Research map / Image generation / Fast diffusion sampling
Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Self-Corrected Flow Distillation, which adds GAN, truncated consistency, reflow, and bidirectional consistency losses to latent consistency distillation of a flow-matching teacher to get consistent one-step and few-step image generation.
Problem
Flow matching models need many function evaluations for sampling, and existing distillation methods give blurry one-step or oversaturated/inconsistent few-step images when applied to latent flow models.
Approach
Applies consistency distillation to latent flow matching and adds a GAN loss on one-step outputs (against blur), a truncated consistency loss for small t (against oversaturation), a reflow loss anchored to the one-step prediction, and a bidirectional consistency loss, so one-step and few-step outputs agree.
Key results
- On CelebA-HQ 256, one-step FID 8.06 and two-step FID 7.67 versus 2-RF one-step 18.03 (rectified) and consistency distillation 41.34 one-step / 23.56 two-step.
- On zero-shot COCO2014, Ours-0.9B reaches FID 11.91 at one step and 11.46 at two steps, beating InstaFlow-1.7B (11.83) at one step.
- On zero-shot COCO2017, Ours-0.9B gets FID 22.09 (1-step) and 21.20 (2-step), with CLIP 0.313 and 0.317.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GAN loss to consistency distillation (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 41.34 -> 11.55 | Helped |
| Add reflow loss to CD + GAN (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 2 NFE | FID: 19.82 -> 11.07 | Helped |
| Add bidirectional consistency loss to CD + GAN + RF (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 and 2 NFE | FID: 10.62 -> 8.06; 11.07 -> 7.67 | Helped |
| Use consistency distillation alone instead of the full method (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 and 2 NFE | FID: 8.06 -> 41.34; 7.67 -> 23.56 | Hurt |
| Set GAN, reflow, and bidirectional loss weights to (1.0, 1.0, 1.0) (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 30.26 | Hurt |
| Set GAN, reflow, and bidirectional loss weights to (0.4, 0.4, 0.4) (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 12.22 | Hurt |
| Set GAN, reflow, and bidirectional loss weights to (0.2, 0.2, 0.2) (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 9.83 | Hurt |
| Set GAN, reflow, and bidirectional loss weights to (1.0, 0.1, 0.1) (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 11.2 | Hurt |
| Set GAN, reflow, and bidirectional loss weights to (0.1, 1.0, 1.0) (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 35.22 | Hurt |
| Set time-truncated threshold t_trunc to 0.2 instead of 0.4 (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 7.50 | Helped |
| Set time-truncated threshold t_trunc to 0.5 instead of 0.4 (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 9.55 | Hurt |
| Set EMA decay μ to 0.95 instead of 0.9 (ablation) | Self-Corrected Flow Distillation · CelebA-HQ 256, 1 NFE | FID: 8.06 -> 7.79 | Helped |
Surprising findings
- Applying the rectified-flow distillation stage hurts multi-step sampling: 2-step FID (30.85) is worse than 1-step FID (12.95) on CelebA-HQ.
- Consistency distillation alone degrades sharply as NFE increases (FID 41.34 at 1 step, 51.34 at 8 steps, 53.69 at 16 steps).
Limitations
- Loss weights and warm-up schedules are sensitive; weights of 1.0 for GAN, RF and BI give FID around 30 at all NFEs.
- Text-to-image evaluation is zero-shot on COCO after LAION training; the paper's compute hardware is not stated.
- Mixed evidence on sample quality: CLIP score of SD Turbo (0.330) is higher than the proposed method.
Experimental setup
Datasets: CelebA-HQ 256 LAION (aesthetic > 6.25 subset) MS COCO 2014 validation MS COCO 2017 validation
Benchmarks: CelebA-HQ 256 MS COCO 2014 (30K prompts, zero-shot) MS COCO 2017 (5K prompts, zero-shot)
Metrics: FID CLIP score precision recall inference time
Baselines: Rectified Flow (2-RF) Consistency Distillation / LCM InstaFlow UFOGen SD Turbo Guided Distillation
Builds on: consistency models rectified flow latent flow matching UFOGen VQGAN warm-up
Closest papers by meaning
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.92)
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.90)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.89)
- Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching (NeurIPS 2026 · similarity 0.89)
- How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · 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.