Research map / Reinforcement learning / RL from feedback
Parallel Diffusion Solver via Residual Dirichlet Policy Optimization
Ruoyu Wang, 李資玉, Beier Zhu, Liangyu Yuan, H L Zhang, Xiaomin Yang, Xiaojun Chang, Chi Zhang
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 |
|---|---|---|---|
| add learned scaling factor o_n (ablation) | EPD-Solver · CIFAR10, NFE 3 and 9 | FID: 13.25 -> 10.40 (NFE 3); 2.79 -> 2.49 (NFE 9) | Improved |
| add learned scaling factors delta_n^k (ablation) | EPD-Solver · CIFAR10, NFE 3 and 9 | FID: 13.02 -> 10.40 (NFE 3); 2.69 -> 2.49 (NFE 9) | Improved |
| add both scaling factors (ablation) | EPD-Solver · CIFAR10, NFE 3 and 9 | FID: 16.01 -> 10.40 (NFE 3); 3.24 -> 2.49 (NFE 9) | Improved |
| use Uniform time schedule instead of LogSNR (ablation) | EPD-Solver · CIFAR10, NFE 3 | FID: 54.07 -> 10.40 | Improved |
| use Uniform time schedule instead of Polynomial (ablation) | EPD-Solver · CIFAR10, NFE 3 | FID: 11.10 -> 10.40 | Improved |
| use DPM-Solver-2 as teacher instead of Heun (ablation) | EPD-Solver · CIFAR10, NFE 3 | FID: 15.91 -> 10.40 | Improved |
| use DPM-Solver-2 as teacher instead of iPNDM (ablation) | EPD-Solver · CIFAR10, NFE 3 | FID: 13.69 -> 10.40 | Improved |
| increase parallel directions K from 1 to 2 (ablation) | EPD-Solver · FFHQ, NFE 3 | FID: 26.0 -> 22.7 | Improved |
| increase parallel directions K from 2 to 3 (ablation) | EPD-Solver · ImageNet, NFE 9 | FID: 4.20 -> 4.18 | No clear effect |
| add parallel directions K=2 vs K=1 for latency (ablation) | EPD-Solver · SD1.5, 20 steps | latency (s): 0.6048 -> 0.6252 | No clear effect |
| add RL fine-tuning (Dirichlet policy) after distillation (ablation) | EPD-Solver · SD v1.5 | ImageReward: 3.0405 -> 3.1121 | Improved |
| use EPD-Solver with 10 steps instead of iPNDM 50 steps | iPNDM · SD v1.5, HPSv2.1 | HPSv2.1: 0.2474 -> 0.2482 | Improved |
Closest papers by meaning
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.90)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.90)
- TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics (NeurIPS 2025 · similarity 0.90)
- Learning to Solve Generative ODEs Beyond the Linear Span (NeurIPS 2026 · similarity 0.89)
- S4S: Solving for a Fast Diffusion Model Solver (ICML 2025 · similarity 0.89)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.89)
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.
