Research map / Image generation / Fast diffusion sampling
MixFlow Training: Alleviating Exposure Bias with Slowed Interpolation Mixture
Hui Li, Jiayue Lyu, Fu-Yun Wang, Kaihui Cheng, Siyu Zhu, Jingdong Wang
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 |
|---|---|---|---|
| post-train with MixFlow slowed interpolation mixture | SiT-B/2 · ImageNet 256x256 (w/o guidance) | gFID: 17.97 -> 15.64 | Improved |
| post-train with MixFlow | SiT-XL/2 · ImageNet 256x256 (w/ guidance) | gFID: 2.15 -> 1.99 | Improved |
| add MixFlow post-training | RAE-XL · ImageNet 256x256 | gFID: 1.51 -> 1.43 (w/o guidance); 1.13 -> 1.10 (w/ guidance) | Improved |
| add MixFlow post-training | REPA-XL/2 · ImageNet 256x256 | gFID: 6.90 -> 6.28 (w/o guidance); 1.65 -> 1.59 (w/ guidance) | Improved |
| sample slowed timestep mt from U[0,t] instead of U[0,1] (ablation) | MixFlow SiT-B · ImageNet 256x256 (w/o guidance) | gFID: 18.27 -> 16.57 | Improved |
| sample t from Beta(2,1) instead of U[0,1] (ablation) | MixFlow SiT-B · ImageNet 256x256 | gFID: 16.57 -> 15.64 (w/o guidance); 4.25 -> 3.93 (w/ guidance) | Improved |
| standard post-training for same extra steps instead of MixFlow (ablation) | SiT-B · ImageNet 256x256 | gFID: 17.97 -> 17.96 (w/o guidance); 4.46 -> 4.46 (w/ guidance) | No clear effect |
| vary mixture range gamma from 0.7 to 1.0 (ablation) | MixFlow SiT-B · ImageNet 256x256 | gFID: 15.65, 15.64, 15.64, 15.64 (w/o guidance) | No clear effect |
| use gamma=0 (no mixture) (ablation) | MixFlow SiT-B · ImageNet 256x256 | gFID: 17.97 -> 18.00 (w/o guidance) | Hurt |
| MixFlow vs Time-Shift sampler | SiT-B/2 · ImageNet 256x256 (w/o guidance) | gFID: 17.97 -> 18.12 (Time-Shift) | Hurt |
| fewer sampling steps (250 to 20) (ablation) | MixFlow SiT-B/2 · ImageNet 256x256 (w/o guidance) | gFID gain: gain 2.33 -> 2.58 | Improved |
| add MixFlow post-training | SD 3.5-medium fine-tuned · GenEval/DPG-Bench/T2I-CompBench | Avg scores: DPG-Bench 84.87 -> 87.49 (40 steps) | Improved |
Closest papers by meaning
- Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps (ICLR 2024 · similarity 0.89)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.88)
- Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment (NeurIPS 2024 · similarity 0.87)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.87)
- Cross Flow: One-Step Generation Across Latent and Pixel Spaces (NeurIPS 2026 · similarity 0.87)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.87)
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.
