Research map / Image generation / Fast diffusion sampling
Masked Auto-Regressive Variational Acceleration: Fast Inference Makes Practical Reinforcement Learning
Yuxuan Gu, Weimin Bai, Yifei Wang, Weijian Luo, He Sun
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 |
|---|---|---|---|
| distill the diffusion chain into one step (GSIM), Base model | MAR-B · ImageNet 256x256 | inference time per image (s): 20.10 -> 0.61 (about 32.95x) | Improved |
| distill the diffusion chain into one step, Base model | MAR-B · ImageNet 256x256 | FID: 2.60 -> 3.06 | Hurt |
| distill into one step, Huge model | MAR-H · ImageNet 256x256 | FID: 2.06 -> 2.00 | Improved |
| set distillation CFG scale to 1.2 instead of 1 (no CFG) (ablation) | MARVAL-B · ImageNet 256x256 | FID: 3.75 -> 3.06 | Improved |
| raise distillation CFG scale from 1.2 to 4 (ablation) | MARVAL-B · ImageNet 256x256 | FID: 3.06 -> 13.84 | Hurt |
| raise distillation CFG scale from 1.2 to 4 (ablation) | MARVAL-B · ImageNet 256x256 | Inception Score: 220.2 -> 316.8 | Improved |
| add RL fine-tuning with PickScore reward (ablation) | MARVAL (Base) · ImageNet class names | ImageReward: -0.142 -> -0.110 | Improved |
| add RL fine-tuning with PickScore reward (ablation) | MARVAL (Huge) · ImageNet class names | ImageReward: -0.107 -> -0.064 | Improved |
| add RL fine-tuning with PickScore reward (ablation) | MARVAL (Base) · ImageNet class names | CLIP score: 29.40 -> 29.47 | No clear effect |
| compress 20-step DDIM to 1 step and add RL | DC-AR-512 · JourneyDB T2I | ImageReward: 0.5876 -> 0.6903 | Improved |
| couple RL directly with one-step GSIM samples | MARVAL · design argument | stability | Hurt |
Closest papers by meaning
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.89)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.88)
- MAR-3D: Progressive Masked Auto-regressor for High-Resolution 3D Generation (CVPR 2025 · similarity 0.88)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 2025 · similarity 0.87)
- Search-Augmented Masked Diffusion Models for Constrained Generation (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.
