Research map / Image generation / Fast diffusion sampling
Towards Stabilized and Efficient Diffusion Transformers through Long-Skip-Connections with Spectral Constraints
Guanjie Chen, Xinyu Zhao, Yucheng Zhou, Xiaoye Qu, Tianlong Chen, Yu Cheng
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 long skip connections to DiT-XL and train | DiT-XL · ImageNet 256 class-to-image (cfg=1.0) | FID: 9.49 -> 8.37 | Improved |
| add long skip connections to Latte | Latte · UCF101 class-to-video | FVD: 155.22 -> 141.30 | Improved |
| add long skip connections to Latte (text-to-video continual training) | Latte · VBench text-to-video (Latte, 50-step DDIM) | VBench score (%): 76.14 -> 75.60 | No clear effect |
| cache at timesteps 700->50 instead of 950->300 (ablation) | Skip-DiT · VBench text-to-video, caching timestep ablation (Skip-DiT) | PSNR: 20.58 -> 29.52 | Improved |
| cache at timesteps 700->50 instead of 950->300 (ablation) | Skip-DiT · VBench text-to-video, caching timestep ablation (Skip-DiT) | VBench score (%): 75.48 -> 75.51 | No clear effect |
| cache at 700->50 instead of 900->50 (ablation) | Skip-DiT · VBench text-to-video, caching timestep ablation (Skip-DiT) | PSNR: 22.13 -> 29.52 | Improved |
| cache block 27 linked via LSC 0 rather than LSC 2 (ablation) | Skip-DiT · Connection selection for caching (Latte T2V) | Similarity (%): 99.66 -> 99.88 | Improved |
| integrate PAB caching with Skip-DiT instead of Latte (ablation) | PAB235 · Compatibility with caching methods | VBench drop: 2.37 -> 1.81 | Improved |
| integrate PAB caching with Skip-DiT instead of Latte (ablation) | PAB469 · Compatibility with caching methods | VBench drop: 5.94 -> 3.96 | Improved |
| use Skip-DiT instead of Latte at 10 sampling steps | Latte · Few-step text-to-video (10 steps) | Delta VBench (%): 1.73 -> 0.94 | Improved |
| use Skip-DiT instead of Latte at 5 sampling steps | Latte · Few-step text-to-video (5 steps) | SSIM: 0.53 -> 0.50 | Hurt |
| increase cache interval n from 2 to 6 (ablation) | Skip-DiT · UCF101 class-to-video (Latte) | FVD: 141.42 -> 151.77 | Hurt |
Closest papers by meaning
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.91)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.90)
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.89)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.89)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 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.
