Research map / Image generation / Fast diffusion sampling
Variational Flow Maps: Make Some Noise for One-Step Conditional Generation
Abbas Mammadov, So Takao, Bohan Chen, Ricardo Baptista, Morteza Mardani, Yee Whye Teh, Julius Berner
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 |
|---|---|---|---|
| jointly train flow map with the noise adapter instead of freezing theta (ablation) | VFM · ImageNet 256 box inpainting | FID: 136.12 -> 33.34 | Improved |
| jointly train flow map with the noise adapter instead of freezing theta (ablation) | VFM · ImageNet 256 Gaussian deblurring | FID: 172.74 -> 51.05 | Improved |
| add the mean flow loss to constrain theta (vs unconstrained-theta) (ablation) | VFM · 2D checkerboard posterior | off-manifold samples | Improved |
| add the KL term on the adapter (ablation) | VFM · 2D checkerboard | training stability | Improved |
| use an EMA of theta in the observation loss (ablation) | VFM · 2D checkerboard | support accuracy | Improved |
| increase tau to 1 or above (from tau at or below sigma) (ablation) | VFM · 2D checkerboard | CRPS and posterior MMD | Improved |
| set alpha to 1 versus below 1 (ablation) | VFM · 2D checkerboard | posterior vs prior MMD | Mixed |
| use alpha < 1 (0.5) for unconditional generation (ablation) | VFM · ImageNet 256 unconditional | FID: FID below 10 with 2-step | Improved |
| average 10 VFM posterior samples instead of a single sample | VFM · ImageNet 256 Gaussian deblurring | PSNR: 21.74 -> 23.92 | Improved |
| average 10 VFM posterior samples instead of a single sample | VFM · ImageNet 256 box inpainting | PSNR: 21.98 -> 22.71 | Improved |
| use VFM instead of guidance baselines | Latent DPS · ImageNet 256 box inpainting | FID: 62.89 -> 33.34 | Improved |
| use VFM single sample instead of Latent DAPS | Latent DAPS · ImageNet 256 box inpainting | PSNR: 23.98 -> 21.98 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- FlowDPS — FlowDPS : Flow-Driven Posterior Sampling for Inverse Problems (ICCV 2025) · better on 2 of 4 shared results
- FlowChef — FlowChef: Steering of Rectified Flow Models for Controlled Generations (ICCV 2025) · better on 2 of 4 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
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.90)
- Inverse Flow and Consistency Models (ICML 2025 · similarity 0.90)
- Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching (NeurIPS 2024 · similarity 0.89)
- Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.89)
- Solving Inverse Problems with FLAIR (NeurIPS 2025 · similarity 0.89)
- Fast samplers for Inverse Problems in Iterative Refinement models (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.
