Research map / Image generation / Fast diffusion sampling
Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers
Wenhao Sun, Ji Li, Zhaoqiang Liu
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 JiT spatial acceleration (18 NFE) instead of the 50-NFE baseline | FLUX.1-dev · Text-to-image (~4x tier) | ImageReward: 1.004 -> 1.017 | Improved |
| use JiT instead of 7-NFE FLUX.1-dev at equal speed | FLUX.1-dev · Text-to-image (~7x tier) | ImageReward: 0.7474 -> 0.9746 | Improved |
| use JiT instead of RALU at ~7x | FLUX.1-dev · Text-to-image (~7x tier) | HPSv2.1: 29.71 -> 29.02 | Hurt |
| use JiT at 18 NFE instead of the 50-NFE baseline | FLUX.1-dev · Text-to-image (~4x tier) | HPSv2.1: 30.39 -> 29.77 | Hurt |
| use JiT instead of 7-NFE FLUX.1-dev in blind human comparisons | FLUX.1-dev · User study | preference for JiT (%): 93.1% preference | Improved |
| add spatial approximation (interpolated velocity) for inactive tokens in SAG-ODE (ablation) | JiT · Ablation | HPSv2.1: 24.18 -> 26.90 | Improved |
| add spatial approximation for inactive tokens in SAG-ODE (ablation) | JiT · Ablation | T2I-CompBench: 0.3414 -> 0.3727 | Improved |
| use importance-guided token activation instead of static fixed grid (ablation) | JiT · Ablation | HPSv2.1: 26.51 -> 26.90 | Improved |
| use noise-aware DMF target instead of naive interpolation (ablation) | JiT · Ablation | HPSv2.1: 26.04 -> 26.90 | Improved |
| use noise-aware DMF target instead of naive interpolation (ablation) | JiT · Ablation | T2I-CompBench: 0.3602 -> 0.3727 | Improved |
Closest papers by meaning
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.90)
- PDE-SSM: A Spectral State Space Approach to Spatial Mixing in Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.90)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · 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.
