1 引言
强化学习已成为将模型与下游目标对齐的核心范式。在语言模型中,DPO(Rafailov 等人,2023)和 GRPO(Shao 等人,2024)等强化学习方法显著提升了对齐能力(Ouyang 等人,2022)和推理能力(Guo 等人,2025)。近期,这些进展已扩展到图像和视频生成领域(Liu 等人,2025;Wallace 等人,2024;Wang 和 Yu,2025;Xue 等人,2025a;Zheng 等人,2026),其中流匹配模型(Lipman 等人,2023;Liu 等人,2023)是主流的生成框架。在这些方法中,Flow-GRPO(Liu 等人,2025)和 DanceGRPO(Xue 等人,2025b)通过将确定性 ODE 采样转化为随机 SDE 轨迹,并引入 PPO 风格的比率裁剪以实施信任区域优化,展现了强劲的性能。
信任区域方法的理论基础源于信任区域策略优化(TRPO)(Schulman 等人,2015),该方法建立了一个策略改进界限:当策略更新保持在由新旧策略之间的散度定义的信任区域内时,可以保证单调改进。PPO(Schulman 等人,2017)随后引入了比率裁剪,作为 TRPO 的一种计算高效的一阶近似。然而,正如 Qi 等人(2026)所指出的,每次裁剪决策都基于真实总变分散度的单样本蒙特卡洛估计,而非散度本身。在流模型连续且高维的潜在空间中,这种估计噪声被显著放大,导致比率分布出现系统性左移,其均值低于 1(Wang 等人,2025)。我们证明这种偏差是高斯策略所固有的:因此,标准 PPO 裁剪范围实际上变得不对称,无法充分约束正优势样本的过度优化,同时却过度裁剪负优势样本。
| FLUX.1-dev | Flow-GRPO | Flow-CPS | GRPO-Guard | Flow-DPPO | |
|---|---|---|---|---|---|
| 七个绿色牛角包 | | | | | |
| 一只蓝色狗站在三只白色羊上面,后面有七根白色蜡烛 | | | | | |
| 一只蓝色长颈鹿,位于七只粉色时钟的右侧,大象的后面。 | | | | | |
为了缓解这种偏差,GRPO-Guard(Wang 等人,2025)提出了对比率分布进行归一化。虽然这种重新居中处理减轻了症状,但并未解决根本原因:该比率仍然是真实策略散度的一个有噪声的、逐样本的代理指标。我们观察到,流模型提供了一种结构上的优势,可以完全规避这个问题。由于每个步骤的策略都是高斯分布,其均值由速度网络决定,方差是固定的且依赖于调度,因此新旧策略之间的 KL 散度可以简化为一个精确的、确定性的量,该量可以通过训练期间已经执行的两次前向传播计算得出。与大语言模型设置不同——在那种情况下 DPPO(Qi 等人,2026)必须对大型词汇表采用近似的散度缩减——流模型可以在不增加额外成本的情况下实现精确的散度计算。这促使我们使用直接的 KL 近端信任区域约束来替代比率裁剪。
基于这一洞察,我们提出了 Flow-DPPO(流散度近端策略优化),该方法用基于散度的掩码替代了比率裁剪。该掩码仅在两个条件同时满足时阻止梯度更新:(1)优势函数和比率表明更新正使策略偏离旧策略,且(2)精确 KL 散度已超过阈值。这种设计在直接强制执行信任区域的同时,保留了 PPO 有益的非对称结构:使策略向旧策略靠近的更新永远不会被阻止,从而加速从过度修正中恢复。在多种基础模型上的大量实验表明,Flow-DPPO 实现了更优的奖励优化、更高的 KL 近端效率、更强的抗灾难性遗忘鲁棒性、可缓解奖励破解的均衡多目标优化,以及支持更高样本效率的稳定多轮训练。图 1 展示了定性生成结果,表明 Flow-DPPO 在实现具有竞争力的组合准确性的同时,保持了显著高于现有方法的视觉质量。
2 预备知识
流匹配(Lipman 等人,2023;Liu 等人,2023)学习一个连续时间速度场,该速度场将样本从简单源分布传输到数据分布。具体而言,设 ,并定义一个插值路径 ,其中 和 决定了数据与噪声之间的概率路径。这一构造导出了一个条件分布 。流匹配的目标是训练一个时间相关的向量场 来匹配目标速度,该速度由 和泛函 给出。随后,通过最小化回归目标来训练模型
| (1) |
其中 是一个权重函数。训练完成后,通过求解常微分方程 生成样本。在实践中,简单的数值求解器(如欧拉离散化)通常足以实现高质量采样(Karras 等人,2022;Lu 等人,2022;Song 等人,2021a)。一个值得注意的特例是整流流(Liu 等人,2023),它使用线性条件路径 和 。在此选择下,目标速度简化为 。本文全程采用这一线性调度。
2.1 面向流匹配模型的强化学习微调
对于文本条件流匹配模型,给定一个条件提示词,生成过程从高斯潜变量开始,逐步将其转换为干净的样本。在每个时间步,流模型预测一个速度场,该速度场指定了确定的生成方向。将 GRPO(Shao 等人,2024)等强化学习算法应用于流匹配模型,需要在每个去噪步骤中引入一个由采样器诱导的随机策略。Flow-GRPO(Liu 等人,2025)通过 ODE 到 SDE 的转换构建了这样的策略,该转换将概率流 ODE 转化为具有相同边际分布的等效 SDE(Albergo 和 Vanden-Eijnden,2023;Albergo 等人,2024;Song 等人,2021b):,其中 表示维纳过程增量,, 是一个控制噪声水平的标量超参数。应用欧拉-丸山离散化方法得到 Flow-SDE 采样器:
| (2) |
其中 。另一种替代方案是系数保持采样(CPS)(Wang 和 Yu,2025),它减少了 Flow-SDE 中过度的噪声注入,并更好地保留了调度器的插值结构。设 分别表示预测的干净样本和噪声分量。CPS 将潜变量更新为
| (3) |
其中 且 控制随机性。因此,Flow-SDE 和 CPS 都诱导出高斯形式的单步策略,可写为
| (4) |
其中 和 的具体形式取决于采样器。上述生成过程可以建模为一个有限时域的马尔可夫决策过程(MDP)(Black 等人,2024;Fan 等人,2023;Liu 等人,2025)。为了将离散决策过程与底层的连续时间流区分开来,我们使用 表示 MDP 状态索引,使用 表示流的反向时间变量。设 为反向时间的离散化,使得状态 对应于流时间 。步骤 的状态为 。注意 。对于 ,动作是下一个潜变量样本 ,由采样器诱导的策略 生成。给定采样的动作后,转移是确定性的,下一个状态为 。轨迹从 和 开始,在 处终止,其中 。
在完整的生成过程之后,会提供一个标量奖励。强化学习微调通过一个KL正则化项来最大化期望的终末奖励,该项会惩罚偏离预训练参考策略的行为:,其中 表示由 诱导出的轨迹, 控制正则化强度。这种KL惩罚有助于防止奖励作弊,并减轻预训练模型能力的灾难性遗忘。
Flow-GRPO(Liu 等人,2025)将GRPO应用于上述MDP。给定一个提示词 ,当前策略会生成一组样本 。它们的奖励在组内进行归一化,以获得相对优势:。在实践中,每次策略优化迭代都从展开一批数据开始,然后将这批数据分成若干小批次进行多次梯度更新。这个过程引入了策略过时性:在第一次更新之后,正在优化的策略已经偏离了生成数据的行为策略。为了控制这种离策略漂移,应用了一种信任区域机制。遵循PPO(Schulman 等人,2017),使用裁剪后的替代目标函数来优化策略:
| (5) |
为简洁起见,我们省略了KL惩罚项,并且每步的重要性比率定义为 。由于Flow-SDE和CPS都按照公式(4)定义了高斯每步策略,因此对数比率具有相同的封闭形式表达式:
| (6) |
因此,两种采样器都可以在同一个GRPO框架内进行优化,区别仅在于所诱导的随机策略的参数化方式不同。
3 方法论
在本节中,我们首先推导出一个策略改进界限,该界限证明了流模型的信任区域方法的合理性。然后,我们证明比率裁剪是对真实散度约束的一个有噪声的近似。最后,我们提出Flow-DPPO,它利用精确的KL计算来强制执行一个确定性的散度掩码,从而产生一个更紧且无方差的信任区域约束。
3.1 面向流匹配模型的信任区域策略优化
受 Schulman 等人(2017)以及 Qi 等人(2026)的启发,我们将信任域框架适配到第 2.1 节定义的流模型微调场景中。该场景与经典的折扣强化学习范式有两个重要区别。首先,这是一个无折扣的回合制任务,决策步数有限。其次,由于采用终端奖励结构,优势函数是在轨迹层面而非每一步进行估计。这些特性需要一种定制化的策略改进保证。我们遵循第 2.1 节定义的马尔可夫决策过程。
定理 1(流模型的性能差异恒等式)。
在具有决策步数的有限时域流模型马尔可夫决策过程中,令 表示期望奖励。对于任意两个策略 和 ,性能差异可分解为:,其中代理目标为
| (7) |
而误差项为
代理目标代表了对真实改进的一阶近似,而误差项则捕捉了每步策略变化之间的高阶交互作用。为了得到一个实用的优化目标,我们对这个误差项进行约束。
定理 2(流模型的策略改进界)。
在具有决策步数的有限时域流模型马尔可夫决策过程中,策略改进的下界为:
| (8) |
其中 是每步总变差散度的最大值, 是最大绝对奖励值。
详细推导请参见附录 B;附录 B.3 给出了一个更紧的、与 呈线性关系的界。该界在结构上类似于 Qi 等人(2026)推导的大语言模型策略改进界。它为流模型微调中的信任域方法提供了严格的理论依据:约束每步散度可以控制惩罚项,并保证单调改进。与 TRPO(Schulman 等人,2015)类似,我们可以求解以下约束优化问题以确保稳定的学习:
| (9) |
注记 3(高斯设定下的精确散度)。
对于公式(4)中的高斯每步策略,总变差散度是均值位移的单调函数:
| (10) |
其中 Φ 是标准正态分布的累积分布函数。因此,将 TV 散度约束在某个阈值以下,等价于对适当的参数进行约束,而这正是 Flow-DPPO 所采用的散度度量。此外,Pinsker 不等式确保了我们的基于 KL 散度的约束同样能对 TV 散度进行上界约束:当每步的 KL 散度满足条件时,我们有相应的 TV 散度上界。在高斯等协方差情况下,由于 KL 散度和 TV 散度都是参数 φ 的单调函数,其逆命题也成立。因此,我们的方法从 KL 散度和 TV 散度两个角度都具有理论依据。与大语言模型场景不同(在 LLM 场景中,离散词汇表需要近似散度计算(Qi 等人,2026)),流模型的 Gaussian 结构可以零额外成本提供精确的每步散度。
3.2 Flow-GRPO 中比率裁剪的陷阱
Flow-GRPO 采用 PPO 风格的比率裁剪来强制执行信任区域。为了与 Flow-GRPO 的符号表示(Liu 等人,2025)保持一致,在本小节及后续小节中,我们使用流时间(等价于第 2.1 节 MDP 索引中的时间步)来索引去噪步骤。裁剪条件旨在防止新策略与旧策略偏离过远。然而,概率比率本质上是一个充满噪声的代理指标,无法准确反映真实的策略散度。根据总变差散度的定义,
| (11) |
因此,每个单独的比率仅仅是 TV 散度的一个单样本蒙特卡洛估计。虽然策略改进界(定理 2)要求约束 TV 散度,但比率裁剪约束的是这个充满噪声的逐样本代理指标。Qi 等人(2026)在 LLM 场景中已经指出了这个问题;我们现在表明,由于流模型具有高维连续动作空间,由此产生的病态问题在流模型中尤为严重。
回顾公式 (6),对数比率如下:
| (12) |
由于 ε 是从标准正态分布中采样的,我们可以写出 ε = z,其中 z 服从标准正态分布。代入并令相关变量为特定值:
| (13) |
第一项是一个均值为零、方差为 2D 的随机变量。这表明对数比率主要由噪声主导:信号项(确定性的第二项)恰好是 KL 散度的负值,但它被一个标准差与信号本身同阶的噪声项所污染。这一分析得出了两个关键见解:
-
高方差。由于随机采样,该比率本身具有噪声。即使真实的 KL 散度处于中等水平,单个比率样本也可能极端(要么非常大,要么非常小),从而触发虚假的裁剪。
-
依赖噪声的裁剪。一个更新是否被裁剪,很大程度上取决于采样过程中抽取的随机噪声,而非真实的策略散度。两条具有相同策略参数但噪声实现不同的轨迹,可能会收到完全不同的裁剪决策。
相比之下,真实的 KL 散度是仅由策略参数决定的确定性函数,不受采样噪声影响。这启发了我们的方法:用直接的散度约束取代基于噪声比率的信任区域。详细的方差分析见附录 D。
3.3 面向流模型的散度近端策略优化
我们现在推导流模型设定下新旧策略之间的散度,并提出我们的 Flow-DPPO 算法。
精确 KL 散度。由于新旧策略均为具有相同方差但不同均值的高斯分布,KL 散度具有封闭形式(推导过程见附录 C):
| (14) |
对于 Flow-SDE(对应公式 (2)),有 ,得到:
| (15) |
对于 CPS(对应公式 (3)),当 时:
| (16) |
备注 4。
在大语言模型设定中,DPPO(Qi 等人,2026)必须通过词表分布的二元或 Top-K 约简来近似真实散度,因为计算 token 上的精确 TV 或 KL 散度在内存上不可行。而在流模型中,高斯策略结构能以可忽略的成本得到精确散度,即速度网络两次前向传播结果的平方差。这使得基于散度的信任区域对流模型而言,比大语言模型更为自然。
Flow-DPPO 掩码。我们将 Flow-DPPO 目标函数定义为:
| (17) |
其中基于散度的掩码为:
| (18) |
其中 且 为散度阈值。
非对称设计。公式 (18) 中的掩码保留了使 PPO 有效的非对称结构。它仅阻止那些已经偏离旧策略的更新:
-
当 且 时:梯度将策略推离旧策略(增加本已增大的动作概率)。若散度超过 ,掩码会阻止这一更新。
-
当 且 时:梯度减小本已减小的动作概率,同样使策略远离旧策略。若散度超过 ,掩码会阻止这一更新。
-
在所有其他情况下( 或 ):梯度使策略向旧策略靠近。无论散度水平如何,这些有益的更新都不会被阻止。
这种不对称性确保了信任区域约束不会阻碍策略恢复:当策略偏离过远时,纠正性更新仍不受限制。我们在附录 E 中提供了该方向条件的理论依据,并讨论了更精细的掩码变体。
4 实验
模型与基线方法。我们采用 Stable Diffusion 3.5 Medium(Esser 等人,2024;Stability AI,2024)(SD3.5)、FLUX2-klein-base-9B(Black Forest Labs,2026)(FLUX2-9B)和 FLUX.1-dev(Black Forest Labs,2024)作为基础模型,以覆盖不同的架构与规模。我们将我们的方法与四种有竞争力的基线方法进行比较:Flow-GRPO(Liu 等人,2025)、Flow-CPS(Wang 和 Yu,2025)、GRPO-Guard(Wang 等人,2025)和 Diffusion-NFT(Zheng 等人,2026)。具体而言,我们评估了本方法的两种变体:Flow-DPPO(使用 Flow-GRPO 的 SDE 采样)和 Flow-DPPO+CPS(使用 CPS 调度的 SDE 采样)。详细配置见附录 F。
指标与数据集。分别选取 GenEval2(Kamath 等人,2025)和 PickScore(Kirstain 等人,2023)作为域内数据集和域外(OOD)数据集。对于 GenEval2,我们按照官方模板生成 2 万条合成训练提示词,并在官方发布的 800 条提示词上进行评估。为监测分布偏移下的灾难性遗忘,我们在训练过程中跟踪 PickScore(Kirstain 等人,2023)、CLIP(Radford 等人,2021)得分和 HPSv2(Wu 等人,2023)得分。我们报告了单奖励优化(仅 GenEval2)和多奖励训练的结果,其中多奖励训练采用 GDPO(Liu 等人,2026)以等奖励权重聚合优势值。
| 域内(GenEval2) | 域外(PickScore) | ||||||
| 模型 | GenEval2 | CLIP | PickScore | HPSv2 | CLIP | PickScore | HPSv2 |
| 预训练基线(强化学习前) | |||||||
| SD3.5-medium | 12.4 | 0.250 | 21.00 | 0.213 | 0.244 | 19.99 | 0.210 |
| FLUX2-klein-base-9B | 25.4 | 0.281 | 20.92 | 0.228 | 0.254 | 20.05 | 0.230 |
| FLUX.1-dev | 23.3 | 0.297 | 23.26 | 0.315 | 0.276 | 21.91 | 0.304 |
| SD3.5-medium,多奖励强化学习微调 | |||||||
| Flow-GRPO | 39.9 | 0.358 | 25.09 | 0.399 | 0.273 | 22.07 | 0.349 |
| Flow-CPS | 44.6 | 0.359 | 25.51 | 0.407 | 0.265 | 22.08 | 0.343 |
| GRPO-Guard | 47.8 | 0.353 | 25.64 | 0.409 | 0.272 | 22.32 | 0.354 |
| Diffusion-NFT | 42.5 | 0.334 | 25.30 | 0.394 | 0.269 | 22.52 | 0.355 |
| Flow-DPPO | 48.1 | 0.345 | 25.63 | 0.409 | 0.273 | 22.58 | 0.360 |
| Flow-DPPO + CPS | 51.6 | 0.369 | 25.72 | 0.415 | 0.279 | 22.51 | 0.361 |
| FLUX2-klein-base-9B,多奖励强化学习微调 | |||||||
| Flow-GRPO | 46.8 | 0.371 | 25.61 | 0.412 | 0.277 | 22.62 | 0.357 |
| Flow-CPS | 47.1 | 0.361 | 25.70 | 0.416 | 0.276 | 22.85 | 0.364 |
| GRPO-Guard | 49.0 | 0.375 | 25.27 | 0.411 | 0.269 | 21.99 | 0.349 |
| Diffusion-NFT | 47.3 | 0.336 | 24.87 | 0.389 | 0.274 | 22.47 | 0.351 |
| Flow-DPPO | 57.7 | 0.364 | 25.76 | 0.418 | 0.282 | 22.90 | 0.368 |
| Flow-DPPO + CPS | 55.2 | 0.386 | 26.15 | 0.427 | 0.287 | 22.97 | 0.370 |
4.1 主要结果
性能与泛化能力。如表1所示,Flow-DPPO变体在两个基础模型和所有评估指标上始终优于所有基线,在GenEval2奖励上提升尤为显著。在单奖励设置(仅优化GenEval2)下,图2表明我们提出的变体不仅在FLUX2-9B上相比基线取得了更优性能,还展现出更稳定的训练轨迹。这些实证优势在SD3.5(图7)和FLUX.1-dev(图9)上同样成立。
我们将这种优越性归因于 Flow-DPPO 中基于散度的精确掩码。通过减轻落入信任区域之外的样本(这些样本容易受到奖励破解的影响)的干扰,Flow-DPPO 保持了更稳健的优化梯度。这种约束防止模型以牺牲其他奖励为代价过度利用单个奖励,从而在多个优化目标之间实现更优的平衡,并促进稳定收敛。图 4 中的多奖励训练曲线进一步证实了这一点,其中 Flow-DPPO 变体在 SD3.5 上的大多数指标上始终优于所有基线,且未牺牲任何单个目标。
域外行为与灾难性遗忘。为了研究灾难性遗忘,我们分析了 OOD 指标(PickScore、CLIP 和 HPSv2)以及与预训练模型的 KL 散度。如图 2 所示,随着 RL 优化推动模型追求更高的视觉质量,所有方法的 OOD 指标最初均有所上升。然而,随着训练的进行,这些指标下降,表明模型以牺牲 OOD 知识为代价过度拟合了域内奖励(GenEval2)。值得注意的是,Flow-DPPO 变体表现出显著更低的 OOD 退化,表明灾难性遗忘得到了有效缓解。图 4.2 中的定性结果进一步支持了这一点,证明我们的方法在 OOD 提示词上能更好地保持视觉保真度。一致地,表 2 显示 Flow-DPPO 变体在大多数设置下保持了更低的 KL 散度。这种减小的分布偏移与 OOD 指标趋势相符,共同表明对奖励破解和遗忘具有更强的抵抗力。最终,这些结果凸显了基于散度的掩码充当了安全边界,使模型能够从奖励中学习,而不会失去其原有的生成质量或陷入分布崩溃。
4.2 分析
非对称掩码与散度阈值。我们使用 SD3.5 结合 CPS 采样,研究了散度阈值和非对称掩码在 Flow-DPPO 中的影响(图 3)。在没有非对称掩码的情况下,训练过程会崩溃,因为信任域正则化失效;具体来说,落在信任域之外的样本在很大程度上被忽略,从而阻碍了优化进程。相反,非对称掩码将这些样本约束回信任域内,从而稳定了训练轨迹。关于散度阈值,较宽松的阈值会导致稳定性下降和收敛效果欠佳。较严格的阈值最初会减慢学习速度,但由于更严格的信任域强制执行,最终能带来更优的稳定性和略好的性能。
多轮训练与样本效率。考虑到 rollout 的高计算成本,我们研究了样本重复使用频率如何影响 SD3.5 上的优化效率。具体来说,我们改变了两个因素:(i) 每次 rollout 的组数,以及 (ii) 每次 rollout 的训练轮数(内循环)。后者决定了每个样本的重复使用频率。例如,两次内循环意味着每个 rollout 批次被用于连续两次梯度更新步骤。
| FLUX2-9B | SD3.5 | ||||
| 方法 | 单一 | 多重 | +CFG | 单一 | 多重 |
| Flow-SDE 调度 | |||||
| Flow-GRPO | 0.77 | 0.79 | 1.36 | 2.34 | 3.81 |
| GRPO-Guard | 1.07 | 1.01 | 1.63 | 2.05 | 3.33 |
| Flow-DPPO | 0.17 | 0.49 | 0.51 | 1.16 | 2.49 |
| CPS 调度 | |||||
| Flow-CPS | 0.24 | 1.66 | 1.51 | 2.41 | 3.18 |
| Flow-DPPO + CPS | 0.68 | 0.70 | 0.83 | 1.60 | 2.52 |
图 5:SD3.5 上的多轮训练(左:Flow-SDE,右:CPS)。Flow-DPPO 变体在多轮训练(G64-I2 和 G32-I2)下显示出持续的长期收益,而基线则趋于平稳甚至退化。
| FLUX2-9B | Flow-GRPO | Flow-CPS | GRPO-Guard | Flow-DPPO | Flow-DPPO+CPS | |
| 域内 | ||||||
| 背景中有一头石猪,猪前面有两只黑猫,再前面有六匹黄马 | | | | | | |
| 下方有五辆彩色自行车,自行车上方有两把石吉他,最高处有一只企鹅 | | | | | | |
| 域外 | ||||||
| 一个手持弓箭的太阳精灵,面向镜头,身处丛林瀑布场景中 | | | | | | |
| 人们在巴西烧烤,高清拍摄,佳能 EOS 5D Mark IV 单反相机 | | | | | | |
图 6:在 FLUX2-9B 上,采用单奖励设置并在相同训练迭代次数下对每个提示词控制种子的定性比较。Flow-DPPO 和 Flow-DPPO + CPS 在域内性能上保持竞争力,奖励破解现象更少,同时在域外提示词上表现出显著更少的灾难性遗忘。
虽然我们的主要实验采用 64 组、1 次内循环(G64-I1),但我们进一步探索了两种面向效率的设置:G32-I2(采样计算量减半,样本重复使用两次)和 G64-I2(标准采样计算量,训练强度加倍)。如图 4.2 所示,基线方法(Flow-GRPO、Flow-CPS)在多轮训练下难以获得持续增益,常常导致性能停滞或退化。相比之下,Flow-DPPO 的变体成功地在多次更新中重复使用采样样本,实现了持续的长周期性能提升。这一优势源于基于散度的掩码,它将更新限制在信任区域内,从而确保了高效的样本利用率。这为采样计算成本高昂的场景(例如长视频生成)提供了一个有前景的方向。
5 结论
我们证明了流模型中的比率裁剪是散度的一个有噪声且有偏的代理指标。为解决此问题,我们提出了一种基于散度的掩码,在零额外成本下使用精确的 KL 散度。在多种基础模型、采样调度和奖励目标上,Flow-DPPO 在奖励优化和灾难性遗忘方面始终优于基线方法。此外,Flow-DPPO 实现了比率裁剪方法无法做到的稳定多轮训练,为采样成本高昂的场景(例如长视频生成)提供了一个有前景的方向。
参考文献
- M. S. Albergo、M. Goldstein、N. M. Boffi、R. Ranganath 与 E. Vanden-Eijnden(2024)《具有数据依赖耦合的随机插值器》。收录于《国际机器学习大会》,第 921–937 页。引用于 §2.1。
- M. S. Albergo 与 E. Vanden-Eijnden(2023)《利用随机插值器构建归一化流》。收录于《第十一届国际学习表征大会》。引用于 §2.1。
- Black Forest Labs(2024)《FLUX.1:宣布成立 Black Forest Labs》。注释:https://blackforestlabs.ai/announcing-black-forest-labs/ 引用于图 1、图 1、§4。
- Black Forest Labs(2026)《FLUX.2 [klein]:迈向交互式视觉智能》。注释:https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence 模型权重:https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B 引用于 §4。
- K. Black、M. Janner、Y. Du、I. Kostrikov 与 S. Levine(2024)《使用强化学习训练扩散模型》。收录于《第十二届国际学习表征大会》。引用于 §2.1。
- P. Esser、S. Kulal、A. Blattmann、R. Entezari、J. Müller、H. Saini、Y. Levi、D. Lorenz、A. Sauer、F. Boesel 等人(2024)《扩展整流流 Transformer 以实现高分辨率图像合成》。收录于《第四十一届国际机器学习大会》。引用于 §4。
- Y. Fan、O. Watkins、Y. Du、H. Liu、M. Ryu、C. Boutilier、P. Abbeel、M. Ghavamzadeh、K. Lee 与 K. Lee(2023)《用于微调文生图扩散模型的强化学习》。收录于《第三十七届神经信息处理系统大会》。引用于 §2.1。
- D. Guo、D. Yang、H. Zhang、J. Song、P. Wang、Q. Zhu、R. Xu、R. Zhang、S. Ma、X. Bi 等人(2025)《DeepSeek-R1:通过强化学习激励大语言模型的推理能力》。arXiv 预印本 arXiv:2501.12948。引用于 §1。
- S. Kakade 与 J. Langford(2002)《近似最优的近似强化学习》。收录于《第十九届国际机器学习大会论文集》,第 267–274 页。引用于附录 B。
- A. Kamath、K. Chang、R. Krishna、L. Zettlemoyer、Y. Hu 和 M. Ghazvininejad(2025)《GenEval 2:解决文本到图像评估中的基准漂移问题》。arXiv 预印本 arXiv:2512.16853。引用自:表 4、表 4、图 1、图 1、第 4 节。
- T. Karras、M. Aittala、T. Aila 和 S. Laine(2022)《阐明基于扩散的生成模型的设计空间》。收录于《神经信息处理系统进展》。引用自:第 2 节。
- Y. Kirstain、A. Polyak、U. Singer、S. Matiana、J. Penna 和 O. Levy(2023)《Pick-a-pic:一个用于文本到图像生成的用户偏好开放数据集》。《神经信息处理系统进展》第 36 卷,第 36652–36663 页。引用自:第 4 节。
- J. Li、Y. Cui、T. Huang、Y. Ma、C. Fan、Y. Cheng、M. Yang、Z. Zhong 和 L. Bo(2025)《Mixgrpo:利用混合 ODE-SDE 释放基于流的 GRPO 效率》。arXiv 预印本 arXiv:2507.21802。引用自:第 F.2 节。
- Y. Lipman、R. T. Q. Chen、H. Ben-Hamu、M. Nickel 和 M. Le(2023)《用于生成建模的流匹配》。收录于《第十一届国际学习表征会议》。引用自:第 1 节、第 2 节。
- J. Liu、G. Liu、J. Liang、Y. Li、J. Liu、X. Wang、P. Wan、D. Zhang 和 W. Ouyang(2025)《Flow-GRPO:通过在线强化学习训练流匹配模型》。arXiv 预印本 arXiv:2505.05470。引用自:图 1、图 1、第 1 节、第 2.1 节、第 2.1 节、第 2.1 节、第 3.2 节、第 4 节。
- S. Liu、X. Dong、X. Lu、S. Diao、P. Belcak、M. Liu、M. Chen、H. Yin、Y. F. Wang、K. Cheng 等(2026)《GDPO:面向多奖励强化学习优化的组奖励解耦归一化策略优化》。arXiv 预印本 arXiv:2601.05242。引用自:第 4 节。
- X. Liu、C. Gong 和 qiang liu(2023)《流直且快:学习使用修正流生成和传输数据》。收录于《第十一届国际学习表征会议》。引用自:第 1 节、第 2 节、第 2 节。
- C. Lu、Y. Zhou、F. Bao、J. Chen、C. Li 和 J. Zhu(2022)《DPM-solver:一种用于约 10 步内扩散概率模型采样的快速 ODE 求解器》。收录于《神经信息处理系统进展》。引用自:第 2 节。
- L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray 等人(2022)《利用人类反馈训练语言模型遵循指令》。Advances in neural information processing systems 35,第 27730–27744 页。引用于:§1。
- P. Qi, X. Zhou, Z. Liu, T. Pang, C. Du, M. Lin 和 W. S. Lee(2026)《重新思考大语言模型强化学习中的信任区域》。arXiv 预印本 arXiv:2602.04879。引用于:附录 B、§1、§1、§3.1、§3.1、§3.2、备注 3、备注 4。
- A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark 等人(2021)《从自然语言监督中学习可迁移的视觉模型》。收录于 International conference on machine learning,第 8748–8763 页。引用于:§4。
- R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon 和 C. Finn(2023)《直接偏好优化:你的语言模型其实是一个奖励模型》。Advances in neural information processing systems 36,第 53728–53741 页。引用于:§1。
- J. Schulman, S. Levine, P. Abbeel, M. Jordan 和 P. Moritz(2015)《信任区域策略优化》。收录于 International conference on machine learning,第 1889–1897 页。引用于:附录 B、§1、§3.1。
- J. Schulman, F. Wolski, P. Dhariwal, A. Radford 和 O. Klimov(2017)《近端策略优化算法》。arXiv 预印本 arXiv:1707.06347。引用于:§1、§2.1、§3.1。
- Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu 等人(2024)《DeepSeekMath:推动开放语言模型数学推理的极限》。arXiv 预印本 arXiv:2402.03300。引用于:§1、§2.1。
- J. Song, C. Meng 和 S. Ermon(2021a)《去噪扩散隐式模型》。收录于 International Conference on Learning Representations。引用于:§2。
- Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon 和 B. Poole(2021b)《基于分数的随机微分方程生成建模》。收录于 International Conference on Learning Representations。引用于:§2.1。
- Stability AI(2024)《Stable Diffusion 3.5》。注释:https://stability.ai/news/introducing-stable-diffusion-3-5 模型权重:https://huggingface.co/stabilityai/stable-diffusion-3.5-medium 引用于:§4。
- B. Wallace、M. Dang、R. Rafailov、L. Zhou、A. Lou、S. Purushwalkam、S. Ermon、C. Xiong、S. Joty 和 N. Naik(2024)《使用直接偏好优化的扩散模型对齐》。收录于《IEEE/CVF 计算机视觉与模式识别会议论文集》,第 8228–8238 页。引用自:§1。
- F. Wang 和 Z. Yu(2025)《面向流匹配强化学习的保系数采样》。arXiv 预印本 arXiv:2509.05952。引用自:附录 A、§C.4、图 1、图 1、§1、§2.1、§4。
- J. Wang、J. Liang、J. Liu、H. Liu、G. Liu、J. Zheng、W. Pang、A. Ma、Z. Xie、X. Wang 等人(2025)《Grpo-guard:通过调节裁剪缓解流匹配中的隐式过优化》。arXiv 预印本 arXiv:2510.22319。引用自:图 1、图 1、§1、§1、§4。
- X. Wu、Y. Hao、K. Sun、Y. Chen、F. Zhu、R. Zhao 和 H. Li(2023)《人类偏好评分 v2:评估文生图合成中人类偏好的可靠基准》。arXiv 预印本 arXiv:2306.09341。引用自:§4。
- S. Xue、C. Ge、S. Zhang、Y. Li 和 Z. Ma(2025a)《优势加权匹配:在扩散模型中对齐强化学习与预训练》。arXiv 预印本 arXiv:2509.25050。引用自:§1。
- Z. Xue、J. Wu、Y. Gao、F. Kong、L. Zhu、M. Chen、Z. Liu、W. Liu、Q. Guo、W. Huang 等人(2025b)《Dancegrpo:在视觉生成中释放 GRPO 的潜力》。arXiv 预印本 arXiv:2505.07818。引用自:§1。
- K. Zheng、H. Chen、H. Ye、H. Wang、Q. Zhang、K. Jiang、H. Su、S. Ermon、J. Zhu 和 M. Liu(2026)《DiffusionNFT:基于前向过程的在线扩散强化学习》。收录于《第十四届国际学习表征会议》。引用自:图 7、图 7、§G.3.1、§1、§4。
Flow-DPPO 附录:面向流匹配模型的散度近端策略优化
附录 A Flow-DPPO 算法
我们总结了完整的 Flow-DPPO 训练流程。该算法采用 CPS 采样框架(Wang 和 Yu,2025)生成轨迹,使用组相对优势估计,并在策略优化过程中应用基于散度的掩码。
计算开销。散度计算需要在训练时对速度网络进行一次额外的前向传播,以获取 。然而,计算对数比率时已经需要这次前向传播,因此散度的计算不会带来额外成本:它仅仅是已计算出的某个差值的平方范数。
附录 B 流模型的策略改进界
我们将经典的策略改进理论(Kakade 和 Langford, 2002; Schulman 等人, 2015)适配到流模型去噪的有限时域、无折扣设定中,遵循 Qi 等人 (2026) 在大语言模型场景下的方法。我们使用第 2.1 节引入的 MDP 符号:决策步骤由 索引,状态为 ,动作为 ,终端奖励为 。
B.1 性能差异恒等式的证明
证明 [定理 1 的证明] 我们首先通过定义来表达性能差异。由于奖励仅是终端状态和提示词 的函数,我们有:
其中积分遍历所有轨迹(为符号清晰起见,我们省略了确定性转移结构)。
证明的核心是轨迹概率差异的 telescoping 恒等式。由于 ,我们应用代数恒等式 :
将其代入性能差异,并转换为在 下的期望:
我们通过加减未来比率乘积设为 1 的项来分解该表达式:
至此证明完成。
B.2 策略改进界的证明
引理 5(轨迹级总变差散度的界)。
设 和 是流模型 MDP 的两个策略。设 和 表示从状态 开始的未来子轨迹上的分布。那么:
其中期望是对从 出发所访问的状态 求取的。
证明 令 且 ,其中 。我们有:
应用伸缩恒等式(由三角不等式推导得出)并积分:
对于每个由 索引的项,对未来的动作进行积分得到 1(因为 是归一化的),剩下:
内层积分是 ,外层积分定义了在策略 下对状态的期望。因此:
证明 [定理 2 的证明] 从定理 1 出发,我们从精确的性能差异恒等式开始:
我们的目标是给出 的上界。我们首先用其最大绝对值 来约束奖励:
| (19) | ||||
对未来子轨迹的内层期望恰好是未来轨迹分布之间的两倍 TV 散度:
应用引理 5 并用 约束每一项:
代回式 (19):
计算该和: 。因此:
代入性能差异恒等式,得到所需界限:
证明完毕。
B.3 一个更紧的策略改进界限
定理 2 中关于时间步长的二次依赖关系可能过于悲观。通过利用 这一事实,我们推导出一个关于 呈线性关系的更紧界限。
从式 (19) 的中间步骤出发,内层期望为 。我们不应用引理 5,而是直接使用通用界限 :
结合两个界限,策略改进满足复合保证:
其中 。二次界限对于较小的策略变化更紧,而线性界限对于较大的更新或更长的时间步长更紧。
B.4 与高斯逐步骤散度的联系
对于式 (4) 中的高斯策略, 和 ,TV 散度具有如下闭式解:
其中 是标准正态分布的累积分布函数。由于 是严格单调递增的,TV 约束等价于:
这正式确立了 Flow-DPPO 掩码(当 时阻止更新)实现了一个信任区域约束,该约束(在单调重缩放的意义上)等价于约束逐步骤 TV 散度。因此,策略改进界限(定理 2)为 Flow-DPPO 提供了严格的理论保证:通过强制实施逐步骤散度阈值,惩罚项保持可控,从而确保单调的策略改进。
附录 C 高斯策略之间的 KL 散度
在本节中,我们推导了流模型中新旧策略之间的 KL 散度,并建立了其与策略改进界中使用的 TV 散度之间的联系。
C.1 一般高斯 KL 散度
设 和 是 中两个具有相同协方差的各向同性高斯分布。KL 散度为:
| (20) |
注意,当协方差相同时,该式关于均值是对称的。
C.2 高斯设定下 KL 与 TV 之间的联系
对于同一对高斯分布,TV 散度为:
由于 KL 和 TV 都是单一量 的单调函数,对其中一个进行阈值化等价于对另一个进行阈值化。具体来说,约束 等价于 ,进而等价于 。这表明,我们的掩码中使用的平方距离是一种统一的散度度量,在单调变换的意义上等价于 KL 和 TV 散度。
C.3 在 Flow-SDE 中的应用
对于 Flow-SDE(式 (2)),每步策略为 ,其中:
均值的差值为:
代入式 (20),其中 :
| (21) |
C.4 在 CPS 中的应用
对于 CPS(式 (3)),策略均值为 ,方差为 。使用 和 ,均值的差值为:
令 。则:
| (22) |
在先前的工作(Wang and Yu, 2025)中,为了数值稳定性去掉了归一化,将散度简化为 。我们则保留了式 (22) 中的完整归一化:因为 在后续去噪步骤中会缩小,因子 会在速度变化较小时放大对输出影响最大的散度,从而提供更严格的约束,防止分布坍缩。
附录 D 比率方差分析
我们对流模型中对数比率的方差进行了详细分析。
由式 (13),,其中 且 。由此可得:
因此 。当 KL 适中时(例如 ),对数比率的标准差为 ,这意味着单个对数比率样本在均值 附近波动约 。就比率本身而言,这大致对应一个 的乘法展开范围。
关于裁剪的含义。在典型的裁剪参数(即裁剪范围)下,对数裁剪范围为。将此窄范围与对数比标准差进行比较,我们发现即使对于适中的 KL 值,也会有相当一部分样本纯粹因噪声而被裁剪,而非因为真实散度过大。这为用直接散度测量替代基于比率的裁剪提供了严格的理论依据。
附录 E 迈向预测性散度掩码
我们回顾 Flow-DPPO 中的非对称掩码(公式 (18))。当两个条件同时成立时,该掩码会阻断梯度(即):(i) 散度已超过信任区域阈值,且 (ii) 方向性条件表明优化会将策略进一步推离。具体而言,当(梯度会进一步增加本已偏高的比率)或(梯度会进一步降低本已偏低的比率)时,方向性条件被触发。这两种情况可简洁地统一为:
| (23) |
虽然该设计在实践中有效,但方向性指标只是判断即将到来的梯度步是否会增加散度的一种启发式代理。在基于比率的信任区域(如 PPO 裁剪)中,这种符号检验有充分依据:直接反映了重要性比率单样本蒙特卡洛估计的偏差,因此 的符号能可靠地指示代理目标是否会使比率进一步偏离 1。然而,在基于散度的信任区域中(约束条件施加于 上),这种关联性就不那么直接了。比率是在单个采样动作上评估的随机量,而 衡量的是对所有动作进行积分的分布距离。正值并不能保证梯度步会增加,负值也不能保证会减少。
在本节中,我们利用流模型策略的高斯结构,推导出一个更具原理性的掩码准则。我们首先预测单步梯度更新会如何改变(§E.1),得到一个闭式表达式,该表达式分解为一阶方向项和二阶幅度项。一阶项的符号给出了一个精确的方向准则,该准则在小散度区间内恢复了当前的符号检验,但在策略已经发生漂移时揭示了一个修正。完整的表达式进一步考虑了步长和梯度幅度,从而得到一个预测性掩码(§E.2),该掩码直接预测更新后的散度是否会超过某个阈值。
E.1 预测更新后散度
固定一个去噪步骤,状态为,为简洁起见省略时间索引。记,,,。采样的动作为,其中。
我们推导代理目标上的单步梯度更新如何改变散度。关于的策略梯度为:
设有效学习率为,更新后的均值为。令。预测的更新后散度为:
| (24) |
由比值分解(式(13))可知,,因此。一阶项因此简化为。式(24)中的三项具有明确的含义:(1) 当前散度;(2) 一阶项,其符号决定了梯度步长是增大还是减小散度;(3) 非负的二阶项,随步长和梯度幅度增大而增大,始终对产生正向贡献。
一阶方向准则。散度变化的方向主要由一阶项的符号决定。由于且,该符号等于:
| (25) |
当该值为正时,梯度步长增大;为负时,则减小。等价地,这也是内积的符号,证实了代理梯度投影到了散度增大的方向上。
当前掩码的恢复。在小发散区域(即信任区域有效时的典型工作范围)内,修正项和判据简化为 。由于 ,这等价于 ,正是式(23)中的方向条件。因此,在该区域内,当前的 Flow-DPPO 掩码实现了正确的一阶发散增大判据。
修正项。当 不可忽略时(即策略已发生显著漂移),真实的发散变化方向是 而非 。减去项 改变了决策边界:样本必须满足 (而非仅仅 )时,正优势梯度才会被归类为发散增大。直观而言,当策略已偏离 后,中等程度的比率升高并不一定会进一步推高发散;只有足够大的比率才会。这给出了掩码的第一个自然改进:用 替代 作为方向指示器,我们称之为一阶预测掩码:
| (26) |
该掩码仅使用训练过程中已计算出的量(, , ),且除现有阈值 外无需额外超参数。
E.2 预测掩码
基于式(24),我们定义(完整)预测掩码,当预测的更新后发散度将超过 时阻止更新:
| (27) |
与一阶掩码的比较。一阶掩码(式(26))仅考虑发散变化的方向,仍依赖独立的阈值条件 。完整预测掩码将两者统一为单一不等式:梯度是增大还是减小发散自动编码在预测值 中,阈值比较则应用于预测发散(而非当前发散)。这带来两个后果。第一,当 时,即使发散增大的步骤也可能被允许,只要预测的 仍低于 。第二,当 接近 时,二阶项可能使 超过 ,即使一阶方向是"安全"的(即一阶掩码不会触发),从而正确阻止信任区域边界附近的大梯度步骤。
恢复现有掩码。在极限情况下,二阶项消失,简化为要求一阶方向为正且满足特定条件。结合小散度近似,这恰好恢复了当前的 Flow-DPPO 掩码(公式 (23))。
E.3 关于掩码变体的讨论
掩码的层级结构。这三种掩码构成了一个保真度递增的自然层级:
当前掩码成本最低(无需额外计算),当信任区域在整个训练过程中保持较小时即可满足需求。一阶掩码在无需额外超参数的情况下优化了方向决策。完整预测掩码则需要额外估计有效学习率,但能提供定量的散度预测。
局部近似。该分析将视为自由向量,而实际中它是神经网络的输出。由于共享参数的存在,实际变化会与其他所有输入的变化耦合。因此,预测掩码是一种局部近似,当有效学习率较小且网络雅可比矩阵在单步内近似保持不变时,其精度最高。
我们将预测掩码的实证验证留待未来工作。本分析的关键贡献有两方面:它为现有的非对称条件提供了理论依据(证明该条件是小散度区间内正确的一阶准则),并规划了一条原则性路径,用于实现更精细的信任区域约束,该路径利用了流模型策略的高斯结构。
附录 F 实验细节
F.1 计算资源
所有实验均在 NVIDIA H20 96GB GPU 上进行。表 1 中的主要结果总计需要约 9 万 GPU 小时(涵盖 SD3.5、FLUX2-klein-base-9B 和 FLUX1-dev 的所有方法与奖励配置)。包括所有消融实验、多轮次实验及辅助运行在内,本文报告的所有实验总计算成本约为 14 万 GPU 小时。
F.2 超参数
所有模型均使用 LoRA。我们对 SD3.5、FLUX2-9B 和 FLUX.1-dev 均采用 LoRA。所有模型的学习率设置与先前工作保持一致。我们将训练分辨率设为 ,SD3.5 的去噪步数设为 10,FLUX2-9B 的去噪步数设为 14。
关于 GRPO 设置,所有方法均使用组大小 16,每轮训练组数为 64。Flow-GRPO 和 Flow-CPS 的 PPO 裁剪阈值设为 ,GRPO-Guard 的 PPO 裁剪阈值设为 ,均遵循官方建议。由于 KL 缩放因子不同,Flow-DPPO 和 Flow-DPPO+CPS 的 KL 裁剪阈值分别设为 和 。我们在所有基线方法和提出的方法上应用了 MixGRPO(Li 等人,2025)提出的策略,以实现更快的收敛和更好的性能。具体来说,我们混合使用 ODE 和 SDE 采样,并从去噪步骤的前半部分中随机选择 3 步进行 SDE 采样。SDE 采样的噪声水平(CPS 采样中的 )设为 。
对于 Diffusion-NFT,其余超参数(如 EMA 调度)均遵循 SD3.5 的官方实现。
附录 G 补充实验结果
G.1 补充训练曲线
我们在图 7 中提供了 SD3.5 在单奖励设置下的训练曲线(多奖励设置见正文图 4)。我们还在图 8 中提供了 FLUX2-9B 的多奖励训练曲线,并在图 9 中提供了 FLUX.1-dev 的训练曲线。
我们还在图 9 中提供了 FLUX.1-dev 的训练曲线。
G.2 KL 散度曲线
图 12 展示了在所有六种训练设置和两种 SDE 调度下,当前模型与参考(预训练)模型之间每步的 KL 散度。相应的训练末期数值在正文表 2 中报告。
G.3 消融实验
G.3.1 无分类器引导
先前研究发现 CFG 对训练收敛和性能有显著影响(Zheng 等人,2026)。在此,我们研究 CFG 对 FLUX2-9B 上 Flow-DPPO 训练的影响,如图 10 所示,其中 CFG 尺度按照官方建议设置为 4.0。使用 CFG 时,Flow-DPPO 变体在训练奖励(GenEval2)上仍能达到最先进的性能,并减轻了在域外提示词上的灾难性遗忘,这与先前讨论中的观察结果一致。这表明基于散度的掩码在 CFG 下具有鲁棒性,并能持续提供强劲性能。
G.3.2 参考 KL 正则化强度
我们在多奖励GDPO设置下,采用CPS调度,对FLUX2-klein-base-9B模型的正则化项强度(由参数控制)进行了消融实验。图11展示了训练奖励曲线,图13展示了与预训练模型的KL散度。适度的正则化强度()能进一步缓解辅助目标(PickScore、HPSv2等)上的早期奖励作弊现象,从而平衡各奖励之间的梯度,并在最终GenEval2性能上相比无正则化基线带来额外提升,同时不降低任何单个奖励在训练结束时的表现。
| FLUX2-9B | SD3.5 | FLUX.1-dev | ||||
|---|---|---|---|---|---|---|
| 方法 | 单奖励 | 多奖励 | +CFG | 单奖励 | 多奖励 | 单奖励 |
| Flow-GRPO | 84.5 | 46.8 | 54.6 | 56.6 | 39.9 | 87.8 |
| Flow-CPS | 82.7 | 47.1 | 89.0 | 74.8 | 44.6 | 91.2 |
| GRPO-Guard | 82.8 | 49.0 | 78.8 | 85.8 | 47.8 | 87.6 |
| Diffusion-NFT | – | 47.3 | – | 64.5 | 42.5 | – |
| Flow-DPPO | 85.1 | 57.7 | 87.4 | 78.9 | 48.1 | 90.7 |
| Flow-DPPO + CPS | 92.6 | 55.2 | 91.0 | 84.1 | 51.6 | 91.6 |
G.4 GenEval2上的定量总结
为了补充上述各设置的训练曲线图,表4和表3报告了每种方法在GenEval2上的训练结束Soft TIFA得分。表4还额外报告了在域内GenEval2提示集和留出的域外PickScore验证提示集上,训练结束时的辅助CLIP、PickScore和HPSv2奖励,并通过堆叠六个模块来同时呈现SD3.5-medium和FLUX2-klein-base-9B的上下文:最先进文生图系统的已发表参考数值、对应的预训练基线得分(无强化学习),以及两种基础模型在单奖励(仅GenEval2)和多奖励(GenEval2 + CLIP + PickScore + HPSv2)配置下应用的五种强化学习微调算法。表3则将每种方法的Soft TIFA比较扩展到本文报告的所有五种训练设置。
| 域内(GenEval2) | 域外(PickScore) | ||||||
| 模型 | GenEval2 | CLIP | PickScore | HPSv2 | CLIP | PickScore | HPSv2 |
| 最先进的 T2I 模型 | |||||||
| SD3.5-large | 22.8 | – | – | – | – | – | – |
| Bagel + CoT | 23.1 | – | – | – | – | – | – |
| Qwen-Image | 33.8 | – | – | – | – | – | – |
| Gemini 2.5 Flash Image | 44.6 | – | – | – | – | – | – |
| 预训练基线(RL 之前) | |||||||
| SD3.5-medium | 12.4 | 0.250 | 21.00 | 0.213 | 0.244 | 19.99 | 0.210 |
| FLUX2-klein-base-9B | 25.4 | 0.281 | 20.92 | 0.228 | 0.254 | 20.05 | 0.230 |
| FLUX.1-dev | 23.3 | 0.297 | 23.26 | 0.315 | 0.276 | 21.91 | 0.304 |
| SD3.5-medium,单奖励 RL 微调 | |||||||
| Flow-GRPO | 56.6 | 0.297 | 21.21 | 0.219 | 0.252 | 19.33 | 0.206 |
| Flow-CPS | 74.8 | 0.313 | 21.68 | 0.235 | 0.260 | 19.94 | 0.220 |
| GRPO-Guard | 85.8 | 0.328 | 22.03 | 0.252 | 0.265 | 19.94 | 0.214 |
| Diffusion-NFT | 64.5 | 0.307 | 21.69 | 0.251 | 0.262 | 20.24 | 0.239 |
| Flow-DPPO | 78.9 | 0.319 | 22.06 | 0.263 | 0.265 | 20.45 | 0.253 |
| Flow-DPPO + CPS | 84.1 | 0.316 | 21.99 | 0.262 | 0.272 | 20.50 | 0.246 |
| SD3.5-medium,多奖励 RL 微调 | |||||||
| Flow-GRPO | 39.9 | 0.358 | 25.09 | 0.399 | 0.273 | 22.07 | 0.349 |
| Flow-CPS | 44.6 | 0.359 | 25.51 | 0.407 | 0.265 | 22.08 | 0.343 |
| GRPO-Guard | 47.8 | 0.353 | 25.64 | 0.409 | 0.272 | 22.32 | 0.354 |
| Diffusion-NFT | 42.5 | 0.334 | 25.30 | 0.394 | 0.269 | 22.52 | 0.355 |
| Flow-DPPO | 48.1 | 0.345 | 25.63 | 0.409 | 0.273 | 22.58 | 0.360 |
| Flow-DPPO + CPS | 51.6 | 0.369 | 25.72 | 0.415 | 0.279 | 22.51 | 0.361 |
| FLUX2-klein-base-9B,单奖励 RL 微调 | |||||||
| Flow-GRPO | 84.5 | 0.314 | 21.82 | 0.276 | 0.264 | 20.84 | 0.280 |
| Flow-CPS | 82.7 | 0.311 | 21.82 | 0.261 | 0.275 | 21.15 | 0.267 |
| GRPO-Guard | 82.8 | 0.312 | 20.52 | 0.210 | 0.230 | 18.45 | 0.167 |
| Flow-DPPO | 85.1 | 0.331 | 22.22 | 0.294 | 0.278 | 21.27 | 0.285 |
| Flow-DPPO + CPS | 92.6 | 0.315 | 21.97 | 0.279 | 0.265 | 20.79 | 0.272 |
| FLUX2-klein-base-9B,多奖励 RL 微调 | |||||||
| Flow-GRPO | 46.8 | 0.371 | 25.61 | 0.412 | 0.277 | 22.62 | 0.357 |
| Flow-CPS | 47.1 | 0.361 | 25.70 | 0.416 | 0.276 | 22.85 | 0.364 |
| GRPO-Guard | 49.0 | 0.375 | 25.27 | 0.411 | 0.269 | 21.99 | 0.349 |
| Diffusion-NFT | 47.3 | 0.336 | 24.87 | 0.389 | 0.274 | 22.47 | 0.351 |
| Flow-DPPO | 57.7 | 0.364 | 25.76 | 0.418 | 0.282 | 22.90 | 0.368 |
| Flow-DPPO + CPS | 55.2 | 0.386 | 26.15 | 0.427 | 0.287 | 22.97 | 0.370 |
| FLUX.1-dev,单奖励 RL 微调 | |||||||
| Flow-GRPO | 87.8 | 0.331 | 23.03 | 0.311 | 0.291 | 21.85 | 0.311 |
| Flow-CPS | 91.2 | 0.328 | 23.20 | 0.317 | 0.288 | 21.98 | 0.307 |
| GRPO-Guard | 87.6 | 0.333 | 22.69 | 0.293 | 0.286 | 21.03 | 0.276 |
| Flow-DPPO | 90.7 | 0.331 | 23.15 | 0.323 | 0.290 | 21.60 | 0.300 |
| Flow-DPPO + CPS | 91.6 | 0.331 | 23.29 | 0.322 | 0.289 | 21.91 | 0.305 |
1 Introduction
Reinforcement learning (RL) has emerged as a core paradigm for aligning models with downstream objectives. In language models, RL methods such as DPO (Rafailov et al., 2023) and GRPO (Shao et al., 2024) have substantially improved alignment (Ouyang et al., 2022) and reasoning capabilities (Guo et al., 2025). Recently, these advances have been extended to image and video generation Liu et al. (2025); Wallace et al. (2024); Wang and Yu (2025); Xue et al. (2025a); Zheng et al. (2026), where flow matching models (Lipman et al., 2023; Liu et al., 2023) represent the dominant generative framework. Among them, Flow-GRPO (Liu et al., 2025) and DanceGRPO (Xue et al., 2025b) demonstrated strong performance by transforming deterministic ODE sampling into stochastic SDE trajectories and introducing PPO-style ratio clipping to enforce trust-region optimization.
The theoretical foundation of trust-region methods originates from Trust Region Policy Optimization (TRPO) (Schulman et al., 2015), which establishes a policy improvement bound: monotonic improvement is guaranteed when policy updates remain within a trust region defined by the divergence between the old and new policies. PPO (Schulman et al., 2017) later introduced ratio clipping as a computationally efficient first-order approximation to TRPO. However, as noted by Qi et al. (2026), each clipping decision is based on a single-sample Monte Carlo estimate of the true Total Variation (TV) divergence, rather than the divergence itself. In the continuous and high-dimensional latent space of flow models, this estimation noise becomes substantially amplified, leading to a systematic left shift in the ratio distribution, with its mean falling below one (Wang et al., 2025). We show that this bias is intrinsic to Gaussian policies: the standard PPO clipping range therefore becomes effectively asymmetric, failing to adequately constrain over-optimization for positive-advantage samples while excessively clipping negative-advantage ones.
| FLUX.1-dev | Flow-GRPO | Flow-CPS | GRPO-Guard | Flow-DPPO | |
|---|---|---|---|---|---|
| seven green croissants | | | | | |
| a blue dog on top of three white sheep behind seven white candles | | | | | |
| a blue giraffe behind seven pink clocks to the right of an elephant | | | | | |
To mitigate this bias, GRPO-Guard (Wang et al., 2025) proposed normalizing the ratio distribution. While this re-centering alleviates the symptom, it does not address the root cause: the ratio remains a noisy, per-sample proxy for the true policy divergence. We observe that flow models offer a structural advantage that sidesteps this problem entirely. Because each per-step policy is Gaussian with a mean determined by the velocity network and a fixed, schedule-dependent variance , the KL divergence between old and new policies reduces to , which is an exact, deterministic quantity that can be computed from two forward passes already performed during training. Unlike the LLM setting, where DPPO (Qi et al., 2026) must resort to approximate divergence reductions over large vocabularies, flow models admit exact divergence computation at no additional cost. This motivates replacing ratio clipping with a direct KL-proximal trust region constraint.
Building on this insight, we propose Flow-DPPO (Flow Divergence Proximal Policy Optimization), which replaces ratio clipping with a divergence-based mask. The mask blocks gradient updates only when two conditions are jointly met: (1) the advantage and ratio indicate that the update is moving the policy away from the old policy, and (2) the exact KL divergence already exceeds a threshold. This design directly enforces the trust region while preserving the beneficial asymmetric structure of PPO: updates that move the policy towards the old policy are never blocked, accelerating recovery from overshooting. Extensive experiments on various base models demonstrate that Flow-DPPO achieves superior reward optimization, improved KL-proximal efficiency, stronger robustness to catastrophic forgetting, balanced multi-objective optimization that mitigates reward hacking, and stable multi-epoch training that enables higher sample efficiency. Figure 1 presents qualitative generation results demonstrating that Flow-DPPO achieves competitive compositional accuracy while preserving notably higher visual quality than existing methods.
2 Preliminaries
Flow matching (Lipman et al., 2023; Liu et al., 2023) learns a continuous-time velocity field that transports samples from a simple source distribution to the data distribution. Specifically, let , and define an interpolating path with , where and determine the probability path between data and noise. This construction induces a conditional distribution . The goal of flow matching is to train a time-dependent vector field to match the target velocity, which is given by and the functinal . The model is then trained by minimizing the regression objective
| (1) |
where is a weighting function. After training, samples are generated by solving the ODE . In practice, simple numerical solvers such as Euler discretization are often sufficient for high-quality sampling (Karras et al., 2022; Lu et al., 2022; Song et al., 2021a). A notable special case is rectified flow (Liu et al., 2023), which uses the linear conditional path and . Under this choice, the target velocity reduces to . We adopt this linear schedule throughout the paper.
2.1 RL Fine-Tuning for Flow Matching Models
For text-conditional flow matching models, given a conditioning prompt , generation starts from a Gaussian latent and progressively transforms it into a clean sample . At each timestep , the flow model predicts a velocity field , which specifies a deterministic generation direction. Applying RL algorithms such as GRPO (Shao et al., 2024) to flow matching models requires a sampler-induced stochastic policy at each denoising step. Flow-GRPO (Liu et al., 2025) constructs such a policy via an ODE-to-SDE conversion, which transforms the probability-flow ODE into an equivalent SDE with the same marginals (Albergo and Vanden-Eijnden, 2023; Albergo et al., 2024; Song et al., 2021b): , where denotes Wiener process increments, , and is a scalar hyperparameter controlling the noise level. Applying Euler–Maruyama discretization yields the Flow-SDE sampler:
| (2) |
with . An alternative is Coefficients-Preserving Sampling (CPS) (Wang and Yu, 2025), which reduces the excessive noise injection in Flow-SDE and better preserves the interpolation structure of the scheduler. Let denote the predicted clean sample and noise component, respectively. CPS updates the latent as
| (3) |
where and controls the stochasticity. Both Flow-SDE and CPS therefore induce Gaussian per-step policies written as
| (4) |
where the specific forms of and depend on the sampler. The above generative process can be formulated as a finite-horizon Markov Decision Process (MDP) (Black et al., 2024; Fan et al., 2023; Liu et al., 2025). To distinguish the discrete decision process from the underlying continuous-time flow, we use for the MDP state index and for the reverse-time variable of the flow. Let be a discretization of reverse time, so that state corresponds to flow time . The state at step is . Note that . For , the action is the next latent sample, , drawn from the sampler-induced policy . Given the sampled action, the transition is deterministic, with next state . The rollout starts from and , and terminates at where .
After the full generative process, a scalar reward is provided. RL fine-tuning maximizes the expected terminal reward with a KL regularization term that penalizes deviation from the pretrained reference policy : , where denotes a trajectory induced by and controls the regularization strength. This KL penalty discourages reward hacking and mitigates catastrophic forgetting of the pretrained model’s capabilities.
Flow-GRPO (Liu et al., 2025) applies GRPO to the above MDP. Given a prompt , the current policy generates a group of samples . Their rewards are normalized within the group to obtain relative advantages: . In practice, each policy optimization iteration begins by rolling out a batch of data, which is then split into several minibatches for multiple gradient steps. This procedure introduces policy staleness: after the first update, the optimizing policy has already diverged from the behavior policy that generated the data. To control this off-policy drift, a trust region mechanism is applied. Following PPO (Schulman et al., 2017), the policy is optimized using the clipped surrogate objective
| (5) |
where we omit the KL penalty term for brevity, and the per-step importance ratio is defined as . Since both Flow-SDE and CPS define Gaussian per-step policies as in Eq. (4), the log-ratio admits the same closed-form expression:
| (6) |
Therefore, both samplers can be optimized within the same GRPO framework, differing only in the parameterization of the induced stochastic policy.
3 Methodology
In this section, we first derive a policy improvement bound that justifies trust-region methods for flow models. Then, we show that ratio clipping is a noisy proxy for the true divergence constraint. Finally, we present Flow-DPPO, which leverages exact KL computation to enforce a deterministic divergence mask, yielding a tighter and variance-free trust-region constraint.
3.1 Trust-Region Policy Optimization for Flow Matching Models
Inspired by Schulman et al. (2017); Qi et al. (2026), we adapt the trust region framework to the flow model fine-tuning setting defined in Section˜2.1. This setting differs from the classical discounted RL paradigm in two important ways. First, the problem is an undiscounted episodic task with a finite horizon of decision steps. Second, due to the terminal reward structure, advantages are estimated at the trajectory level rather than per step. These properties necessitate a tailored policy improvement guarantee. We follow the MDP defined in Section˜2.1.
Theorem 1(Performance Difference Identity for Flow Models).
In the finite-horizon flow model MDP with decision steps, let denote the expected reward. For any two policies and , the performance difference decomposes as: , where the surrogate objective is
| (7) |
and the error term is
The surrogate represents a first-order approximation to the true improvement, while the error term captures higher-order interactions between per-step policy changes. To yield a practical optimization objective, we bound this error term.
Theorem 2(Policy Improvement Bound for Flow Models).
In the finite-horizon flow model MDP with decision steps, the policy improvement is lower-bounded by:
| (8) |
where is the maximum per-step Total Variation divergence, and is the maximum absolute reward.
Please refer to Appendix B for the detailed derivation; a tighter bound linear in is given in Appendix B.3. This bound is structurally analogous to the policy improvement bound for LLMs derived in Qi et al. (2026). It provides a rigorous justification for trust-region methods in flow model fine-tuning: constraining the per-step divergence controls the penalty term and guarantees monotonic improvement. Similar to TRPO (Schulman et al., 2015), we can solve the following constrained optimization problem to ensure stable learning:
| (9) |
Remark 3(Exact Divergence in the Gaussian Setting).
For the Gaussian per-step policies in Eq. (4), the TV divergence is a monotone function of the mean displacement:
| (10) |
where is the standard normal CDF. Constraining the TV divergence below a threshold is therefore equivalent to constraining for an appropriate , which is precisely the divergence measure that Flow-DPPO employs. Moreover, the Pinsker inequality ensures that our KL-based constraint also upper-bounds the TV divergence: when the per-step , we have . In the Gaussian equal-covariance case, the converse also holds since KL and TV are both monotone functions of . Thus, our method is theoretically justified from both the KL and TV perspectives. Unlike the LLM setting, where the discrete vocabulary requires approximate divergence computations (Qi et al., 2026), the Gaussian structure of flow models provides exact per-step divergence at zero additional cost.
3.2 Pitfalls of Ratio Clipping in Flow-GRPO
Flow-GRPO adopts PPO-style ratio clipping to enforce a trust region. For consistency with the Flow-GRPO notation (Liu et al., 2025), in this and the following subsections we index denoising steps by the flow time (equivalently, in the MDP indexing of Section˜2.1). The clipping condition is intended to prevent the new policy from deviating too far from the old one. However, the probability ratio is a fundamentally noisy proxy for the true policy divergence. By definition of the Total Variation divergence,
| (11) |
so each individual is merely a single-sample Monte Carlo estimate of . While the policy improvement bound (Theorem˜2) calls for constraining , ratio clipping constrains this noisy per-sample surrogate instead. This issue was identified by Qi et al. (2026) in the LLM setting; we now show that the resulting pathology is particularly severe in flow models due to the high-dimensional continuous action space.
Recall from Eq. (6) that the log-ratio is:
| (12) |
Since is sampled from , we can write where . Substituting and letting :
| (13) |
The first term, , is a zero-mean random variable with variance . This reveals that the log-ratio is dominated by noise: the signal (the deterministic second term ) is exactly the negative of the KL divergence, but it is corrupted by a noise term whose standard deviation is of the same order as the signal itself. This analysis yields two key insights:
-
High variance. The ratio is inherently noisy due to the stochastic sample . Even when the true KL divergence is moderate, individual ratio samples can be extreme (either very large or very small), triggering spurious clipping.
-
Noise-dependent clipping. Whether an update is clipped depends heavily on the random noise drawn during sampling, rather than the true policy divergence. Two trajectories with identical policy parameters but different noise realizations may receive entirely different clipping decisions.
In contrast, the true KL divergence is a deterministic function of the policy parameters alone, unaffected by the sampling noise. This motivates our approach: replace the noisy ratio-based trust region with a direct divergence constraint. A detailed variance analysis is provided in Appendix D.
3.3 Divergence Proximal Policy Optimization for Flow Models
We now derive the divergence between old and new policies in the flow model setting and present our Flow-DPPO algorithm.
Exact KL divergence. Since both and are Gaussians with the same variance but different means, the KL divergence admits the closed form (see Appendix C for derivation):
| (14) |
For Flow-SDE (corresponding to Eq. (2)), , giving:
| (15) |
For CPS (corresponding to Eq. (3)), with :
| (16) |
Remark 4.
In the LLM setting, DPPO (Qi et al., 2026) must approximate the true divergence via Binary or Top-K reductions of the vocabulary distribution, as computing exact TV or KL over tokens is memory-prohibitive. In flow models, the Gaussian policy structure yields exact divergence at negligible cost, namely the squared difference between two forward passes of the velocity network. This makes divergence-based trust regions strictly more natural for flow models than for LLMs.
The Flow-DPPO mask. We define the Flow-DPPO objective as:
| (17) |
where the divergence-based mask is:
| (18) |
with and a divergence threshold.
Asymmetric design. The mask in Eq. (18) preserves the asymmetric structure that makes PPO effective. It only blocks updates that are already moving away from the old policy:
-
When and : the gradient is pushing the policy further from (increasing an already-increased action probability). The mask blocks this if the divergence exceeds .
-
When and : the gradient is decreasing an already-decreased action probability, again moving away from the old policy. The mask blocks this if divergence exceeds .
-
In all other cases ( or ): the gradient is moving the policy towards the old policy. These beneficial updates are never blocked, regardless of the divergence level.
This asymmetry ensures that the trust region constraint does not impede recovery: when the policy has drifted too far, corrective updates remain uninhibited. We provide a justification of this directional condition and discuss refined mask variants in Appendix E.
4 Experiments
Models and Baselines. We employ Stable Diffusion 3.5 Medium (Esser et al., 2024; Stability AI, 2024) (SD3.5), FLUX2-klein-base-9B (Black Forest Labs, 2026) (FLUX2-9B) and FLUX.1-dev (Black Forest Labs, 2024) as base models to cover diverse architectures and scales. We compare our method against four competitive baselines: Flow-GRPO (Liu et al., 2025), Flow-CPS (Wang and Yu, 2025), GRPO-Guard (Wang et al., 2025) and Diffusion-NFT (Zheng et al., 2026). Specifically, we evaluate two variants of our approach: Flow-DPPO (using SDE sampling from Flow-GRPO) and Flow-DPPO+CPS (using CPS-scheduled SDE sampling). Detailed configurations are deferred to Appendix F.
Metrics and Datasets. GenEval2 (Kamath et al., 2025) and PickScore (Kirstain et al., 2023) are selected as in-domain and out-of-domain (OOD) datasets, respectively. For GenEval2, we follow the official template to generate 20k synthetic training prompts and evaluate on the 800 officially released prompts. To monitor catastrophic forgetting under distribution shifts, we track PickScore (Kirstain et al., 2023), CLIP (Radford et al., 2021) score, and HPSv2 (Wu et al., 2023) during training. We report results for both single-reward optimization (GenEval2 only) and multi-reward training, where GDPO (Liu et al., 2026) aggregates advantages with equal reward weights.
| In-Domain (GenEval2) | Out-of-Domain (PickScore) | ||||||
| Model | GenEval2 | CLIP | PickScore | HPSv2 | CLIP | PickScore | HPSv2 |
| Pretrained baselines (before RL) | |||||||
| SD3.5-medium | 12.4 | 0.250 | 21.00 | 0.213 | 0.244 | 19.99 | 0.210 |
| FLUX2-klein-base-9B | 25.4 | 0.281 | 20.92 | 0.228 | 0.254 | 20.05 | 0.230 |
| FLUX.1-dev | 23.3 | 0.297 | 23.26 | 0.315 | 0.276 | 21.91 | 0.304 |
| SD3.5-medium, multi-reward RL fine-tuning | |||||||
| Flow-GRPO | 39.9 | 0.358 | 25.09 | 0.399 | 0.273 | 22.07 | 0.349 |
| Flow-CPS | 44.6 | 0.359 | 25.51 | 0.407 | 0.265 | 22.08 | 0.343 |
| GRPO-Guard | 47.8 | 0.353 | 25.64 | 0.409 | 0.272 | 22.32 | 0.354 |
| Diffusion-NFT | 42.5 | 0.334 | 25.30 | 0.394 | 0.269 | 22.52 | 0.355 |
| Flow-DPPO | 48.1 | 0.345 | 25.63 | 0.409 | 0.273 | 22.58 | 0.360 |
| Flow-DPPO + CPS | 51.6 | 0.369 | 25.72 | 0.415 | 0.279 | 22.51 | 0.361 |
| FLUX2-klein-base-9B, multi-reward RL fine-tuning | |||||||
| Flow-GRPO | 46.8 | 0.371 | 25.61 | 0.412 | 0.277 | 22.62 | 0.357 |
| Flow-CPS | 47.1 | 0.361 | 25.70 | 0.416 | 0.276 | 22.85 | 0.364 |
| GRPO-Guard | 49.0 | 0.375 | 25.27 | 0.411 | 0.269 | 21.99 | 0.349 |
| Diffusion-NFT | 47.3 | 0.336 | 24.87 | 0.389 | 0.274 | 22.47 | 0.351 |
| Flow-DPPO | 57.7 | 0.364 | 25.76 | 0.418 | 0.282 | 22.90 | 0.368 |
| Flow-DPPO + CPS | 55.2 | 0.386 | 26.15 | 0.427 | 0.287 | 22.97 | 0.370 |
4.1 Main results
Performance and Generalization. As summarized in Table 1, Flow-DPPO variants consistently outperform all baselines across both base models and all evaluation metrics, with particularly substantial gains in the GenEval2 reward. In the single-reward setting (optimizing GenEval2 only), Figure 2 demonstrates that our proposed variants not only achieve superior performance on FLUX2-9B compared to baselines but also exhibit a more stable training trajectory. These empirical advantages persist across SD3.5 (Figure 7) and FLUX.1-dev (Figure 9).
We attribute this superiority to the precise divergence-based mask in Flow-DPPO. By mitigating the influence of samples falling outside the trust region, which are susceptible to reward hacking, Flow-DPPO maintains a more robust optimization gradient. This constraint prevents the model from excessively exploiting individual rewards at the expense of others, thereby achieving a superior balance across multiple optimization objectives and fostering stable convergence. This is further corroborated by the multi-reward training curves in Figure 4, where Flow-DPPO variants consistently outperform all baselines across most metrics on SD3.5, without sacrificing any individual objective.
Out-of-domain Behavior and Catastrophic Forgetting. To investigate catastrophic forgetting, we analyze OOD metrics (PickScore, CLIP, and HPSv2) and the KL divergence from the pre-trained model. As illustrated in Figure 2, OOD metrics initially increase across all methods as RL optimization drives the model toward higher visual quality. However, as training progresses, these metrics decline, indicating that the model overfits the in-domain reward (GenEval2) at the expense of OOD knowledge. Notably, Flow-DPPO variants exhibit significantly less OOD degradation, suggesting that catastrophic forgetting is effectively mitigated. Qualitative results in Figure 4.2 further support this, demonstrating that our methods better preserve visual fidelity on OOD prompts. Consistently, Table 2 shows that Flow-DPPO variants maintain a lower KL divergence in most settings. This reduced distribution drift aligns with OOD metric trends, collectively indicating stronger resistance to reward hacking and forgetting. Ultimately, these results highlight that the divergence-based mask acts as a safety boundary, allowing the model to learn from rewards without losing its original generative quality or falling into distribution collapse.
4.2 Analysis
Asymmetric Masking and Divergence Threshold. We investigate the impact of the divergence threshold and asymmetric masking in Flow-DPPO using SD3.5 with CPS sampling (Figure 3). Without asymmetric masking, the training process collapses as the trust-region regularization becomes ineffective; specifically, samples falling outside the trust region are largely ignored, preventing optimization progress. Conversely, asymmetric masking constrains these samples back within the trust region, thereby stabilizing the trajectory. Regarding the divergence threshold, a looser threshold () results in diminished stability and suboptimal convergence. A tighter threshold () initially slows down learning but fosters superior stability and slightly better final performance due to more rigorous trust-region enforcement.
Multi-epoch Training and Sample Efficiency. Given the high computational cost of rollouts, we investigate how sample reuse frequency affects optimization efficiency on SD3.5. Specifically, we vary two factors: (i) the number of groups per rollout, and (ii) the number of training epochs per rollout (inner loops). The latter determines the reuse frequency of each sample. For instance, two inner loops imply that each rollout batch is utilized for two consecutive gradient steps.
| FLUX2-9B | SD3.5 | ||||
| Method | Single | Multi | +CFG | Single | Multi |
| Flow-SDE schedule | |||||
| Flow-GRPO | 0.77 | 0.79 | 1.36 | 2.34 | 3.81 |
| GRPO-Guard | 1.07 | 1.01 | 1.63 | 2.05 | 3.33 |
| Flow-DPPO | 0.17 | 0.49 | 0.51 | 1.16 | 2.49 |
| CPS schedule | |||||
| Flow-CPS | 0.24 | 1.66 | 1.51 | 2.41 | 3.18 |
| Flow-DPPO + CPS | 0.68 | 0.70 | 0.83 | 1.60 | 2.52 |
Figure 5: Multi-epoch training on SD3.5 (Left: Flow-SDE, Right: CPS). Flow-DPPO variants show consistent long-term gains under multi-epoch training (G64-I2 and G32-I2), while baselines plateau or even degrade.
| FLUX2-9B | Flow-GRPO | Flow-CPS | GRPO-Guard | Flow-DPPO | Flow-DPPO+CPS | |
| In-Domain | ||||||
| a stone pig in background, two black cats in front of the pig, and six yellow horses in front | | | | | | |
| five colorful bicycles below, two stone guitars above them, and a penguin at the highest point | | | | | | |
| Out-of-Domain | ||||||
| a sun elf with a bow, facing the camera, in a jungle waterfall scene | | | | | | |
| people at a barbecue in Brazil, captured in HD, Canon EOS 5D Mark IV DSLR | | | | | | |
Figure 6: Qualitative comparison on FLUX2-9B with single-reward setting and controlled seeds for each prompt at the same training iteration. Flow-DPPO and Flow-DPPO + CPS retain competitive in-domain performance with less reward hacking while exhibiting notably less catastrophic forgetting on out-of-domain prompts.
While our main experiments use 64 groups with 1 inner loop (G64-I1), we further explore two efficiency-oriented settings: G32-I2 (half the rollout computation with samples reused twice) and G64-I2 (standard rollout computation with doubled training intensity). As shown in Figure 4.2, baseline methods (Flow-GRPO, Flow-CPS) struggle to achieve sustained gains under multi-epoch training, often leading to performance plateaus or degradation. In contrast, Flow-DPPO variants successfully reuse rollout samples across multiple updates, yielding consistent long-term performance improvements. This advantage stems from the divergence-based mask, which constrains updates within the trust region, ensuring efficient sample utilization. This offers a promising direction for scenarios where rollouts are computationally expensive, such as long-video generation.
5 Conclusion
We show ratio clipping in flow models is a noisy, biased proxy for divergence. To address this, we propose a divergence-based mask using the exact KL at zero extra cost. Across multiple base models, sampling schedules, and reward objectives, Flow-DPPO consistently achieves superior performance than baselines in terms of reward optimization and catastrophic forgetting. Furthermore, Flow-DPPO enables stable multi-epoch training where ratio clipping degrades, offering a promising direction for scenarios with expensive rollouts, such as long-video generation.
References
- M. S. Albergo, M. Goldstein, N. M. Boffi, R. Ranganath, and E. Vanden-Eijnden (2024) Stochastic interpolants with data-dependent couplings. In International Conference on Machine Learning, pp. 921–937. Cited by: §2.1.
- M. S. Albergo and E. Vanden-Eijnden (2023) Building normalizing flows with stochastic interpolants. In The Eleventh International Conference on Learning Representations, Cited by: §2.1.
- Black Forest Labs (2024) FLUX.1: announcing black forest labs. Note: https://blackforestlabs.ai/announcing-black-forest-labs/ Cited by: Figure 1, Figure 1, §4.
- Black Forest Labs (2026) FLUX.2 [klein]: towards interactive visual intelligence. Note: https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligenceModel weights: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B Cited by: §4.
- K. Black, M. Janner, Y. Du, I. Kostrikov, and S. Levine (2024) Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, Cited by: §2.1.
- P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. (2024) Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, Cited by: §4.
- Y. Fan, O. Watkins, Y. Du, H. Liu, M. Ryu, C. Boutilier, P. Abbeel, M. Ghavamzadeh, K. Lee, and K. Lee (2023) Reinforcement learning for fine-tuning text-to-image diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, Cited by: §2.1.
- D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. (2025) Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: §1.
- S. Kakade and J. Langford (2002) Approximately optimal approximate reinforcement learning. In Proceedings of the nineteenth international conference on machine learning, pp. 267–274. Cited by: Appendix B.
- A. Kamath, K. Chang, R. Krishna, L. Zettlemoyer, Y. Hu, and M. Ghazvininejad (2025) GenEval 2: addressing benchmark drift in text-to-image evaluation. arXiv preprint arXiv:2512.16853. Cited by: Table 4, Table 4, Figure 1, Figure 1, §4.
- T. Karras, M. Aittala, T. Aila, and S. Laine (2022) Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, Cited by: §2.
- Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy (2023) Pick-a-pic: an open dataset of user preferences for text-to-image generation. Advances in neural information processing systems 36, pp. 36652–36663. Cited by: §4.
- J. Li, Y. Cui, T. Huang, Y. Ma, C. Fan, Y. Cheng, M. Yang, Z. Zhong, and L. Bo (2025) Mixgrpo: unlocking flow-based grpo efficiency with mixed ode-sde. arXiv preprint arXiv:2507.21802. Cited by: §F.2.
- Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023) Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, Cited by: §1, §2.
- J. Liu, G. Liu, J. Liang, Y. Li, J. Liu, X. Wang, P. Wan, D. Zhang, and W. Ouyang (2025) Flow-grpo: training flow matching models via online rl. arXiv preprint arXiv:2505.05470. Cited by: Figure 1, Figure 1, §1, §2.1, §2.1, §2.1, §3.2, §4.
- S. Liu, X. Dong, X. Lu, S. Diao, P. Belcak, M. Liu, M. Chen, H. Yin, Y. F. Wang, K. Cheng, et al. (2026) Gdpo: group reward-decoupled normalization policy optimization for multi-reward rl optimization. arXiv preprint arXiv:2601.05242. Cited by: §4.
- X. Liu, C. Gong, and qiang liu (2023) Flow straight and fast: learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, Cited by: §1, §2, §2.
- C. Lu, Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu (2022) DPM-solver: a fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, Cited by: §2.
- L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. (2022) Training language models to follow instructions with human feedback. Advances in neural information processing systems 35, pp. 27730–27744. Cited by: §1.
- P. Qi, X. Zhou, Z. Liu, T. Pang, C. Du, M. Lin, and W. S. Lee (2026) Rethinking the trust region in llm reinforcement learning. arXiv preprint arXiv:2602.04879. Cited by: Appendix B, §1, §1, §3.1, §3.1, §3.2, Remark 3, Remark 4.
- A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. (2021) Learning transferable visual models from natural language supervision. In International conference on machine learning, pp. 8748–8763. Cited by: §4.
- R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn (2023) Direct preference optimization: your language model is secretly a reward model. Advances in neural information processing systems 36, pp. 53728–53741. Cited by: §1.
- J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz (2015) Trust region policy optimization. In International conference on machine learning, pp. 1889–1897. Cited by: Appendix B, §1, §3.1.
- J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §1, §2.1, §3.1.
- Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al. (2024) Deepseekmath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: §1, §2.1.
- J. Song, C. Meng, and S. Ermon (2021a) Denoising diffusion implicit models. In International Conference on Learning Representations, Cited by: §2.
- Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole (2021b) Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, Cited by: §2.1.
- Stability AI (2024) Stable diffusion 3.5. Note: https://stability.ai/news/introducing-stable-diffusion-3-5Model weights: https://huggingface.co/stabilityai/stable-diffusion-3.5-medium Cited by: §4.
- B. Wallace, M. Dang, R. Rafailov, L. Zhou, A. Lou, S. Purushwalkam, S. Ermon, C. Xiong, S. Joty, and N. Naik (2024) Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8228–8238. Cited by: §1.
- F. Wang and Z. Yu (2025) Coefficients-preserving sampling for reinforcement learning with flow matching. arXiv preprint arXiv:2509.05952. Cited by: Appendix A, §C.4, Figure 1, Figure 1, §1, §2.1, §4.
- J. Wang, J. Liang, J. Liu, H. Liu, G. Liu, J. Zheng, W. Pang, A. Ma, Z. Xie, X. Wang, et al. (2025) Grpo-guard: mitigating implicit over-optimization in flow matching via regulated clipping. arXiv preprint arXiv:2510.22319. Cited by: Figure 1, Figure 1, §1, §1, §4.
- X. Wu, Y. Hao, K. Sun, Y. Chen, F. Zhu, R. Zhao, and H. Li (2023) Human preference score v2: a solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341. Cited by: §4.
- S. Xue, C. Ge, S. Zhang, Y. Li, and Z. Ma (2025a) Advantage weighted matching: aligning rl with pretraining in diffusion models. arXiv preprint arXiv:2509.25050. Cited by: §1.
- Z. Xue, J. Wu, Y. Gao, F. Kong, L. Zhu, M. Chen, Z. Liu, W. Liu, Q. Guo, W. Huang, et al. (2025b) Dancegrpo: unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818. Cited by: §1.
- K. Zheng, H. Chen, H. Ye, H. Wang, Q. Zhang, K. Jiang, H. Su, S. Ermon, J. Zhu, and M. Liu (2026) DiffusionNFT: online diffusion reinforcement with forward process. In The Fourteenth International Conference on Learning Representations, Cited by: Figure 7, Figure 7, §G.3.1, §1, §4.
Appendix of Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models
Appendix A The Flow-DPPO Algorithm
We summarize the complete Flow-DPPO training procedure. The algorithm adopts the CPS sampling framework (Wang and Yu, 2025) for trajectory generation, uses group-relative advantage estimation, and applies the divergence-based mask during policy optimization.
Computational overhead. The divergence computation requires one additional forward pass of the velocity network to obtain at training time. However, this forward pass is already required for computing the log ratio, so the divergence comes at zero additional cost: it is simply the squared norm of a difference that is already computed.
Appendix B Policy Improvement Bound for Flow Models
We adapt the classical policy improvement theory (Kakade and Langford, 2002; Schulman et al., 2015) to the finite-horizon, undiscounted setting of flow model denoising, following the approach of Qi et al. (2026) for the LLM regime. We use the MDP notation introduced in Section˜2.1: decision steps indexed by , states , actions , and terminal reward .
B.1 Proof of Performance Difference Identity
Proof [Proof of Theorem˜1] We begin by expressing the performance difference via its definition. Since the reward is only a function of the terminal state and the prompt , we have:
where the integral is over all trajectories (we omit the deterministic transition structure for notational clarity).
The core of the proof is the telescoping identity for the difference in trajectory probabilities. Since , we apply the algebraic identity :
Substituting into the performance difference and converting to an expectation under :
We decompose this expression by adding and subtracting the term where the future ratio product is set to 1:
This completes the proof.
B.2 Proof of Policy Improvement Bound
Lemma 5(Bound on Trajectory-Level TV Divergence).
Let and be two policies for the flow model MDP. Let and denote the distributions over future sub-trajectories starting from state . Then:
where the expectation is over states visited under starting from .
Proof Let and , where . We have:
Applying the telescoping identity (which follows from the triangle inequality) and integrating:
For each term indexed by , integrating out the future actions yields 1 (since is normalized), leaving:
The inner integral is , and the outer integral defines an expectation over states under policy . Thus:
Proof [Proof of Theorem˜2] From Theorem˜1, we start with the exact performance difference identity:
Our goal is to upper-bound . We begin by bounding the reward by its maximum absolute value :
| (19) | ||||
The inner expectation over future sub-trajectories is exactly twice the TV divergence between future trajectory distributions:
Applying Lemma˜5 and bounding each term by :
Substituting back into Eq. (19):
Evaluating the sum: . Therefore:
Substituting into the performance difference identity yields the desired bound:
This completes the proof.
B.3 A Tighter Policy Improvement Bound
The quadratic dependence on the horizon in Theorem˜2 can be overly pessimistic. By exploiting the fact that , we derive a tighter bound that is linear in .
Starting from the intermediate step in Eq. (19), the inner expectation is . Instead of applying Lemma˜5, we directly use the universal bound :
Combining both bounds, the policy improvement satisfies the composite guarantee:
where . The quadratic bound is tighter for small policy changes, while the linear bound is tighter for larger updates or longer horizons.
B.4 Connection to Gaussian Per-Step Divergence
For the Gaussian policies in Eq. (4), and , the TV divergence admits the closed form:
where is the standard normal CDF. Since is strictly monotonically increasing, the TV constraint is equivalent to:
This formally establishes that the Flow-DPPO mask, which blocks updates when , implements a trust-region constraint equivalent (up to a monotone rescaling) to constraining the per-step TV divergence. The policy improvement bound (Theorem˜2) thus provides a rigorous theoretical guarantee for Flow-DPPO: by enforcing a per-step divergence threshold, the penalty term remains controlled, ensuring monotonic policy improvement.
Appendix C KL Divergence Between Gaussian Policies
In this section, we derive the KL divergence between old and new policies in flow models and establish its connection to the TV divergence used in the policy improvement bound.
C.1 General Gaussian KL Divergence
Let and be two isotropic Gaussians in with the same covariance. The KL divergence is:
| (20) |
Note that this is symmetric in the means: when the covariances are identical.
C.2 Connection Between KL and TV in the Gaussian Setting
For the same pair of Gaussians, the TV divergence is:
Since both KL and TV are monotone functions of the single quantity , thresholding one is equivalent to thresholding the other. Specifically, the constraint is equivalent to , which in turn is equivalent to . This shows that the squared distance used in our mask is a unified divergence measure equivalent (up to monotone transformations) to both KL and TV divergences.
C.3 Application to Flow-SDE
For Flow-SDE (Eq. (2)), the per-step policy is where:
The difference in means is:
Substituting into Eq. (20) with :
| (21) |
C.4 Application to CPS
For CPS (Eq. (3)), the policy mean is and the variance is . Using and , the difference in means is:
Let . Then:
| (22) |
In previous work (Wang and Yu, 2025), the normalization is dropped for numerical stability, reducing the divergence to . We instead retain the full normalization in Eq. (22): because shrinks at later denoising steps, the factor amplifies the divergence where small velocity changes most affect the output, yielding a tighter constraint that prevents distribution collapse.
Appendix D Ratio Variance Analysis
We provide a detailed analysis of the variance of the log-ratio in flow models.
From Eq. (13), , where and . It follows that:
Thus . When the KL is moderate (e.g., ), the standard deviation of the log-ratio is , meaning that individual log-ratio samples fluctuate by around the mean of . In terms of the ratio itself, this corresponds to roughly a multiplicative spread.
Implication for clipping. With a typical clip parameter (i.e., clip range ), the log-clip range is . Comparing this narrow range with the log-ratio standard deviation of , we see that even for modest KL values, a significant fraction of samples will be clipped purely due to noise, not because the true divergence is excessive. This provides rigorous justification for replacing ratio-based clipping with direct divergence measurement.
Appendix E Towards a Predictive Divergence Mask
We recall the asymmetric mask in Flow-DPPO (Eq. (18)). The mask blocks the gradient (i.e., ) when two conditions hold simultaneously: (i) the divergence already exceeds the trust-region threshold, and (ii) a directional condition signals that the optimization would push the policy further away from . Concretely, the directional condition triggers when (the gradient would further increase an already-elevated ratio) or (the gradient would further decrease an already-reduced ratio). These two cases can be compactly unified as:
| (23) |
While this design is effective in practice, the directional indicator is a heuristic proxy for whether the upcoming gradient step will increase the divergence. In a ratio-based trust region (e.g., PPO clipping), this sign test is well-motivated: directly reflects the deviation of the single-sample Monte Carlo estimate of the importance ratio, so the sign of faithfully indicates whether the surrogate objective would drive the ratio further from unity. However, in a divergence-based trust region where the constraint is on , the connection is less direct. The ratio is a stochastic quantity evaluated at a single sampled action, whereas measures a distributional distance that integrates over all actions. A positive does not guarantee that the gradient step will increase , nor does a negative value guarantee a decrease.
In this section we exploit the Gaussian structure of flow model policies to derive a more principled masking criterion. We first predict how a single gradient step changes (§E.1), obtaining a closed-form expression that decomposes into a first-order directional term and a second-order magnitude term. The sign of the first-order term yields an exact directional criterion , which recovers the current sign test in the small-divergence regime but reveals a correction when the policy has already drifted. The full expression further accounts for the step size and gradient magnitude, leading to a predictive mask (§E.2) that directly forecasts whether the post-update divergence will exceed .
E.1 Predicting Post-Update Divergence
Fix a denoising step with state and suppress the time index for brevity. Write , , , and . The sampled action is with .
We derive how a single gradient step on the surrogate objective changes the divergence . The policy gradient with respect to is:
With effective learning rate , the updated mean is . Let . The predicted post-update divergence is:
| (24) |
From the ratio decomposition (Eq. (13)), , which gives . The first-order term thus simplifies to . The three terms in Eq. (24) have clear interpretations: (1) the current divergence ; (2) a first-order term whose sign determines whether the gradient step increases or decreases the divergence; (3) a non-negative second-order term that grows with the step size and gradient magnitude , always contributing positively to .
The first-order directional criterion. The direction of divergence change is mainly determined by the sign of the first-order term. Since and , this sign equals:
| (25) |
When this is positive, the gradient step increases ; when negative, it decreases . Equivalently, this is the sign of the inner product , confirming that the surrogate gradient projects onto the divergence-increasing direction.
Recovery of the current mask. In the small-divergence regime (which is the typical operating range when the trust region is effective), the correction and the criterion simplifies to . Since , this is equivalent to , which is exactly the directional condition in Eq. (23). Thus, the current Flow-DPPO mask implements the correct first-order divergence-increasing criterion in this regime.
The correction term. When is non-negligible (i.e., the policy has already drifted appreciably), the true divergence-change direction is rather than . The subtracted term shifts the decision boundary: a sample must have (rather than merely ) before the positive-advantage gradient is classified as divergence-increasing. Intuitively, when the policy has already moved away from , a moderately elevated ratio does not necessarily push it further; only sufficiently large ratios do. This yields a first natural refinement of the mask: replacing with as the directional indicator, which we call the first-order predictive mask:
| (26) |
This mask uses only quantities already computed during training (, , ) and requires no additional hyperparameters beyond the existing threshold .
E.2 The Predictive Mask
Based on Eq. (24), we define the (full) predictive mask that blocks updates whenever the predicted post-update divergence would exceed :
| (27) |
Comparison with the first-order mask. The first-order mask (Eq. (26)) only considers the direction of divergence change and still relies on the separate threshold condition . The full predictive mask unifies both into a single inequality: whether the gradient increases or decreases divergence is automatically encoded in the predicted value , and the threshold comparison is applied to the predicted (rather than current) divergence. This has two consequences. First, when , even a divergence-increasing step may be permitted if the predicted remains below . Second, when is close to , the second-order term may push above even when the first-order direction is “safe” (i.e., the first-order mask would not fire), correctly blocking large gradient steps near the trust-region boundary.
Recovery of the existing mask. In the limit , the second-order term vanishes and reduces to requiring that the first-order direction is positive and . Combined with the small-divergence approximation (), this exactly recovers the current Flow-DPPO mask (Eq. (23)).
E.3 Discussion on Mask Variants
Hierarchy of masks. The three masks form a natural hierarchy of increasing fidelity:
The current mask is the cheapest (no additional computation) and suffices when the trust region keeps small throughout training. The first-order mask refines the directional decision with zero additional hyperparameters. The full predictive mask additionally requires an effective learning rate estimate but provides quantitative divergence prediction.
Local approximation. The analysis treats as a free vector, whereas in practice it is the output of a neural network. The actual change in is coupled to changes at all other inputs through shared parameters. The predictive mask is thus a local approximation that is most accurate when the effective learning rate is small and the network Jacobian is approximately preserved across one step.
We leave empirical validation of the predictive masks to future work. The key contribution of this analysis is twofold: it provides a theoretical justification for the existing asymmetric condition (showing it is the correct first-order criterion in the small-divergence regime), and it charts a principled path toward more refined trust-region enforcement that exploits the Gaussian structure of flow model policies.
Appendix F Experimental Details
F.1 Computational Resources.
All experiments are conducted on NVIDIA H20 96GB GPUs. The main results in Table 1 require approximately 90K GPU hours in total (across SD3.5, FLUX2-klein-base-9B, and FLUX1-dev with all methods and reward configurations). Including all ablation studies, multi-epoch experiments, and auxiliary runs, the overall computational cost for all experiments reported in this paper is approximately 140K GPU hours.
F.2 Hyperparameters.
LoRA is used for all models. We use LoRA and for SD3.5, and for FLUX2-9B and FLUX.1-dev. The learning rate is set to for all models aligning to previous works. We set the training resolution to , number of denoising steps to 10 for SD3.5 and 14 for FLUX2-9B.
For GRPO setting, we use group size 16 and number of groups 64 per epoch for all methods. The PPO clip threshold is set to for Flow-GRPO and Flow-CPS, and for GRPO-Guard, following the official recommendation. The thresholds for KL-clipping are set to for Flow-DPPO and for Flow-DPPO+CPS due to their different KL-scaling factors. We applied the stragegy proposed in MixGRPO (Li et al., 2025) on all baselines and proposed methods for faster convergence and better performance. Specifically, we mix ODE and SDE sampling and randomly select 3 steps out of first half of the denoising steps for SDE sampling. The noise level for SDE sampling ( in CPS sampling) is set to .
For Diffusion-NFT, we follow the official implementation for SD3.5 for the rest of the hyperparameters, such as EMA schedule.
Appendix G Additional Experimental Results
G.1 Additional Training Curves
We provide the training curves on SD3.5 for the single-reward setting in Figure 7 (the multi-reward setting is in Figure 4 in the main body). We also provide the FLUX2-9B multi-reward training curves in Figure 8 and FLUX.1-dev in Figure 9.
We additionally provide training curves on FLUX.1-dev in Figure 9.
G.2 KL Divergence Curves
Figure 12 visualises the per-step KL divergence between the current and reference (pre-trained) model across all six training settings and two SDE schedules. The corresponding end-of-training values are reported in Table 2 of the main body.
G.3 Ablation Studies
G.3.1 Classifier-Free Guidance
Previous works found that CFG heavily affects the training convergence and performance (Zheng et al., 2026). Here, we study the effect of CFG on the training of Flow-DPPO on FLUX2-9B, as shown in Figure 10, where the CFG scale is set to 4.0 following the official recommendation. With CFG, Flow-DPPO variants still achieve state-of-the-art performance on the training reward (GenEval2) and mitigate catastrophic forgetting on the out-of-domain prompts, consistent with the observations in previous discussions. This shows that the divergence-based mask is robust under CFG and continues to deliver strong performance.
G.3.2 Reference KL Regularization Strength
We ablate the strength of the regularization term (controlled by ) on FLUX2-klein-base-9B under the multi-reward GDPO setting with CPS scheduling. Figure 11 shows the training reward curves and Figure 13 shows the KL divergence from the pretrained model. A moderate regularization strength () further mitigates early-stage reward hacking on auxiliary objectives (PickScore, HPSv2, etc.), thereby balancing the gradients across rewards and yielding an additional improvement in final GenEval2 performance over the unregularized baseline, without degrading end-of-training performance on any individual reward.
| FLUX2-9B | SD3.5 | FLUX.1-dev | ||||
|---|---|---|---|---|---|---|
| Method | Single | Multi | +CFG | Single | Multi | Single |
| Flow-GRPO | 84.5 | 46.8 | 54.6 | 56.6 | 39.9 | 87.8 |
| Flow-CPS | 82.7 | 47.1 | 89.0 | 74.8 | 44.6 | 91.2 |
| GRPO-Guard | 82.8 | 49.0 | 78.8 | 85.8 | 47.8 | 87.6 |
| Diffusion-NFT | – | 47.3 | – | 64.5 | 42.5 | – |
| Flow-DPPO | 85.1 | 57.7 | 87.4 | 78.9 | 48.1 | 90.7 |
| Flow-DPPO + CPS | 92.6 | 55.2 | 91.0 | 84.1 | 51.6 | 91.6 |
G.4 Quantitative Summary on GenEval2
To complement the per-setting training-curve figures above, Tables 4 and 3 report the end-of-training Soft TIFA score on GenEval2 for each method. Table 4 additionally reports end-of-training ancillary CLIP, PickScore, and HPSv2 rewards on both the in-domain GenEval2 prompt set and the held-out out-of-domain PickScore validation prompts, contextualising both SD3.5-medium and FLUX2-klein-base-9B by stacking six blocks: published reference numbers for state-of-the-art text-to-image systems, the corresponding pretrained-baseline scores (no RL), and the five RL fine-tuning algorithms applied to each base model under both the single-reward (GenEval2-only) and multi-reward (GenEval2 + CLIP + PickScore + HPSv2) configurations. Table 3 then expands the per-method Soft TIFA comparison to all five training settings reported in this paper.
| In-Domain (GenEval2) | Out-of-Domain (PickScore) | ||||||
| Model | GenEval2 | CLIP | PickScore | HPSv2 | CLIP | PickScore | HPSv2 |
| State-of-the-Art T2I Models | |||||||
| SD3.5-large | 22.8 | – | – | – | – | – | – |
| Bagel + CoT | 23.1 | – | – | – | – | – | – |
| Qwen-Image | 33.8 | – | – | – | – | – | – |
| Gemini 2.5 Flash Image | 44.6 | – | – | – | – | – | – |
| Pretrained baselines (before RL) | |||||||
| SD3.5-medium | 12.4 | 0.250 | 21.00 | 0.213 | 0.244 | 19.99 | 0.210 |
| FLUX2-klein-base-9B | 25.4 | 0.281 | 20.92 | 0.228 | 0.254 | 20.05 | 0.230 |
| FLUX.1-dev | 23.3 | 0.297 | 23.26 | 0.315 | 0.276 | 21.91 | 0.304 |
| SD3.5-medium, single-reward RL fine-tuning | |||||||
| Flow-GRPO | 56.6 | 0.297 | 21.21 | 0.219 | 0.252 | 19.33 | 0.206 |
| Flow-CPS | 74.8 | 0.313 | 21.68 | 0.235 | 0.260 | 19.94 | 0.220 |
| GRPO-Guard | 85.8 | 0.328 | 22.03 | 0.252 | 0.265 | 19.94 | 0.214 |
| Diffusion-NFT | 64.5 | 0.307 | 21.69 | 0.251 | 0.262 | 20.24 | 0.239 |
| Flow-DPPO | 78.9 | 0.319 | 22.06 | 0.263 | 0.265 | 20.45 | 0.253 |
| Flow-DPPO + CPS | 84.1 | 0.316 | 21.99 | 0.262 | 0.272 | 20.50 | 0.246 |
| SD3.5-medium, multi-reward RL fine-tuning | |||||||
| Flow-GRPO | 39.9 | 0.358 | 25.09 | 0.399 | 0.273 | 22.07 | 0.349 |
| Flow-CPS | 44.6 | 0.359 | 25.51 | 0.407 | 0.265 | 22.08 | 0.343 |
| GRPO-Guard | 47.8 | 0.353 | 25.64 | 0.409 | 0.272 | 22.32 | 0.354 |
| Diffusion-NFT | 42.5 | 0.334 | 25.30 | 0.394 | 0.269 | 22.52 | 0.355 |
| Flow-DPPO | 48.1 | 0.345 | 25.63 | 0.409 | 0.273 | 22.58 | 0.360 |
| Flow-DPPO + CPS | 51.6 | 0.369 | 25.72 | 0.415 | 0.279 | 22.51 | 0.361 |
| FLUX2-klein-base-9B, single-reward RL fine-tuning | |||||||
| Flow-GRPO | 84.5 | 0.314 | 21.82 | 0.276 | 0.264 | 20.84 | 0.280 |
| Flow-CPS | 82.7 | 0.311 | 21.82 | 0.261 | 0.275 | 21.15 | 0.267 |
| GRPO-Guard | 82.8 | 0.312 | 20.52 | 0.210 | 0.230 | 18.45 | 0.167 |
| Flow-DPPO | 85.1 | 0.331 | 22.22 | 0.294 | 0.278 | 21.27 | 0.285 |
| Flow-DPPO + CPS | 92.6 | 0.315 | 21.97 | 0.279 | 0.265 | 20.79 | 0.272 |
| FLUX2-klein-base-9B, multi-reward RL fine-tuning | |||||||
| Flow-GRPO | 46.8 | 0.371 | 25.61 | 0.412 | 0.277 | 22.62 | 0.357 |
| Flow-CPS | 47.1 | 0.361 | 25.70 | 0.416 | 0.276 | 22.85 | 0.364 |
| GRPO-Guard | 49.0 | 0.375 | 25.27 | 0.411 | 0.269 | 21.99 | 0.349 |
| Diffusion-NFT | 47.3 | 0.336 | 24.87 | 0.389 | 0.274 | 22.47 | 0.351 |
| Flow-DPPO | 57.7 | 0.364 | 25.76 | 0.418 | 0.282 | 22.90 | 0.368 |
| Flow-DPPO + CPS | 55.2 | 0.386 | 26.15 | 0.427 | 0.287 | 22.97 | 0.370 |
| FLUX.1-dev, single-reward RL fine-tuning | |||||||
| Flow-GRPO | 87.8 | 0.331 | 23.03 | 0.311 | 0.291 | 21.85 | 0.311 |
| Flow-CPS | 91.2 | 0.328 | 23.20 | 0.317 | 0.288 | 21.98 | 0.307 |
| GRPO-Guard | 87.6 | 0.333 | 22.69 | 0.293 | 0.286 | 21.03 | 0.276 |
| Flow-DPPO | 90.7 | 0.331 | 23.15 | 0.323 | 0.290 | 21.60 | 0.300 |
| Flow-DPPO + CPS | 91.6 | 0.331 | 23.29 | 0.322 | 0.289 | 21.91 | 0.305 |