 3.63 | 285.5 |
| Mask | 2.4 | [0.44,1] | 1.23 | 185.62 | 4.3 |
| Mask | 8.0 | [0.00,1] | 0.50 | 18.13 | 220.2 |
In the ablation study, we validate the effectiveness of our global refinement mechanism, termed Refine. We contrast its performance with a conventional mask-based generation pipeline like MaskGIT [maskgit] or BERT [bert], where previously generated tokens are fixed. The results are striking: using identical decoding hyperparameters, the mask-based approach collapses into generating nonsensical outputs (FID=185.62), as detailed in Tab. 7. Even with optimal decoding parameters found via grid search (higher CFG, lower temperature ), the mask-based method (FID=18.13) still lags significantly behind our approach (FID=3.63). This experiment clearly demonstrates that our refinement paradigm effectively mitigates error propagation, a critical weakness in standard AR, and thus achieves superior performance.
4.5.3 Complexity-Aware Sampling
We evaluate the efficacy of our complexity-aware sampling on GRNbit-B, with hyperparameters set to . Following standard settings in diffusion models, we set the maximum number of refinement steps to . To strike a balance between performance and efficiency, we empirically set the minimum number of steps to . We synthesize 63K images and plot the distribution of their allocated generation steps in a histogram (Fig. 9). The results demonstrate that our proposed method enables GRN to dynamically allocate computational resources based on varying levels of complexity. As observed, different examples are assigned different numbers of refinement steps, ranging from 20 to 50. Over 97.9% of samples require fewer than 50 refinement steps. Intriguingly, around 6.3% of images are generated using a minimum of 20 steps, which suggests the model possesses high confidence in these particular predictions. Compared to the fixed-step baseline (50 steps for all samples), our complexity-aware sampling reduces FID from 3.56 to 3.47 while offering significant speedups (max 2.5, avg 1.25). Furthermore, we apply complexity-aware sampling to the text-to-image generation task and set . The qualitative results in Fig. 8 visually confirm that our method effectively enables complexity-aware, adaptive-step generation. For future work, we plan to explore refinement-step distillation, which is naturally compatible with complexity-aware sampling and enables more efficient visual generation.
4.5.4 Training and Inference Cost
As shown in Tab. 3, we report Gflops for one forward pass in C2I. Training cost is comparable to one forward pass while inference cost scales with steps. For T2I, GRN-2B takes 2-8s per 10241024 image with an average of 4s (vs. Infinity 1s, FluxDev 12s, Qwen-Image 33s). For T2V, GRN takes 125s for a 480p, 81-frame video (vs. Wan2.1 480s, Wan2.2 437s). Note that these baselines use optimizations from Diffusers, whereas GRN runs in bare PyTorch eager mode. We test these results using the same devices.
We also conduct ablation studies on the bit prediction target for GRNbit and the decoding parameters. For brevity, please refer to Appendix E.4 and Appendix E.5 for more details.
5 Conclusion
In this paper, we introduce GRN, a next-generation visual synthesis framework characterized by a global refinement mechanism and complexity-aware generation. We propose Hierarchical Binary Quantization (HBQ) to develop a series of discrete image and video tokenizers that are on par with their continuous counterparts while using the same number of latent channels and offering a significantly higher compression rate. In the generation phase, GRN sets new state-of-the-art results in both image reconstruction and class-conditional image generation. Extensive experiments demonstrate that, at equivalent scales, GRN surpasses existing autoregressive and diffusion-based approaches in both text-to-image and text-to-video generation tasks.
Moreover, as an autoregressive framework built entirely on discrete tokens, we believe GRN could be integrated more naturally into existing large language models. Unified learning over discrete text and visual tokens could substantially promote multimodal understanding and generation. At a fundamental level, GRN resolves the issues of quantization loss and error accumulation that have long limited previous visual autoregressive generative models. We believe it has the potential to emerge as a strong competitor to the currently dominant Transfusion [transfusion] architecture.
6 Limitations
This work also has several limitations. Due to limited computational resources, we have not scaled up the training compute or model size to the level of leading visual generation models. In addition, for the text-to-video generation task, we observe that GRN performs better in human-related scenarios. The generated videos may sometimes lack rich visual details and exhibit distortions. We believe that these limitations could be alleviated by balancing the data distribution and scaling up the model size.
7 Acknowledgements
We would like to thank Ruibiao Lu for his contributions to data collection and the video demo, and Hui Wu for his valuable advice on infrastructure.
References
Appendix A Algorithm for Hierarchical Binary Quantization
We outline the procedure for our proposed Hierarchical Binary Quantization (HBQ) in Alg. 1. The HBQ binary tokens are ordered from coarse to fine. Early tokens in the sequence correspond to core semantic concepts, while later tokens introduce high-frequency details, progressively enriching the representation. Furthermore, our multi-round quantization process generates binary tokens sequentially, with each round corresponding to a single bit. This inherent structure makes HBQ particularly well-suited for direct, bitwise prediction tasks. In contrast to methods like FSQ [fsq], which quantize vectors holistically, our approach offers a more natural framework for bitwise generation, i.e., GRNbit.
VAE Encoding
restrict data range to [-1,1]
initialize bucket centroids
obtain quantized feature
Straight-Through Estimator
VAE Decoding
Appendix B Algorithm for Generative Refinement Network
Alg. 2 outlines the pseudo-code for a single training step. The model input, denoted as , is a hybrid feature map composed of a subset of ground truth tokens and a complementary subset of random tokens. Taking as input, GRN is trained to predict the complete set of ground truth tokens. Despite its simplicity, this training strategy implicitly teaches the model to differentiate between reliable (ground truth) and unreliable (random) ones. Consequently, the model learns to preserve reliable tokens while refining the unreliable ones.
The sampling procedure of GRN is detailed in Alg. 3 (we omit complexity-aware sampling for clarity). The process is analogous to human drawing, where a state is iteratively refined. At each step , the current state consists of the already drawn content, represented by , and the remaining blank regions filled with random tokens, represented by . The model then predicts a complete set of tokens based on the current state. A randomly selected subset of these new predictions is then used to update the state for the next refinement step. This straightforward random selection mechanism elegantly unifies three essential operations into a single framework:
-
Filling: Introducing predicted tokens into previously blank areas.
-
Refining: Improving the quality of previously predicted tokens.
-
Erasing: Replacing previously predicted tokens with random ones.
To better illustrate this process, both Fig. 4 in the paper and the fourth column of Fig. 10 in the appendix visualize the sampling process of GRN. We especially highlight concrete examples of the filling, refining, and erasing tokens described above.
Appendix C Difference with Other Autoregressive Models
In Fig. 10, we compare GRN with conventional autoregressive models, including GPT-Style AR models (next-token prediction) [llamagen], VAR (next-scale prediction) [keyuVAR], and Masked AR models (BERT-Style) [maskgit]. Conventional AR models are constrained by a fixed generation order, where previously generated tokens are immutable. This often leads to issues like error propagation. In stark contrast, GRN employs a flexible global refinement strategy, leveraging its unique filling, refining, and erasing mechanism. This allows our model to iteratively revise and enhance the entire visual representation, effectively mitigating the error propagation issue inherent in conventional autoregressive models.
Appendix D Implementation Details
Model Architecture. Tab. 8 summarizes the architectural details of our proposed C2I, T2I, and T2V models. Following the methodology of JiT [jitpaper], we implement four variants for the C2I task, with model sizes scaling from 130M to 2B parameters. For T2I and T2V generation, we introduce a new 2B-parameter architecture specifically designed to meet the FlexAttention requirement, which requires the head dimension to be a multiple of 128. Furthermore, our models support sequence packing to accelerate training, and NaViT [navit] to handle arbitrary aspect ratios and resolutions.
Visual Tokenizer. For the C2I generation task, we employ an image-only visual tokenizer trained on the OpenImages dataset [openimages]. This tokenizer features a latent dimension of 16 and utilizes 4 rounds of HBQ. It compresses a 256 256 image into 16 16 16 4 binary tokens with a spatial stride of 16, achieving a state-of-the-art reconstruction FID of 0.56 on the ImageNet benchmark. In contrast, our T2I and T2V models share a unified visual tokenizer, which is jointly optimized on a mixture of images and videos. This unified tokenizer is configured with a 64-dimensional latent space, a spatial stride of 16, a temporal stride of 4, and also undergoes 4 rounds of HBQ, attaining a FVD of 26.3 and a PSNR of 34.73 on our video reconstruction benchmark. Additional results for the released tokenizers are provided in Tab. 1 and Tab. 2 in the paper, as well as Tab. 9 in the appendix.
Training. For our C2I models (GRN-B/L/H/G), we conduct training on the ImageNet dataset using 256×256 resolution images. The models are trained for 600 epochs, equivalent to 750K iterations, with a batch size of 1024. We employ a constant learning rate of 2e-4 throughout the training process. We apply a 10% condition dropping rate to enable Classifier-Free Guidance. For our GRN-T2I and GRN-T2V models, we adopt a coarse-to-fine training strategy. For instance, the GRN-T2I model is trained for 150K iterations at 256256 resolution and 60K iterations at 10241024 resolution with batch sizes of 15400 and 2048, respectively, using corresponding learning rates of 2e-4 and 2e-5. Other hyperparameters, such as the constant learning rate schedule, zero weight decay, and 10% condition drop, remain consistent across these models.
Sampling. During the inference phase, we utilize Classifier-Free Guidance to enhance sample quality and adherence to conditioning. For our ImageNet-trained models, we dynamically start CFG based on an optimal threshold found within the range of [0, 0.5]. The CFG strength is swept within [1.0, 3.0] to find the best results. For our text-conditional models (GRN-T2I and GRN-T2V), we apply CFG throughout the entire sampling process. We search for the optimal CFG strength within the range of [1.0, 4.0] and a temperature range of [0.5, 1.5] to generate diverse and high-fidelity results. For the benchmark experiments on the C2I, T2I, and T2V tasks, we use 50 fixed refinement steps. In Sec. 4.5.3, we apply complexity-aware sampling to the C2I and T2I tasks. The proposed complexity-aware sampling method is controlled by parameters and , where is optimized in the range of [300, 1200] and is set accordingly to determine overall sampling steps.
| GRN-B | GRN-L | GRN-H | GRN-G | GRN-T2I | GRN-T2V | |
| Architecture | ||||||
| Depth | 12 | 24 | 32 | 40 | 28 | 28 |
| Hidden dim | 768 | 1024 | 1280 | 1664 | 2304 | 2304 |
| Heads | 12 | 16 | 16 | 16 | 18 | 18 |
| Parameter | 130M | 458M | 952M | 2B | 2B | 2B |
| Data | ImageNet 256256 images | Images | Videos | |||
| Tokenizer | ||||||
| Tokenizer Stride | 1616 | 16164 | ||||
| Tokenizer dim | 16 | 64 | ||||
| HBQ Round | 4 | 4 | ||||
| Training | ||||||
| Epoch | 600 | N/A | ||||
| Iter | 750K | 150K/60K | 150K/9K | |||
| Batch Size | 1024 | 15400/2048 | 4096/1350 | |||
| LR | 2e-4 | 2e-4/2e-5 | 2e-4/2e-5 | |||
| LR Schedule | constant | constant | ||||
| Weight Decay | 0 | 0 | ||||
| Condition Drop | 10% | 10% | ||||
| Sampling | ||||||
| , | optimal in [300,1200], is set accordingly | |||||
| sweep range | [1.0,1.5] | [0.5,1.5] | ||||
| CFG sweep range | [1.0, 3.0] | [1.0,4.0] | ||||
| CFG interval | optimal in [0, 0.5] to start CFG | not used | ||||
Appendix E More Ablation Studies
E.1 GAN Loss in Tokenizer
Fig. 5 illustrates a clear trend: reconstruction quality steadily improves as more HBQ rounds are introduced. Although four to six rounds are sufficient for decent performance, an eight-round model closes the gap, achieving results nearly the same as the baseline without quantization.
We further investigate the impact of varying the GAN loss weight, . For this study, we fine-tune a baseline model (an HBQ tokenizer with 64 latent channels and four quantization rounds) using different values. As detailed in Tab. 9, increasing the GAN loss weight from 0.001 to 0.02 significantly improves perceptual quality, reducing the rFVD from 48.6 to 28.6. However, this comes at the cost of a drop in reconstruction fidelity, with the PSNR decreasing from 34.05 to 33.73. Empirically, we found that a weight of 0.005 strikes an effective balance between these two metrics. Consequently, this setting is applied to our GRN-T2I and GRN-T2V models.
| Method | Channel | Stride | Compress | rFVD | LPIPS | SSIM | PSNR | |
| Wan 2.1 (patchify) | N/A | 64 | 16164 | 24 | 19.5 | 0.058 | 0.929 | 34.10 |
| HBQ (baseline) | 0.001 | 64 | 16164 | 96 | 50.6 | 0.084 | 0.930 | 33.97 |
| HBQ () | 0.001 | 64 | 16164 | 96 | 48.6 | 0.083 | 0.930 | 34.05 |
| HBQ () | 0.005 | 64 | 16164 | 96 | 30.0 | 0.078 | 0.928 | 33.98 |
| HBQ () | 0.02 | 64 | 16164 | 96 | 28.6 | 0.081 | 0.925 | 33.73 |
| HBQ () | 0.1 | 64 | 16164 | 96 | 29.5 | 0.084 | 0.923 | 33.55 |
| Recons | GRN-B(130M) | GRN-L(458M) | ||||
|---|---|---|---|---|---|---|
| rFID | PSNR | gFID | IS | gFID | IS | |
| HBQ-2 | 1.23 | 21.24 | 3.22 | 304.1 | 2.71 | 309.4 |
| HBQ-4 | 0.54 | 22.81 | 3.54 | 293.0 | 2.63 | 290.1 |
| HBQ-6 | 0.49 | 23.06 | 4.74 | 235.2 | 3.10 | 253.6 |
| HBQ-8 | 0.48 | 23.09 | 5.29 | 208.8 | 3.48 | 235.2 |
E.2 Impacts of HBQ Rounds on Generation Performance
As shown in Tab. 10, although HBQ (M=2) performs best for GRN-B with 130M parameters. HBQ (M=4) surpasses HBQ (M=2) when scaled to GRN-L (458M parameters). We conclude that large models benefit from more HBQ rounds. Another interesting observation is that although the reconstruction metrics remain nearly identical from 6 rounds to 8 rounds, there are notable discrepancies in the generation metrics. This indicates that larger model capacity is indeed required when information is distributed across more distinct bits. Additionally, our experiments reveal that the accuracy of MSBs is substantially higher than that of LSBs. Future comparative studies between HBQ (16dim, M=8) and HBQ (32dim, M=4) would constitute a compelling line of experimentation.
E.3 Binary Selection Map: Random vs. Confidence
| Sampling | Metrics | |||
|---|---|---|---|---|
| Random | Confidence | FID | IS | |
| GRNbit-B | ✓ | 3.63 | 285.5 | |
| GRNbit-B | ✓ | 10.64 | 246.8 | |
As detailed in Eq. 6, the binary selection map is constructed without prior constraints during generation, meaning that we randomly select current predictions to update the state for the next refinement step. To understand the importance of this random sampling, we conducted an experiment using a confidence-based sampling alternative. In this setting, tokens to update are selected based on their prediction confidence. While this prioritizes tokens deemed more ‘correct’, the outcome was a severe performance drop (FID: 3.63 10.64), as detailed in Tab. 11. The reason for this counter-intuitive result lies in the discrepancy between training and inference patterns. Our model is trained to operate on a state where ground truth and random tokens are uniformly distributed. The confidence-based method breaks this assumption by selecting high-confidence tokens that are not uniformly distributed but are instead clustered. This distributional shift moves the input far from the manifold learned during training, resulting in a catastrophic failure of the generative process.
E.4 Bit Prediction Target: Absolute vs. Relative
We investigate two different prediction targets for the binary labels in GRNbit: absolute bits versus relative bits. While absolute bit prediction directly targets the ground-truth bits (), relative bit prediction targets whether the input bit should be flipped. This can be formulated as predicting a residual, , where a ‘1’ indicates a required flip and a ‘0’ indicates preservation. As shown in Fig. 11, our experiments reveal that predicting absolute bits yields superior results, leading to generated images with significantly better structural stability compared to those from relative bit prediction.
E.5 Decoding Hyper-Parameters
| Method | FID | IS | ||||||
| Linear | / | / | 50 | 50 | 50 | 0.0 | 3.56 | 290.3 |
| Cosine | / | / | 50 | 50 | 50 | 0.0 | 3.94 | 272.5 |
| Complexity | 700 | -620 | 20 | 50 | 40 | 7.4 | 3.47 | 287.3 |
| Linear | / | / | 30 | 30 | 30 | 0.0 | 5.06 | 262.3 |
| Cosine | / | / | 30 | 30 | 30 | 0.0 | 7.87 | 224.5 |
| Complexity | 700 | -630 | 20 | 40 | 24 | 4.0 | 3.79 | 270.9 |
| Complexity | 600 | -527 | 20 | 50 | 39 | 6.8 | 3.50 | 286.6 |
| 600 | -527 | 10 | 50 | 39 | 8.2 | 3.53 | 282.1 | |
| 700 | -620 | 20 | 50 | 40 | 7.4 | 3.47 | 287.3 | |
| 700 | -630 | 20 | 40 | 24 | 4.0 | 3.79 | 270.9 | |
| 800 | -714 | 20 | 50 | 40 | 7.9 | 3.47 | 285.9 |
Regarding the decoding parameters, we use GRNbit-B as a representative example. The optimal parameters were found to be , , and an interval of . We empirically observed that a lot of parameter combinations can achieve similar results. Intuitively, increasing or decreasing encourages more diverse generation but can also lead to greater instability. The CFG interval is introduced to restore diversity when applying a higher CFG strength, as it disables CFG during the initial decoding steps, which are crucial for determining the overall semantics. The effect of varying each parameter while keeping the others fixed is detailed in Fig. 12. As shown in Tab. 12, our complexity-aware sampling outperforms linear and cosine schedules in fewer steps. In our complexity-aware sampling, the main parameters are and . We tune for dynamic range and for average steps. and clip the number of steps to a reasonable range and require less tuning. Tab. 12 shows good results with suitable ranges for in [600, 800].
Appendix F More Qualitative Results
F.1 C2I Qualitative Results
Similar to JiT [jitpaper], we present uncurated 256256 samples generated by GRN-G in Fig. 13. To ensure a fair and representative visualization of our model’s capabilities, these images were generated with the same CFG scale 1.7 and CFG interval = [0.3, 1.0] used to achieve the best FID of 1.81. This contrasts with the common approach of using a higher CFG scale for qualitative examples, which may not reflect the model’s real performance as measured by FID.
F.2 T2I Qualitative Results
In Fig. 14, we present 10241024 images generated by GRN-T2I.
F.3 T2V Qualitative Results
In Fig. 15 and Fig. 16, we present more text-to-video generation results of GRN-T2V.