Research map / Image generation / Fast diffusion sampling
Align Your Flow: Scaling Continuous-Time Flow Map Distillation
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 Align Your Flow (AYF), continuous-time flow map distillation with two new objectives (EMD, LMD), autoguidance-distilled teachers, and adversarial finetuning, achieving SOTA few-step ImageNet 64/512 and LoRA-distilled FLUX.1 text-to-image results.
Problem
Consistency models degrade as the number of sampling steps increases (error accumulation in multi-step sampling), and existing few-step distillation methods often rely on adversarial losses that reduce diversity.
Approach
Train flow maps f(x_t, t, s) that jump between any two noise levels along the teacher PF-ODE with continuous-time Eulerian (EMD) and Lagrangian (LMD) objectives, EMD generalizing both continuous-time consistency and flow matching losses; sharpen the teacher with autoguidance during distillation.
Key results
- AYF with adversarial loss reaches FID 1.32 at 1 step and 1.07 at 8 steps on ImageNet 64x64 (recall 0.65), SOTA among few-step methods.
- On ImageNet 512x512, AYF-S (280M) reaches FID 1.70 at 4 steps and 1.87 at 2 steps, on par with the best sCD models at 2 steps using about 18% of their compute.
- A user study shows clear preference for the LoRA-distilled FLUX.1 AYF model over LCM and TCD (SDXL) at 4 steps, comparable to FLUX.1 [schnell].
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase the number of sampling steps from 1 to 2 | AYF · ImageNet 64x64 | FID: 2.98 -> 1.25 | Helped |
| increase the number of sampling steps from 1 to 4 | AYF-S · ImageNet 512x512 | FID: 3.32 -> 1.70 | Helped |
| add adversarial finetuning | AYF · ImageNet 64x64 | FID: 2.98 -> 1.32 | Helped |
| add adversarial finetuning | AYF-S · ImageNet 512x512 | FID: 3.32 -> 1.92 | Hurt |
| use AYF-LMD instead of AYF-EMD (ablation) | AYF · Image datasets | sample quality | Hurt |
| remove adaptive weighting (ablation) | AYF training · Image generation | performance | No clear effect |
| use an autoguided teacher instead of a standard teacher (ablation) | AYF distillation · Image generation | sample quality | Helped |
| increase consistency-model sampling steps beyond the best step count | Consistency models · Gaussian data, c=0.5 | Wasserstein-2 distance | Hurt |
Surprising findings
- Consistency models' multi-step sampling provably degrades (Wasserstein-2 grows with step count) even with an arbitrarily accurate Gaussian-case model.
- AYF-LMD is more stable in 2D toy experiments but yields overly smoothed image samples on real image datasets.
Limitations
- One-step performance is slightly worse than dedicated one-step methods such as sCD or SiDA.
- A gap remains between AYF and its multi-step teacher regardless of step count.
- Focuses on distillation and assumes access to a pretrained flow-based teacher.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 64x64 ImageNet 512x512
Metrics: FID recall wall-clock time Gflops user preference
Baselines: sCD sCT SiDA LCM TCD FLUX.1 [schnell] Shortcut models EMD MultiStep-CD
Builds on: Consistency Trajectory Models continuous-time consistency models (sCM) EDM2 autoguidance flow matching
Closest papers by meaning
- How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · similarity 0.93)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.92)
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.90)
- Expanding Flow Maps (NeurIPS 2026 · similarity 0.90)
- Bellman Optimal Stepsize Straightening of Flow-Matching Models (ICLR 2024 · similarity 0.89)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · 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.