Summary (Overview)

• Flow-DPPO replaces the ratio clipping in PPO-style RL fine-tuning of flow matching models with a direct divergence-based mask, enabling more precise trust-region enforcement.
• The key insight is that per‑step policies in flow models are Gaussian with fixed variance, allowing exact and cheap computation of the KL divergence between old and new policies (no Monte‑Carlo noise).
• An asymmetric divergence mask blocks updates only when they simultaneously move away from the old policy and exceed a divergence threshold, preserving beneficial corrective updates.
• Experiments across SD3.5, FLUX.1‑dev, and FLUX2‑9B show that Flow‑DPPO achieves higher rewards, better KL‑proximal efficiency, reduced catastrophic forgetting, balanced multi‑objective optimization, and stable multi‑epoch training compared to Flow‑GRPO, Flow‑CPS, GRPO‑Guard, and Diffusion‑NFT.
• The method provides a theoretically grounded policy improvement bound for flow models, and the Gaussian structure makes divergence‑based trust regions strictly more natural than ratio clipping.

Introduction and Theoretical Foundation

Background

Reinforcement learning (RL) has become a key paradigm for aligning generative models with downstream objectives, first in language (DPO, GRPO) and recently in image/video generation using flow matching models. Methods like Flow‑GRPO and DanceGRPO transform deterministic ODE sampling into stochastic SDE trajectories and apply PPO‑style ratio clipping to enforce trust‑region optimization.

Theoretical Foundation

Trust‑region methods originate from TRPO, which guarantees monotonic improvement when the policy update stays within a divergence‑defined region. PPO approximates this with ratio clipping. The paper adapts trust‑region theory to the finite‑horizon, terminal‑reward MDP used for flow model fine‑tuning.

Theorem 1 (Performance Difference Identity for Flow Models):
For two policies $\pi_\theta$ and $\pi_{\theta_\text{old}}$,

$$J(\pi_\theta)-J(\pi_{\theta_\text{old}}) = L'_{\theta_\text{old}}(\pi_\theta) - \Delta(\pi_{\theta_\text{old}},\pi_\theta),$$

where $L'_{\theta_\text{old}}(\pi_\theta)$ is a first‑order surrogate and $\Delta$ captures higher‑order interactions.

Theorem 2 (Policy Improvement Bound for Flow Models):

$$J(\pi_\theta)-J(\pi_{\theta_\text{old}}) \ge L'_{\theta_\text{old}}(\pi_\theta) - 2\xi(K-1)(K-2)\cdot\bigl(D^{\max}_{\text{TV}}(\pi_{\theta_\text{old}}\|\pi_\theta)\bigr)^2,$$

where $K$ is the number of steps, $\xi$ is the max absolute reward, and $D^{\max}_{\text{TV}}$ is the maximum per‑step Total Variation divergence. This justifies constraining the per‑step divergence.

Gaussian policies in flow models:
Both the Flow‑SDE and CPS samplers induce a Gaussian per‑step policy:

$$p_\theta(x_{t-\Delta t}|x_t,t,c)=\mathcal{N}\bigl(x_{t-\Delta t};\mu_\theta(x_t,t,c),\sigma^2(t)I\bigr).$$

This enables exact computation of the KL divergence (see Methodology).

Methodology

Exact KL Divergence for Flow Models

Because both $\pi_{\theta_\text{old}}$ and $\pi_\theta$ are Gaussians with equal covariance $\sigma^2 I$, the KL divergence is:

$$D_{\text{KL}}\bigl(\pi_{\theta_\text{old}}(\cdot|x_t)\|\pi_\theta(\cdot|x_t)\bigr)=\frac{\|\mu_{\theta_\text{old}}(x_t,t)-\mu_\theta(x_t,t)\|^2}{2\sigma^2}.$$

For Flow‑SDE:

$$D^{\text{SDE}}_{\text{KL}}(\pi_{\theta_\text{old}}\|\pi_\theta)=\frac{\Delta t}{2}\left(\frac{\sigma_t(1-t)}{2t}+\frac{1}{\sigma_t}\right)^2\|v_\theta(x_t,t)-v_{\theta_\text{old}}(x_t,t)\|^2.$$

For CPS ($\sigma_{\text{CPS}}=(t-\Delta t)\sin(\eta\pi/2)$):

$$D^{\text{CPS}}_{\text{KL}}(\pi_{\theta_\text{old}}\|\pi_\theta)=\frac{\|\mu^{\text{CPS}}_\theta(x_t,t)-\mu^{\text{CPS}}_{\theta_\text{old}}(x_t,t)\|^2}{2(t-\Delta t)^2\sin^2(\eta\pi/2)}.$$

Pitfalls of Ratio Clipping

The per‑step importance ratio $r_t^i$ is a noisy single‑sample estimate of $2D_{\text{TV}}$ (Eq. 11). In high‑dimensional continuous action spaces, the log‑ratio is dominated by noise:

$$\log r_t^i(\theta)=\frac{\epsilon^\top d}{\sigma}-\frac{\|d\|^2}{2\sigma^2},\qquad d=\mu_\theta-\mu_{\theta_\text{old}},\ \epsilon\sim\mathcal{N}(0,I).$$

The noise term has standard deviation $\|d\|/\sigma$, of the same order as the signal $-\|d\|^2/(2\sigma^2)$. This leads to spurious clipping decisions that depend on the random noise rather than the true policy divergence.

The Flow‑DPPO Mask

Flow‑DPPO replaces ratio clipping with a divergence‑based mask in the objective:

$$\mathcal{L}_{\text{Flow-DPPO}}(\theta)=\mathbb{E}\left[\frac{1}{G}\sum_{i=1}^G\frac{1}{T}\sum_{t=0}^{T-1}\Bigl(M_t^i\cdot r_t^i(\theta)\cdot\hat{A}^i-\beta D_{\text{KL}}(\pi_\theta\|\pi_{\text{ref}})\Bigr)\right],$$

where the mask $M_t^i$ is:

$$M_t^i = \begin{cases} 0, & \text{if } \hat{A}^i>0 \text{ and } r_t^i>1 \text{ and } D_t>\delta \\ & \text{or } \hat{A}^i<0 \text{ and } r_t^i<1 \text{ and } D_t>\delta, \\ 1, & \text{otherwise,} \end{cases}$$

with $D_t\equiv D_{\text{KL}}(\pi_{\theta_\text{old}}(\cdot|x_t^i)\|\pi_\theta(\cdot|x_t^i))$ and $\delta$ a divergence threshold.
The mask is asymmetric: it blocks updates only when the gradient is moving the policy away from the old policy and the divergence has exceeded $\delta$. Updates that move the policy toward the old policy (e.g., $\hat{A}^i>0, r_t^i<1$) are never blocked, enabling rapid recovery from overshooting.

Empirical Validation / Results

Experimental Setup

• Base models: Stable Diffusion 3.5 Medium, FLUX2‑klein‑base‑9B, FLUX.1‑dev.
• Baselines: Flow‑GRPO, Flow‑CPS, GRPO‑Guard, Diffusion‑NFT.
• Variants: Flow‑DPPO (SDE sampling) and Flow‑DPPO+CPS (CPS‑scheduled SDE).

Key Findings

• Higher rewards with better KL‑proximal efficiency: Flow‑DPPO achieves higher reward scores while maintaining lower KL divergence from the reference policy, indicating more efficient trust‑region enforcement.
• Alleviation of catastrophic forgetting: The asymmetric mask preserves the pretrained model’s capabilities (e.g., visual quality, compositional accuracy) better than ratio‑clipping methods.
• Balanced multi‑objective optimization: Flow‑DPPO mitigates reward hacking by not over‑optimizing for a single objective; it produces outputs with competitive compositional accuracy and notably less image quality degradation (see Figure 1 qualitative comparison).
• Stable multi‑epoch training: Where ratio clipping leads to performance degradation over multiple gradient steps due to policy staleness and noisy clipping, Flow‑DPPO maintains stable improvement across epochs, enabling higher sample efficiency.

Table 1: Comparison of methods on key metrics. Numbers are illustrative; the paper reports consistent improvements across all base models and evaluation settings.

(Full quantitative results with exact scores, FID, CLIP, GenEval2 etc. are provided in the paper.)

Theoretical and Practical Implications

• Theoretical: The policy improvement bound (Theorem 2) and the exact divergence computation (Remark 3) provide a rigorous foundation for trust‑region methods in flow model fine‑tuning. The Gaussian policy structure makes divergence‑based constraints not only feasible but also cheaper and more accurate than the approximate reductions needed in LLMs (DPPO).
• Practical: Flow‑DPPO can be directly integrated into existing flow model RL pipelines (Flow‑SDE, CPS) with minimal overhead (only the divergence mask computation, which reuses the two forward passes already performed). The asymmetric design automatically handles off‑policy updates and prevents catastrophic forgetting without extra regularization tuning.
• Generality: The method is applicable to any flow matching model whose sampler induces a Gaussian per‑step policy, including rectified flow and various SDE schedulers. It also extends to multi‑objective reward settings by maintaining balanced optimization.

Conclusion

Flow‑DPPO replaces the noisy ratio clipping in PPO‑style RL fine‑tuning of flow matching models with an exact, deterministic divergence mask. By leveraging the Gaussian structure of the induced policies, the KL divergence between old and new policies can be computed cheaply and exactly. The asymmetric mask blocks updates only when they move the policy away from the old policy and exceed a divergence threshold, while allowing beneficial corrective updates. Extensive experiments on multiple base models demonstrate that Flow‑DPPO achieves higher rewards, better KL‑proximal efficiency, reduced catastrophic forgetting, balanced multi‑objective optimization, and stable multi‑epoch training compared to existing ratio‑clipping methods.

Future directions: The paper suggests that the divergence‑based approach could be extended to other generative models with reparameterizable policies (e.g., diffusion models with explicit noise schedules), and that the theoretical bounds could be tightened further. The code is publicly available at the provided repository.