RL for Diffusion Models

Diffusion large language models are a compelling alternative to autoregressive models, yet existing RL methods for diffusion treat all denoising steps as equally important and rely on biased, high-variance likelihood estimates. We identify two fundamental weaknesses: the absence of temporal credit assignment across the denoising trajectory, and the systematic bias of mean-field likelihood estimates used for policy optimization. To address these, we propose Denoising-Aware Credit Assignment for GRPO (DACA-GRPO), a lightweight, plug-and-play enhancement for any GRPO-style trainer. DACA-GRPO introduces two complementary mechanisms: Denoising Progress Scores, which extract per-token importance weights from intermediate predictions at no additional forward cost, and Stratified Masking Likelihood, which partitions token positions into strata so that each token is predicted with most of the sequence as context, reducing the mean-field bias. Applied on top of three GRPO base methods, DACA-GRPO achieves consistent improvements across seven benchmarks spanning mathematical reasoning, code generation, constraint satisfaction, and constrained generation, with gains of up to 5.6pp on math reasoning, 7.4pp on code generation, 36.3pp on constraint satisfaction, and 5.9pp on JSON schema adherence.

Reinforcement learning fine-tuning has become the dominant approach for aligning diffusion models with human preferences. However, assessing images is intrinsically a multi-dimensional task, and multiple evaluation criteria need to be optimized simultaneously. Existing practice deal with multiple rewards by training one specialist model per reward, optimizing a weighted-sum reward $R(x)=\sum_k w_k R_k(x)$, or sequentially fine-tuning with a hand-crafted stage schedule. These approaches either fail to produce a unified model that can be jointly trained on all rewards or necessitates heavy manually tuned sequential training. We find that the failure stems from using a naive weighted-sum reward aggregation. This approach suffers from a sample-level mismatch because most rollouts are specialist samples, highly informative for certain reward dimensions but irrelevant for others; consequently, weighted summation dilutes their supervision. To address this issue, we propose MARBLE (Multi-Aspect Reward BaLancE), a gradient-space optimization framework that maintains independent advantage estimators for each reward, computes per-reward policy gradients, and harmonizes them into a single update direction without manually-tuned reward weighting, by solving a Quadratic Programming problem. We further propose an amortized formulation that exploits the affine structure of the loss used in DiffusionNFT, to reduce the per-step cost from K+1 backward passes to near single-reward baseline cost, together with EMA smoothing on the balancing coefficients to stabilize updates against transient single-batch fluctuations. On SD3.5 Medium with five rewards, MARBLE improves all five reward dimensions simultaneously, turns the worst-aligned reward's gradient cosine from negative under weighted summation in 80% of mini-batches to consistently positive, and runs at 0.97X the training speed of baseline training.

This paper studies full-body 3D human motion recovery from head-mounted device signals. Existing diffusion-based methods often rely on global distribution matching, leading to local joint reconstruction errors. We propose MotionGRPO, a novel framework leveraging reinforcement learning post-training to inject fine-grained guidance into the diffusion process. Technically, we model diffusion sampling as a Markov decision process optimized via Group Relative Policy Optimization (GRPO). To this end, we introduce a hybrid reward mechanism that combines a learned conditioned perceptual model for global visual plausibility and explicit constraints for local joint precision. Our key technical insight is that policy optimization in diffusion-based recovery suffers from vanishing gradients due to limited intra-group sample diversity. To address this, we further introduce a noise-injection strategy that explicitly increases sample variance and stabilizes learning. Extensive experiments demonstrate that MotionGRPO achieves state-of-the-art performance with superior visual fidelity

While Reinforcement Learning from Human Feedback (RLHF) has become a pivotal paradigm for text-to-image generation, its application to image editing remains largely unexplored. A key bottleneck is the lack of a robust general reward model for all editing tasks. Existing edit reward models usually give overall scores without detailed checks, ignoring different instruction requirements and causing biased rewards. To address this, we argue that the key is to move from a simple scorer to a reasoning verifier. We introduce Edit-R1, a framework that builds a chain-of-thought (CoT) verifier-based reasoning reward model (RRM) and then leverages it for downstream image editing. The Edit-RRM breaks instructions into distinct principles, evaluates the edited image against each principle, and aggregates these checks into an interpretable, fine-grained reward. To build such an RRM, we first apply supervised fine-tuning (SFT) as a ``cold-start'' to generate CoT reward trajectories. Then, we introduce Group Contrastive Preference Optimization (GCPO), a reinforcement learning algorithm that leverages human pairwise preference data to reinforce our pointwise RRM. After building the RRM, we use GRPO to train editing models with this non-differentiable yet powerful reward model. Extensive experiments demonstrate that our Edit-RRM surpasses powerful VLMs such as Seed-1.5-VL and Seed-1.6-VL as an editing-specific reward model, and we observe a clear scaling trend, with performance consistently improving from 3B to 7B parameters. Moreover, Edit-R1 delivers gains to editing models like FLUX.1-kontext, highlighting its effectiveness in enhancing image editing.

Diffusion models offer superior generation quality at the expense of extensive sampling steps. Distillation methods, with Distribution Matching Distillation (DMD) as a popular example, can mitigate this issue, but performance degradation remains pronounced when sampling steps are limited. Reinforcement learning (RL) has been leveraged to improve the few-step generation quality during distillation, with the potential to even surpass the performance of the teacher model. However, existing approaches are combinatorial in nature, merely integrating an RL process with the distillation process, which introduces unnecessary complexities. To address this gap, we propose AdvDMD, a method that seamlessly unifies DMD distillation and RL. Specifically, AdvDMD employs the adversarially trained discriminator from DMD2 as the reward model, which assigns low scores to generated images and high scores to real ones. It is trained on both intermediate and final states of the denoising process and updated online with the distilled model, enabling a holistic supervision of the sampling trajectories and mitigating reward hacking. We adopt a unified SDE backward simulation and a different training schedule for DMD and RL to enable a more stable and efficient training. Experimental results demonstrate that the 4-step AdvDMD outperforms the original 40-step model for SD3.5 on DPG-Bench, while achieving significant performance gains for SD3 on the GenEval. On Qwen-Image, our 2-step AdvDMD achieves superior performance over TwinFlow.

While large-scale video diffusion models have demonstrated impressive capabilities in generating high-resolution and semantically rich content, a significant gap remains between their pretraining performance and real-world deployment requirements due to critical issues such as prompt sensitivity, temporal inconsistency, and prohibitive inference costs. To bridge this gap, we propose a comprehensive post-training framework that systematically aligns pretrained models with user intentions through four synergistic stages: we first employ Supervised Fine-Tuning (SFT) to transform the base model into a stable instruction-following policy, followed by a Reinforcement Learning from Human Feedback (RLHF) stage that utilizes a novel Group Relative Policy Optimization (GRPO) method tailored for video diffusion to enhance perceptual quality and temporal coherence; subsequently, we integrate Prompt Enhancement via a specialized language model to refine user inputs, and finally address system efficiency through Inference Optimization. Together, these components provide a systematic approach to improving visual quality, temporal coherence, and instruction following, while preserving the controllability learned during pretraining. The result is a practical blueprint for building scalable post-training pipelines that are stable, adaptable, and effective in real-world deployment. Extensive experiments demonstrate that this unified pipeline effectively mitigates common artifacts and significantly improves controllability and visual aesthetics while adhering to strict sampling cost constraints.

Aligning denoising generative models with human preferences or verifiable rewards remains a key challenge. While policy-gradient online reinforcement learning (RL) offers a principled post-training framework, its direct application is hindered by the intractable likelihoods of these models. Prior work therefore either optimizes an induced Markov decision process (MDP) over sampling trajectories, which is stable but inefficient, or uses likelihood surrogates based on the diffusion evidence lower bound (ELBO), which have so far underperformed on visual generation. Our key insight is that the ELBO-based approach can, in fact, be made both stable and efficient. By reducing surrogate variance and controlling gradient steps, we show that this approach can beat MDP-based methods. To this end, we introduce Variational GRPO (V-GRPO), a method that integrates ELBO-based surrogates with the Group Relative Policy Optimization (GRPO) algorithm, alongside a set of simple yet essential techniques. Our method is easy to implement, aligns with pretraining objectives, and avoids the limitations of MDP-based methods. V-GRPO achieves state-of-the-art performance in text-to-image synthesis, while delivering a $2\times$ speedup over MixGRPO and a $3\times$ speedup over DiffusionNFT.

Reinforcement Learning (RL) post-training has become the standard for aligning generative models with human preferences, yet most methods rely on a single scalar reward. When multiple criteria matter, the prevailing practice of ``early scalarization'' collapses rewards into a fixed weighted sum. This commits the model to a single trade-off point at training time, providing no inference-time control over inherently conflicting goals -- such as prompt adherence versus source fidelity in image editing. We introduce ParetoSlider, a multi-objective RL (MORL) framework that trains a single diffusion model to approximate the entire Pareto front. By training the model with continuously varying preference weights as a conditioning signal, we enable users to navigate optimal trade-offs at inference time without retraining or maintaining multiple checkpoints. We evaluate ParetoSlider across three state-of-the-art flow-matching backbones: SD3.5, FluxKontext, and LTX-2. Our single preference-conditioned model matches or exceeds the performance of baselines trained separately for fixed reward trade-offs, while uniquely providing fine-grained control over competing generative goals.

Some of the most performant reinforcement learning algorithms today can be prohibitively expensive as they use test-time scaling methods such as sampling multiple action candidates and selecting the best one. In this work, we propose FASTER, a method for getting the benefits of sampling-based test-time scaling of diffusion-based policies without the computational cost by tracing the performance gain of action samples back to earlier in the denoising process. Our key insight is that we can model the denoising of multiple action candidates and selecting the best one as a Markov Decision Process (MDP) where the goal is to progressively filter action candidates before denoising is complete. With this MDP, we can learn a policy and value function in the denoising space that predicts the downstream value of action candidates in the denoising process and filters them while maximizing returns. The result is a method that is lightweight and can be plugged into existing generative RL algorithms. Across challenging long-horizon manipulation tasks in online and batch-online RL, FASTER consistently improves the underlying policies and achieves the best overall performance among the compared methods. Applied to a pretrained VLA, FASTER achieves the same performance while substantially reducing training and inference compute requirements. Code is available at https://github.com/alexanderswerdlow/faster .

Common image editing tasks typically adopt powerful generative diffusion models as the leading paradigm for real-world content editing. Meanwhile, although reinforcement learning (RL) methods such as Diffusion-DPO and Flow-GRPO have further improved generation quality, efficiently applying Reinforcement Learning from Human Feedback (RLHF) to diffusion-based editing remains largely unexplored, due to a lack of scalable human-preference datasets and frameworks tailored to diverse editing needs. To fill this gap, we propose HP-Edit, a post-training framework for Human Preference-aligned Editing, and introduce RealPref-50K, a real-world dataset across eight common tasks and balancing common object editing. Specifically, HP-Edit leverages a small amount of human-preference scoring data and a pretrained visual large language model (VLM) to develop HP-Scorer--an automatic, human preference-aligned evaluator. We then use HP-Scorer both to efficiently build a scalable preference dataset and to serve as the reward function for post-training the editing model. We also introduce RealPref-Bench, a benchmark for evaluating real-world editing performance. Extensive experiments demonstrate that our approach significantly enhances models such as Qwen-Image-Edit-2509, aligning their outputs more closely with human preference.

Reinforcement learning, particularly Group Relative Policy Optimization (GRPO), has emerged as an effective framework for post-training visual generative models with human preference signals. However, its effectiveness is fundamentally limited by coarse reward credit assignment. In modern visual generation, multiple reward models are often used to capture heterogeneous objectives, such as visual quality, motion consistency, and text alignment. Existing GRPO pipelines typically collapse these rewards into a single static scalar and propagate it uniformly across the entire diffusion trajectory. This design ignores the stage-specific roles of different denoising steps and produces mistimed or incompatible optimization signals. To address this issue, we propose Objective-aware Trajectory Credit Assignment (OTCA), a structured framework for fine-grained GRPO training. OTCA consists of two key components. Trajectory-Level Credit Decomposition estimates the relative importance of different denoising steps. Multi-Objective Credit Allocation adaptively weights and combines multiple reward signals throughout the denoising process. By jointly modeling temporal credit and objective-level credit, OTCA converts coarse reward supervision into a structured, timestep-aware training signal that better matches the iterative nature of diffusion-based generation. Extensive experiments show that OTCA consistently improves both image and video generation quality across evaluation metrics.

Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-step models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.

Masked diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. While reinforcement learning (RL) methods have recently been adapted to dLLM fine-tuning, their objectives typically depend on sequence-level marginal likelihoods, which are intractable for masked diffusion models. To address this, we derive Discrete Tilt Matching (DTM), a likelihood-free method that recasts dLLM fine-tuning as state-level matching of local unmasking posteriors under reward tilting. DTM takes the form of a weighted cross-entropy objective with explicit minimizer, and admits control variates that improve training stability. On a synthetic maze-planning task, we analyze how DTM's annealing schedule and control variates affect training stability and prevent mode collapse. At scale, fine-tuning LLaDA-8B-Instruct with DTM yields strong gains on Sudoku and Countdown while remaining competitive on MATH500 and GSM8K.

Uniform Discrete Diffusion Model (UDM) has recently emerged as a promising paradigm for discrete generative modeling; however, its integration with reinforcement learning remains largely unexplored. We observe that naively applying GRPO to UDM leads to training instability and marginal performance gains. To address this, we propose UDM-GRPO, the first framework to integrate UDM with RL. Our method is guided by two key insights: (i) treating the final clean sample as the action provides more accurate and stable optimization signals; and (ii) reconstructing trajectories via the diffusion forward process better aligns probability paths with the pretraining distribution. Additionally, we introduce two strategies, Reduced-Step and CFG-Free, to further improve training efficiency. UDM-GRPO significantly improves base model performance across multiple T2I tasks. Notably, GenEval accuracy improves from $69\%$ to $96\%$ and PickScore increases from $20.46$ to $23.81$, achieving state-of-the-art performance in both continuous and discrete settings. On the OCR benchmark, accuracy rises from $8\%$ to $57\%$, further validating the generalization ability of our method. Code is available at https://github.com/Yovecent/UDM-GRPO.

High-level autonomous driving requires motion planners capable of modeling multimodal future uncertainties while remaining robust in closed-loop interactions. Although diffusion-based planners are effective at modeling complex trajectory distributions, they often suffer from stochastic instabilities and the lack of corrective negative feedback when trained purely with imitation learning. To address these issues, we propose RAD-2, a unified generator-discriminator framework for closed-loop planning. Specifically, a diffusion-based generator is used to produce diverse trajectory candidates, while an RL-optimized discriminator reranks these candidates according to their long-term driving quality. This decoupled design avoids directly applying sparse scalar rewards to the full high-dimensional trajectory space, thereby improving optimization stability. To further enhance reinforcement learning, we introduce Temporally Consistent Group Relative Policy Optimization, which exploits temporal coherence to alleviate the credit assignment problem. In addition, we propose On-policy Generator Optimization, which converts closed-loop feedback into structured longitudinal optimization signals and progressively shifts the generator toward high-reward trajectory manifolds. To support efficient large-scale training, we introduce BEV-Warp, a high-throughput simulation environment that performs closed-loop evaluation directly in Bird's-Eye View feature space via spatial warping. RAD-2 reduces the collision rate by 56% compared with strong diffusion-based planners. Real-world deployment further demonstrates improved perceived safety and driving smoothness in complex urban traffic.

Reinforcement learning (RL) has emerged as a powerful tool for aligning diffusion models with human preferences, typically by optimizing a single reward function under a KL regularization constraint. In practice, however, human preferences are inherently pluralistic, and aligned models must balance multiple downstream objectives, such as aesthetic quality and text-image consistency. Existing multi-objective approaches either rely on costly multi-objective RL fine-tuning or on fusing separately aligned models at denoising time, but they generally require access to reward values (or their gradients) and/or introduce approximation error in the resulting denoising objectives. In this paper, we revisit the problem of RL fine-tuning for diffusion models and address the intractability of identifying the optimal policy by introducing a step-level RL formulation. Building on this, we further propose Multi-objective Step-level Denoising-time Diffusion Alignment (MSDDA), a retraining-free framework for aligning diffusion models with multiple objectives, obtaining the optimal reverse denoising distribution in closed form, with mean and variance expressed directly in terms of single-objective base models. We prove that this denoising-time objective is exactly equivalent to the step-level RL fine-tuning, introducing no approximation error. Moreover, we provide numerical results, which indicate our method outperforms existing denoising-time approaches.

Achieving precise, versatile whole-body character control in physics-based animation remains challenging. Recent diffusion-based policies generate rich and expressive motions but typically rely on gradient-based test-time guidance to satisfy task objectives, which is slow and can reduce robustness. We introduce NaP-Control (Navigating Diffusion Prior for Versatile and Fast Character Control), abbreviated as NaP. Our method uses reinforcement learning to manipulate the latent noise of a task-agnostic diffusion policy prior, steering it toward task-specific behaviors for fast, robust control with high motion fidelity. In contrast to methods that rely solely on offline training, NaP interacts with the environment during training to correct motions and optimize task rewards, improving success rates and enabling adaptation to challenging scenarios. By directly predicting task-optimized diffusion noise, NaP eliminates iterative guidance during denoising and enables efficient inference. Experiments show that NaP attains higher success rates and faster inference while preserving natural motion across diverse tasks.

Model-based reinforcement learning (MBRL) with autoregressive world models suffers from compounding errors, whereas diffusion world models mitigate this by generating trajectory segments jointly. However, existing diffusion guides are either policy-only, discarding value information, or reward-based, which becomes myopic when the diffusion horizon is short. We introduce Advantage-Guided Diffusion for MBRL (AGD-MBRL), which steers the reverse diffusion process using the agent's advantage estimates so that sampling concentrates on trajectories expected to yield higher long-term return beyond the generated window. We develop two guides: (i) Sigmoid Advantage Guidance (SAG) and (ii) Exponential Advantage Guidance (EAG). We prove that a diffusion model guided through SAG or EAG allows us to perform reweighted sampling of trajectories with weights increasing in state-action advantage-implying policy improvement under standard assumptions. Additionally, we show that the trajectories generated from AGD-MBRL follow an improved policy (that is, with higher value) compared to an unguided diffusion model. AGD integrates seamlessly with PolyGRAD-style architectures by guiding the state components while leaving action generation policy-conditioned, and requires no change to the diffusion training objective. On MuJoCo control tasks (HalfCheetah, Hopper, Walker2D and Reacher), AGD-MBRL improves sample efficiency and final return over PolyGRAD, an online Diffuser-style reward guide, and model-free baselines (PPO/TRPO), in some cases by a margin of 2x. These results show that advantage-aware guidance is a simple, effective remedy for short-horizon myopia in diffusion-model MBRL.

Reinforcement learning (RL) has been successfully applied to autoregressive (AR) and diffusion models. However, extending RL to hybrid AR-diffusion frameworks remains challenging due to interleaved inference and noisy log-probability estimation. In this work, we study masked autoregressive models (MAR) and show that the diffusion head plays a critical role in training dynamics, often introducing noisy gradients that lead to instability and early performance saturation. To address this issue, we propose a stabilized RL framework for MAR. We introduce multi-trajectory expectation (MTE), which estimates the optimization direction by averaging over multiple diffusion trajectories, thereby reducing diffusion-induced gradient noise. To avoid over-smoothing, we further estimate token-wise uncertainty from multiple trajectories and apply multi-trajectory optimization only to the top-k% uncertain tokens. In addition, we introduce a consistency-aware token selection strategy that filters out AR tokens that are less aligned with the final generated content. Extensive experiments across multiple benchmarks demonstrate that our method consistently improves visual quality, training stability, and spatial structure understanding over baseline GRPO and pre-RL models. Code is available at: https://github.com/AMAP-ML/mar-grpo.

Reinforcement-Learning-based post-training has recently emerged as a promising paradigm for aligning text-to-image diffusion models with human preferences. In recent studies, increasing the rollout group size yields pronounced performance improvements, indicating substantial room for further alignment gains. However, scaling rollouts on large-scale foundational diffusion models (e.g., FLUX.1-12B) imposes a heavy computational burden. To alleviate this bottleneck, we explore the integration of FP4 quantization into Diffusion RL rollouts. Yet, we identify that naive quantized pipelines inherently introduce risks of performance degradation. To overcome this dilemma between efficiency and training integrity, we propose Sol-RL (Speed-of-light RL), a novel FP4-empowered Two-stage Reinforcement Learning framework. First, we utilize high-throughput NVFP4 rollouts to generate a massive candidate pool and extract a highly contrastive subset. Second, we regenerate these selected samples in BF16 precision and optimize the policy exclusively on them. By decoupling candidate exploration from policy optimization, Sol-RL integrates the algorithmic mechanisms of rollout scaling with the system-level throughput gains of NVFP4. This synergistic algorithm-hardware design effectively accelerates the rollout phase while reserving high-fidelity samples for optimization. We empirically demonstrate that our framework maintains the training integrity of BF16 precision pipeline while fully exploiting the throughput gains enabled by FP4 arithmetic. Extensive experiments across SANA, FLUX.1, and SD3.5-L substantiate that our approach delivers superior alignment performance across multiple metrics while accelerating training convergence by up to $4.64\times$, unlocking the power of massive rollout scaling at a fraction of the cost.

Diffusion large language models (dLLMs) are emerging as a compelling alternative to dominant autoregressive models, replacing strictly sequential token generation with iterative denoising and parallel generation dynamics. However, their open-source ecosystem remains fragmented across model families and, in particular, across post-training pipelines, where reinforcement learning objectives, rollout implementations and evaluation scripts are often released as paper-specific codebases. This fragmentation slows research iteration, raises the engineering burden of reproduction, and makes fair comparison across algorithms difficult. We present DARE (dLLMs Alignment and Reinforcement Executor), an open framework for post-training and evaluating dLLMs. Built on top of verl~\cite{sheng2024hybridflow} and OpenCompass~\cite{2023opencompass}, DARE unifies supervised fine-tuning, parameter-efficient fine-tuning, preference optimization, and dLLM-specific reinforcement learning under a shared execution stack for both masked and block diffusion language models. Across representative model families including LLaDA, Dream, SDAR, and LLaDA2.x, DARE provides broad algorithmic coverage, reproducible benchmark evaluation, and practical acceleration. Extensive empirical results position that DARE serves as a reusable research substrate for developing, comparing, and deploying post-training methods for current and emerging dLLMs.

Flow-GRPO successfully applies reinforcement learning to flow models, but uses uniform credit assignment across all steps. This ignores the temporal structure of diffusion generation: early steps determine composition and content (low-frequency structure), while late steps resolve details and textures (high-frequency details). Moreover, assigning uniform credit based solely on the final image can inadvertently reward suboptimal intermediate steps, especially when errors are corrected later in the diffusion trajectory. We propose Stepwise-Flow-GRPO, which assigns credit based on each step's reward improvement. By leveraging Tweedie's formula to obtain intermediate reward estimates and introducing gain-based advantages, our method achieves superior sample efficiency and faster convergence. We also introduce a DDIM-inspired SDE that improves reward quality while preserving stochasticity for policy gradients.

Diffusion models achieve state-of-the-art generative performance but are fundamentally bottlenecked by their slow, iterative sampling process. While diffusion distillation techniques enable high-fidelity, few-step generation, traditional objectives often restrict the student's performance by anchoring it solely to the teacher. Recent approaches have attempted to break this ceiling by integrating Reinforcement Learning (RL), typically through a simple summation of distillation and RL objectives. In this work, we propose a novel paradigm by re-conceptualizing distribution matching as a reward, denoted as $R_\text{dm}$. This unified perspective bridges the algorithmic gap between Diffusion Matching Distillation (DMD) and RL, providing several primary benefits. (1) Enhanced Optimization Stability: We introduce Group Normalized Distribution Matching (GNDM), which adapts standard RL group normalization to stabilize $R_\text{dm}$ estimation. By leveraging group-mean statistics, GNDM establishes a more robust and effective optimization direction. (2) Seamless Reward Integration: Our reward-centric formulation inherently supports adaptive weighting mechanisms, allowing for the fluid combination of DMD with external reward models. (3) Improved Sampling Efficiency: By aligning with RL principles, the framework readily incorporates Importance Sampling (IS), leading to a significant boost in sampling efficiency. Extensive experiments demonstrate that GNDM outperforms vanilla DMD, reducing the FID by 1.87. Furthermore, our multi-reward variant, GNDMR, surpasses existing baselines by striking an optimal balance between aesthetic quality and fidelity, achieving a peak HPS of 30.37 and a low FID-SD of 12.21. Ultimately, $R_\text{dm}$ provides a flexible, stable, and efficient framework for real-time, high-fidelity synthesis. Codes are coming soon.

Thanks to their remarkable flexibility, diffusion models and flow models have emerged as promising candidates for policy representation. However, efficient reinforcement learning (RL) upon these policies remains a challenge due to the lack of explicit log-probabilities for vanilla policy gradient estimators. While numerous attempts have been proposed to address this, the field lacks a unified perspective to reconcile these seemingly disparate methods, thus hampering ongoing development. In this paper, we bridge this gap by introducing a comprehensive taxonomy for RL algorithms with diffusion/flow policies. To support reproducibility and agile prototyping, we introduce a modular, JAX-based open-source codebase that leverages JIT-compilation for high-throughput training. Finally, we provide systematic and standardized benchmarks across Gym-Locomotion, DeepMind Control Suite, and IsaacLab, offering a rigorous side-by-side comparison of diffusion-based methods and guidance for practitioners to choose proper algorithms based on the application. Our work establishes a clear foundation for understanding and algorithm design, a high-efficiency toolkit for future research in the field, and an algorithmic guideline for practitioners in generative models and robotics. Our code is available at https://github.com/typoverflow/flow-rl.

Surgical smoke severely degrades intraoperative video quality, obscuring anatomical structures and limiting surgical perception. Existing learning-based desmoking approaches rely on scarce paired supervision and deterministic restoration pipelines, making it difficult to perform exploration or reinforcement-driven refinement under real surgical conditions. We propose PhySe-RPO, a diffusion restoration framework optimized through Physics- and Semantics-Guided Relative Policy Optimization. The core idea is to transform deterministic restoration into a stochastic policy, enabling trajectory-level exploration and critic-free updates via group-relative optimization. A physics-guided reward imposes illumination and color consistency, while a visual-concept semantic reward learned from CLIP-based surgical concepts promotes smoke-free and anatomically coherent restoration. Together with a reference-free perceptual constraint, PhySe-RPO produces results that are physically consistent, semantically faithful, and clinically interpretable across synthetic and real robotic surgical datasets, providing a principled route to robust diffusion-based restoration under limited paired supervision.

Classical on-policy algorithms such as PPO and mirror descent policy optimization provide stable proximal policy updates through tractable action likelihoods, but are typically instantiated with simple Gaussian policies whose expressiveness can be limited in complex continuous-control tasks. Generative policies based on diffusion and flow models provide more expressive action distributions, but they naturally define distributions over multi-step denoising paths whose terminal action density is often intractable, creating a mismatch with likelihood-based on-policy proximal updates. To address this mismatch, we introduce GSB-MDPO (Generalized Schrödinger Bridge Mirror Descent Policy Optimization), which formulates on-policy generative policy optimization as a Generalized Schrödinger Bridge problem over state-conditioned generation paths and instantiates the resulting path-measure update through mirror descent policy optimization. The key insight is that the GSB path-space KL plays the role of the proximal term in MDPO while upper-bounding the terminal action KL, enabling direct control of the executed action distribution without explicit terminal action likelihood evaluation. Experiments on 14 continuous-control tasks across Playground and Gym-MuJoCo demonstrate the empirical effectiveness of GSB-MDPO and support path-space regularization as a principled proximal update for multi-step generative policies.

Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models with human preferences, inspiring the development of reward-centric diffusion reinforcement learning (RDRL) to achieve similar alignment and controllability. While diffusion models can generate high-quality outputs, RDRL remains susceptible to reward hacking, where the reward score increases without corresponding improvements in perceptual quality. We demonstrate that this vulnerability arises from the non-robustness of reward model gradients, particularly when the reward landscape with respect to the input image is sharp. To mitigate this issue, we introduce methods that exploit gradients from a robustified reward model without requiring its retraining. Specifically, we employ gradients from a flattened reward model, obtained through parameter perturbations of the diffusion model and perturbations of its generated samples. Empirically, each method independently alleviates reward hacking and improves robustness, while their joint use amplifies these benefits. Our resulting framework, RSA-FT (Reward Sharpness-Aware Fine-Tuning), is simple, broadly compatible, and consistently enhances the reliability of RDRL.

Aligning Diffusion models has achieved remarkable breakthroughs in generating high-quality, human preference-aligned images. Existing techniques, such as supervised fine-tuning (SFT) and DPO-style preference optimization, have become principled tools for fine-tuning diffusion models. However, SFT relies on high-quality images that are costly to obtain, while DPO-style methods depend on large-scale preference datasets, which are often inconsistent in quality. Beyond data dependency, these methods are further constrained by computational inefficiency. To address these two challenges, we propose Composite Reward Assisted Fine-Tuning (CRAFT), a lightweight yet powerful fine-tuning paradigm that requires significantly reduced training data while maintaining computational efficiency. It first leverages a Composite Reward Filtering (CRF) technique to construct a high-quality and consistent training dataset and then perform an enhanced variant of SFT. We also theoretically prove that CRAFT actually optimizes the lower bound of group-based reinforcement learning, establishing a principled connection between SFT with selected data and reinforcement learning. Our extensive empirical results demonstrate that CRAFT with only 100 samples can easily outperform recent SOTA preference optimization methods with thousands of preference-paired samples. Moreover, CRAFT can even achieve 11-220$\times$ faster convergences than the baseline preference optimization methods, highlighting its extremely high efficiency.

Diffusion Large Language Models (dLLMs) introduce a new paradigm for language generation, which in turn presents new challenges for aligning them with human preferences. In this work, we aim to improve the policy optimization for dLLMs by reducing the cost of the trajectory probability calculation, thereby enabling scaled-up offline policy training. We prove that: (i) under reference policy regularization, the probability ratio of the newly unmasked tokens is an unbiased estimate of that of intermediate diffusion states, and (ii) the probability of the full trajectory can be effectively estimated with a single forward pass of a re-masked final state. By integrating these two trajectory reduction strategies into a policy optimization objective, we propose Trajectory Reduction Policy Optimization (dTRPO). We evaluate dTRPO on 7B dLLMs across instruction-following and reasoning benchmarks. Results show that it substantially improves the core performance of state-of-the-art dLLMs, achieving gains of up to 9.6% on STEM tasks, up to 4.3% on coding tasks, and up to 3.0% on instruction-following tasks. Moreover, dTRPO exhibits strong training efficiency due to its offline, single-forward nature, and achieves improved generation efficiency through high-quality outputs.

Image-conditioned video diffusion models achieve impressive visual realism but often suffer from weakened motion fidelity, e.g., reduced motion dynamics or degraded long-term temporal coherence, especially after fine-tuning. We study motion alignment in video diffusion models post-training. To address this, we introduce pixel-motion rewards based on pixel flux dynamics, capturing both instantaneous and long-term motion consistency. We further propose Smooth Hybrid Fine-tuning (SHIFT), a scalable reward-driven framework that unifies supervised fine-tuning and advantage-weighted fine-tuning. Benefiting from novel adversarial advantages, SHIFT improves convergence speed and mitigates reward hacking. Experiments show that our approach efficiently resolves dynamic-degree collapse in modern video diffusion models supervised fine-tuning. Project page: https://xiye20.github.io/projects/SHIFT/.