Research map / Image generation / Fast diffusion sampling
Temporal Equilibrium MeanFlow: Bridging the Scale Gap for One-Step Generation
Yuanpeng Tu, Yunpeng Chen, Xinyu Zhang, Chao Liao, Hengshuang Zhao
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 temporal equilibrium weighting (both boundary and JVP terms) (ablation) | TEMF · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 6.52 -> 4.31 | Improved |
| use JVP term only or boundary term only in weighting (ablation) | TEMF weighting · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 6.52 -> 5.13 (JVP only); 6.52 -> 4.86 (boundary only) | Improved |
| replace static 25% r != t ratio with exponential dynamic boundary schedule (ablation) | MeanFlow sampling · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 5.72 -> 4.31 | Improved |
| use linear or cosine boundary schedule instead of exponential (ablation) | TEMF · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 4.31 -> 4.86 (linear); 4.31 -> 4.53 (cosine) | Hurt |
| add dynamic velocity mixing for CFG instead of no mixing (ablation) | TEMF · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 5.24 -> 4.31 | Improved |
| use dynamic mixing instead of fixed mixing m=0.25 (ablation) | TEMF · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 4.58 -> 4.31 | Improved |
| incrementally add temporal weighting, dynamic boundary, CFG integration to MeanFlow (ablation) | MeanFlow baseline · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 6.17 -> 5.26 -> 4.93 -> 4.61 -> 4.31 | Improved |
| set r != t sampling ratio to 0.50 rather than extremes (ablation) | TEMF · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 5.95 (0.0) / 4.68 (1.0) -> 4.31 (0.50) | Improved |
| remove boundary term in weighting (lambda1 = 0.0) (ablation) | TEMF weighting · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 4.31 -> 4.75 | Hurt |
| vary lambda2 away from 1.0 (0.5 to 2.0) (ablation) | TEMF weighting · ImageNet 256x256, DiT-B/4, 1-NFE | FID-50K: 4.31 -> 4.58 (0.5); 4.62 (2.0) | Hurt |
| use velocity mix ratio 0.3 instead of 0.1 or 0.7 (ablation) | TEMF CFG training · ImageNet 256x256, DiT-B/4, 1-NFE, w=3.0 | FID-50K: 4.49 (0.1), 4.67 (0.7) -> 4.31 (0.3) | Improved |
| replace MeanFlow with TEMF at the same model size | MeanFlow-XL · ImageNet 256x256, 1-NFE | FID-50K: 3.43 -> 2.81 | 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.
- SoFlow-XL — SoFlow: Solution Flow Models for One-Step Generative Modeling (ICLR 2026) · better on 2 of 2 shared results
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
- Improved Mean Flows: On the Challenges of Fastforward Generative Models (CVPR 2026 · similarity 0.91)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.89)
- AlphaFlow: Understanding and Improving MeanFlow Models (ICLR 2026 · similarity 0.89)
- Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models (NeurIPS 2026 · similarity 0.88)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.87)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.86)
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.
