Research map / Image generation / Fast diffusion sampling
RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers
Ruitong Sun, Tianze Yang, Wei Niu, Jin 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 |
|---|---|---|---|
| use adaptive region-wise rank instead of uniform r=1024 (ablation) | DiT-XL/2 RSTR · ImageNet 512x512 | FID and MACs: FID 2.92 vs 3.01; MACs 34.68T vs 22.37T | Mixed |
| use adaptive rank instead of uniform r=512 (ablation) | DiT-XL/2 RSTR · ImageNet 512x512 | FID: 4.46 -> 3.01 | Improved |
| use adaptive rank instead of uniform r=256 (ablation) | DiT-XL/2 RSTR · ImageNet 512x512 | FID: 3.68 -> 3.01 | Improved |
| add adaptive caching to the sparse schedule (ablation) | DiT-XL/2 RSTR · ImageNet 512x512 (50 steps) | MACs: MACs 30.9 -> 24.9; FID 2.71 -> 2.72 | Improved |
| lengthen reference trajectory from 50 to 1000 steps (ablation) | DiT-XL/2 RSTR · ImageNet 512x512 | FID: 2.84 -> 2.71 | Improved |
| shift high-scale CFG step from 12 to 16 (ablation) | DiT-XL/2 RSTR · ImageNet 256x256 | FID: FID 2.10 -> 2.21; IS 263.0 -> 272.4 | Hurt |
| shift high-scale CFG step from 12 to 8 (ablation) | DiT-XL/2 RSTR · ImageNet 256x256 | FID: FID 2.10 -> 2.04; IS 263.0 -> 256.5 | Improved |
| keep high-scale step 39 rather than removing it (ablation) | DiT-XL/2 RSTR · ImageNet 256x256 | FID: 2.44 vs 2.10 | Improved |
| keep high-scale step 47 rather than removing it (ablation) | DiT-XL/2 RSTR · ImageNet 256x256 | FID: 2.16 vs 2.10 | Improved |
| use sparse schedule instead of constant CFG under 3.3x scale (ablation) | DiT-XL/2 RSTR · ImageNet 256x256 | FID: 16.39 -> 9.20 | Improved |
| transfer DiT-XL/2 schedule to DiT-B/2 (ablation) | RSTR schedule · ImageNet 256x256 DiT-B/2 | FID: 11.52 -> 9.22 | Improved |
| transfer DiT-XL/2 schedule to DiT-S/2 (ablation) | RSTR schedule · ImageNet 256x256 DiT-S/2 | FID: 24.09 -> 20.00 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HarmoniCa — HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025) · better on 1 of 1 shared result
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
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.89)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.88)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.88)
- TinyFusion: Diffusion Transformers Learned Shallow (CVPR 2025 · similarity 0.88)
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.
