Research map / Image generation / Fast diffusion sampling
DuoFlow: JVP-Free Finite-Difference Mean Flows for One-Step Image Generation
Zeming Li, Xiangyu Zhang, Ping Tan, Heung-Yeung Shum
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 |
|---|---|---|---|
| replace JVP-based MeanFlow with DuoFlow finite-difference training | MeanFlow-XL/2 · ImageNet 256x256 (1-NFE) | FID: 3.43 -> 2.48 | Improved |
| replace JVP with one-forward differential estimator | derivative evaluation · SiT-XL/2 micro-benchmark | speedup: 7.17x to 34.83x faster | Improved |
| DuoFlow trainer instead of MeanFlow JVP trainer | XL/2 training · end-to-end wall-clock | seconds per iteration: 0.89 -> 0.35 s/iter | Improved |
| add sign randomization to one-sided differencing (ablation) | DuoFlow · ImageNet 256x256 (80-epoch ablation) | FID: 19.06 -> 14.11 | Improved |
| add interval-conditioned joint velocity estimation (ablation) | DuoFlow · ImageNet 256x256 (80-epoch ablation) | FID: 14.11 -> 11.30 | Improved |
| add progressive self-bootstrapping (ablation) | DuoFlow · ImageNet 256x256 (80-epoch ablation) | FID: 11.30 -> 9.17 | Improved |
| signed differencing alone vs JVP MeanFlow baseline (ablation) | MeanFlow · ImageNet 256x256 (80-epoch ablation) | FID: 14.13 -> 14.11 | No clear effect |
| use fixed kappa=0.9 instead of progressive alignment-gated kappa (ablation) | DuoFlow self-bootstrapping · ImageNet 256x256 (80-epoch ablation) | best FID: 9.17 -> 35.33 | Hurt |
| use progressive kappa instead of fixed 0.5-0.7 (ablation) | DuoFlow self-bootstrapping · ImageNet 256x256 (80-epoch ablation) | best FID: 9.36 -> 9.17 | Improved |
| use extreme step size (0.25x or 50x base) (ablation) | DuoFlow differencing · ImageNet 256x256 (80-epoch ablation) | FID: 9.17 -> 16.52 (0.25x), 14.21 (50x) | Hurt |
| vary base step size from 1x to 5x or 10x (ablation) | DuoFlow differencing · ImageNet 256x256 (80-epoch ablation) | FID: 9.17 -> 9.12 (5x), 9.16 (10x) | No clear effect |
| add RMS-based step rescaling (ablation) | DuoFlow differencing · ImageNet 256x256 (80-epoch ablation) | FID: 9.68 -> 9.17 | Improved |
Closest papers by meaning
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.90)
- Trajectory Consistency for One-Step Generation on Euler Mean Flows (ICML 2026 · similarity 0.89)
- AlphaFlow: Understanding and Improving MeanFlow Models (ICLR 2026 · similarity 0.89)
- Understanding, Accelerating, and Improving MeanFlow Training (CVPR 2026 · similarity 0.89)
- Improved Mean Flows: On the Challenges of Fastforward Generative Models (CVPR 2026 · similarity 0.89)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · 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.
