Research map / Image generation / Fast diffusion sampling
RAPID$^3$: Tri-Level Reinforced Acceleration Policies for Diffusion Transformer
Wangbo Zhao, Yizeng Han, Zhiwei Tang, Jiasheng Tang, Pengfei Zhou, Kai Wang, Bohan Zhuang, Zhangyang Wang, Fan Wang, Yang You
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 |
|---|---|---|---|
| use learned per-image tri-level policies instead of static or hand-designed acceleration | SD3 · COCO, HPS, GenEval | speedup / quality: 2.92x speed; GenEval overall 63.48 vs 61.67 (SD3 9-steps) | Improved |
| use tri-level policies instead of step-only dynamic policy (TPDM) | SD3 · GenEval Overall | Overall: 60.70 -> 63.48 | Improved |
| use RAPID^3 instead of fine-tuned DyFLUX | FLUX · FLUX latency and training cost | latency (s): 13.93 -> 8.30; GPU hours 38,000 -> 400 | Improved |
| add cache-reuse policy to step-skip (ablation) | RAPID^3 · HPS | HPS score: 27.75 -> 27.86 | Improved |
| add sparse-attention policy to step+cache (ablation) | RAPID^3 · HPS | HPS score: 27.86 -> 28.07 | Improved |
| add step-skip policy to cache+sparse (ablation) | RAPID^3 · COCO | Aesthetic score: 5.13 -> 5.26 | Improved |
| add discriminator reward to ImageReward reward (ablation) | RAPID^3 · HPS | HPS score: 27.72 -> 28.07 | Improved |
| add discriminator reward to ImageReward reward (ablation) | RAPID^3 · ImageReward | ImageReward: 0.9605 -> 0.9574 | Hurt |
| use discriminator reward alone without ImageReward (ablation) | RAPID^3 · HPS | HPS score: 28.07 -> 27.96 | Hurt |
| replace manual combination of step, cache and sparse acceleration with learned policies (ablation) | SD3 · COCO CLIP | CLIP score: 31.43 (manual-1) -> 32.09 | Improved |
| replace GRPO with RLOO (ablation) | RAPID^3 · COCO, HPS | HPS score: 28.07 -> 28.06 | No clear effect |
| increase text-only training prompts from 20K to 40K (ablation) | RAPID^3 · HPS | HPS score: 28.07 -> 28.26 | Improved |
Closest papers by meaning
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.91)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.91)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.91)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Dynamic Diffusion Transformer (ICLR 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.
