Research map / Image generation / Fast diffusion sampling
Flow Map Distillation Without Data
Shangyuan Tong, Nanye Ma, Saining Xie, Tommi Jaakkola
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 |
|---|---|---|---|
| distill without any external dataset (prior-only sampling) | FreeFlow-XL/2 · ImageNet 256x256, REPA teacher | FID-50K: 2.16 -> 1.45 | Improved |
| add correction objective Eq. (11) to prediction objective Eq. (9) (ablation) | FreeFlow · ImageNet 256 (SiT-XL/2 teacher) | FID-50K | Improved |
| add correction objective with stronger decay (k=1.0) on prediction gradients (ablation) | FreeFlow · ImageNet 256 DiT-B/2 | FID: 43.53 -> 5.58 | Improved |
| add both objectives (k=1.0) instead of prediction only (ablation) | FreeFlow · ImageNet 256 DiT-B/2 | FID: 11.91 -> 5.58 | Improved |
| use only the correction objective (ablation) | FreeFlow · ImageNet 256 (SiT-XL/2 teacher) | FID-50K | Hurt |
| drop high noise levels from correction r range ([0,0.6] vs [0,1.0]) (ablation) | FreeFlow · ImageNet 256 DiT-B/2 | FID: 6.02 -> 91.82 | Hurt |
| shift r sampling to higher noise (LogitNormal mean 0.8 vs -0.4) (ablation) | FreeFlow · ImageNet 256 DiT-B/2 | FID: 6.24 -> 5.63 | Improved |
| limit guidance to interval [0,0.7] vs applying at all noise levels (ablation) | FreeFlow correction · ImageNet 256 DiT-B/2 | FID: 8.65 -> 5.63 | Improved |
| apply stronger decay (k) on prediction gradient with prediction only (ablation) | FreeFlow · ImageNet 256 DiT-B/2 | FID: 11.91 / 11.71 / 12.40 | No clear effect |
| train on data augmented to increase teacher-data mismatch (ablation) | data-based flow map distillation · ImageNet | FID | Hurt |
| use FreeFlow student to search noise for teacher (Best-of-N) | SiT-XL/2 teacher · ImageNet 256 | FID: 80 NFEs beats CFG at 128 NFEs | Improved |
| distill from REPA teacher instead of SiT-XL/2 teacher | FreeFlow-XL/2 (1 NFE) · ImageNet 512x512 | FID-50K: 2.25 -> 1.49 | Improved |
Closest papers by meaning
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.92)
- How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · similarity 0.92)
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.89)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.88)
- Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching (NeurIPS 2026 · similarity 0.88)
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.
