Flow-GRPO is the first online RL method for flow matching models, raising GenEval accuracy from 63% to 95% and text-rendering accuracy from 59% to 92% with little reward hacking.
hub Canonical reference
Improving Video Generation with Human Feedback
Canonical reference. 73% of citing Pith papers cite this work as background.
53
Pith papers citing it
Background
73% of classified citations
abstract
Video generation has achieved significant advances through rectified flow techniques, but issues like unsmooth motion and misalignment between videos and prompts persist. In this work, we develop a systematic pipeline that harnesses human feedback to mitigate these problems and refine the video generation model. Specifically, we begin by constructing a large-scale human preference dataset focused on modern video generation models, incorporating pairwise annotations across multi-dimensions. We then introduce VideoReward, a multi-dimensional video reward model, and examine how annotations and various design choices impact its rewarding efficacy. From a unified reinforcement learning perspective aimed at maximizing reward with KL regularization, we introduce three alignment algorithms for flow-based models. These include two training-time strategies: direct preference optimization for flow (Flow-DPO) and reward weighted regression for flow (Flow-RWR), and an inference-time technique, Flow-NRG, which applies reward guidance directly to noisy videos. Experimental results indicate that VideoReward significantly outperforms existing reward models, and Flow-DPO demonstrates superior performance compared to both Flow-RWR and supervised fine-tuning methods. Additionally, Flow-NRG lets users assign custom weights to multiple objectives during inference, meeting personalized video quality needs.
hub tools
citation-role summary
background 16
baseline 4
method 2
citation-polarity summary
representative citing papers
PhyEditBench is a new benchmark for physics-aware image editing with real and synthetic instances plus a training-free PhyWorld baseline that uses test-time scaling to outperform SOTA models.
PRISM shows video diffusion models inherently encode preference information in noisy latents, achieving SOTA accuracy and enabling noise-robust early-stage sampling with a correlation to generative performance.
DRL trains a discriminator on data versus base-model samples in pretrained representation space and uses its logit as reward in KL-regularized RL, cutting guidance-free FID from 9.38 to 2.62 on SiT and similar gains on other backbones.
Presents multi-verifier framework and Adaptive Reward Weighting (ARW) for inference-time scaling in joint audio-video generation, reporting gains in alignment and synchronization on VGGSound and JavisBench-mini.
DRM turns a pre-trained diffusion model into a step-wise reward model and uses it for dense RL training (Step-wise GRPO) and guided sampling to improve final image quality.
Geo-Align applies RL with a perceptual reward derived from 3D camera trajectory estimation to improve controllability and fidelity in video generation without paired training data.
KVPO aligns streaming autoregressive video generators with human preferences via ODE-native GRPO, using KV cache for semantic exploration and TVE for velocity-based policy modeling, yielding gains in quality and alignment.
CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight bimanual manipulation tasks.
CaC presents a new spatiotemporal concentrating reward model for video anomalies, built on a novel large-scale dataset and three-stage training with RL and IoU rewards, claiming 25.7% accuracy gains and 11.7% anomaly reduction.
PhyGround is a new benchmark with curated prompts, a 13-law taxonomy, large-scale human annotations, and an open physics-specialized VLM judge for evaluating physical reasoning in generative video models.
RewardHarness self-evolves a tool-and-skill library from 100 preference examples to reach 47.4% accuracy on image-edit evaluation, beating GPT-5, and yields stronger RL-tuned models.
Stream-R1 improves distillation of autoregressive streaming video diffusion models by adaptively weighting supervision with a reward model at both rollout and per-pixel levels.
Hallo-Live achieves 20.38 FPS real-time text-to-audio-video avatar generation with 0.94s latency using asynchronous dual-stream diffusion and HP-DMD preference distillation, matching teacher model quality at 16x higher throughput.
OTCA improves GRPO training for visual generation by estimating step importance in trajectories and adaptively weighting multiple reward objectives.
SOLACE improves text-to-image generation by using intrinsic self-confidence rewards from noise reconstruction accuracy during reinforcement learning post-training without external supervision.
DiNa-LRM introduces a diffusion-native latent reward model using a noise-calibrated Thurstone likelihood on noisy states, matching VLM performance at lower compute in image alignment and preference optimization.
MixGRPO speeds up GRPO for flow-based image generators by restricting SDE sampling and optimization to a sliding window while using ODE elsewhere, cutting training time by up to 71% with better alignment performance.
UnifiedReward is the first unified reward model that jointly assesses multimodal understanding and generation to provide better preference signals for aligning vision models via DPO.
FlowPilot combines anchored flow matching for multimodal action pre-training with human-in-the-loop preference learning to improve long-horizon monocular sidewalk navigation, reporting 42% success in simulation and reduced interruptions in real-world tests.
SpecLoR rectifies the amplitude spectrum of lookahead-estimated clean latents to natural-video priors during early ODE sampling steps, cutting physical artifacts with only four extra NFEs.
SCAIL-2 achieves end-to-end character animation via direct video concatenation, in-context mask conditioning, mode-specific RoPE, the synthetic MotionPair-60K dataset, and Bias-Aware DPO, outperforming prior methods on multiple tasks.
Z-Reward trains a 27B reasoning teacher VLM on score distributions via GDSO and distills it via RISD into a 9B student, reaching 89.6% and 88.6% human preference accuracy with 41.3% optimization gain over SFT baseline.
FlowPRO applies proximalized preference optimization to flow-matching VLAs with intervention-rollback data to reach higher success rates on long-horizon bimanual tasks without rewards or critics.
citing papers explorer
-
Flow-GRPO: Training Flow Matching Models via Online RL
Flow-GRPO is the first online RL method for flow matching models, raising GenEval accuracy from 63% to 95% and text-rendering accuracy from 59% to 92% with little reward hacking.
-
CaC: Advancing Video Reward Models via Hierarchical Spatiotemporal Concentrating
CaC presents a new spatiotemporal concentrating reward model for video anomalies, built on a novel large-scale dataset and three-stage training with RL and IoU rewards, claiming 25.7% accuracy gains and 11.7% anomaly reduction.
-
PhyGround: Benchmarking Physical Reasoning in Generative World Models
PhyGround is a new benchmark with curated prompts, a 13-law taxonomy, large-scale human annotations, and an open physics-specialized VLM judge for evaluating physical reasoning in generative video models.
-
Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation
Stream-R1 improves distillation of autoregressive streaming video diffusion models by adaptively weighting supervision with a reward model at both rollout and per-pixel levels.
-
Hallo-Live: Real-Time Streaming Joint Audio-Video Avatar Generation with Asynchronous Dual-Stream and Human-Centric Preference Distillation
Hallo-Live achieves 20.38 FPS real-time text-to-audio-video avatar generation with 0.94s latency using asynchronous dual-stream diffusion and HP-DMD preference distillation, matching teacher model quality at 16x higher throughput.
-
Learning to Credit the Right Steps: Objective-aware Process Optimization for Visual Generation
OTCA improves GRPO training for visual generation by estimating step importance in trajectories and adaptively weighting multiple reward objectives.
-
SyncDPO: Enhancing Temporal Synchronization in Video-Audio Joint Generation via Preference Learning
SyncDPO improves temporal synchronization in video-audio joint generation using DPO with efficient on-the-fly negative sample construction and curriculum learning.
-
Leveraging Verifier-Based Reinforcement Learning in Image Editing
Edit-R1 builds a CoT-based reasoning reward model (RRM) via SFT and GCPO, then applies it with GRPO to improve image editing models such as FLUX.1-kontext.
-
VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects
VEFX-Bench releases a large human-labeled video editing dataset, a multi-dimensional reward model, and a standardized benchmark that better matches human judgments than generic evaluators.
-
VERTIGO: Visual Preference Optimization for Cinematic Camera Trajectory Generation
VERTIGO post-trains camera trajectory generators with visual preference signals from Unity-rendered previews scored by a cinematically fine-tuned VLM, cutting character off-screen rates from 38% to near zero while improving framing and prompt adherence.
-
SkyReels-V2: Infinite-length Film Generative Model
SkyReels-V2 produces infinite-length film videos via MLLM-based captioning, progressive pretraining, motion RL, and diffusion forcing with non-decreasing noise schedules.
-
A Systematic Post-Train Framework for Video Generation
A post-training pipeline for video generation models combines SFT, RLHF with novel GRPO, prompt enhancement, and inference optimization to improve visual quality, temporal coherence, and instruction following.
-
Reward-Aware Trajectory Shaping for Few-step Visual Generation
RATS lets few-step visual generators surpass multi-step teachers by shaping trajectories with reward-based adaptive guidance instead of strict imitation.
-
World Simulation with Video Foundation Models for Physical AI
Cosmos-Predict2.5 unifies text-to-world, image-to-world, and video-to-world generation in one model trained on 200M clips with RL post-training, delivering improved quality and control for physical AI.
-
Seedance 1.0: Exploring the Boundaries of Video Generation Models
Seedance 1.0 generates 5-second 1080p videos in about 41 seconds with claimed superior motion quality, prompt adherence, and multi-shot consistency compared to prior models.