Research map / Reinforcement learning / Visual and representation RL
Diffusion Reward: Learning Rewards via Conditional Video Diffusion
Tao Huang, Guangqi Jiang, Yanjie Ze, Huazhe Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace VideoGPT with VQ-Diffusion as video model | video prediction for rewards · Real dexterous robot videos | PSNR: 23.35 -> 27.46 | Improved |
| use conditional entropy reward from video diffusion instead of VIPER log-likelihood | Diffusion Reward · MetaWorld (7 tasks) and Adroit (3 tasks) | success rate: +38% (MetaWorld), +35% (Adroit) over best baseline | Improved |
| standardize VIPER log-likelihood reward (VIPER-std) | VIPER · MetaWorld and Adroit | success rate | Improved |
| use conditional entropy rather than log-likelihood as reward, with diffusion or transformer video models (ablation) | Diffusion Reward · MetaWorld (7 tasks) | success rate | Improved |
| set diffusion sampling noise to 0 (deterministic reward) (ablation) | Diffusion Reward · MetaWorld and Adroit | success rate | Hurt |
| vary denoising steps from 2 to 20 (ablation) | Diffusion Reward · Visual RL tasks | success rate | Mixed |
| vary reward coefficient alpha from 1 to 0 (0.95 best) (ablation) | Diffusion Reward · Visual RL tasks | success rate | Mixed |
| increase context length from 1-2 to 4 or 8 historical frames (ablation) | Diffusion Reward · Visual RL tasks | success rate | Hurt |
| remove sparse environmental reward from composite reward (ablation) | Diffusion Reward · Door and Hammer (Adroit) | success rate | Hurt |
| increase number of denoised samples M from 1 to 4 (ablation) | Diffusion Reward · Door (Adroit) | FPS: 86.4 -> 45.9 | Hurt |
| use two independent coefficients instead of a single alpha for entropy and RND rewards (ablation) | Diffusion Reward · Door and Hammer (Adroit) | success rate | No clear effect |
| apply pretrained reward directly to unseen MetaWorld tasks | Diffusion Reward · 5 unseen MetaWorld tasks | success rate | Improved |
Later papers that compared with this method
1 later accepted paper reports Diffusion Reward as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ReLAM: Learning Anticipation Model for Rewarding Visual Robotic Manipulation (ICML 2026 · better on 4 of 5 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Subtask-Aware Visual Reward Learning from Segmented Demonstrations (ICLR 2025 · similarity 0.89)
- Diffusion-Reward Adversarial Imitation Learning (NeurIPS 2024 · similarity 0.88)
- Diffusion-based Curriculum Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 2025 · similarity 0.87)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.87)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
