📢📢 "Align Your Flow: Scaling Continuous-Time Flow Map Distillation" New flow map framework for state-of-the-art few-step generation, w/ the amazing Amirmojtaba Sabour and Sanja Fidler. 🔥 Project page: 📜 Paper: 🧵Thread below... (1/n)

@FidlerSanja 🔸Consistency models enable distillation of diffusion and flow matching models into fast single-step generators, without adversarial objectives. However, they inevitably degrade in multi-step sampling, which we show empirically and also prove analytically. (2/n)

🔸Flow Maps (aka. consistency trajectory models) instead distill into fast generators that can jump from any noise level to any other noise level in the noise-to-data mapping, also enabling few-step sampling. They generalize consistency and diffusion/flow matching models. (3/n)

🔸We propose "Align Your Flow" (AYF), a continuous-time flow map framework, with new objectives, for scalable, state-of-the-art flow map distillation, enabling high-performance generation under any sampling step budgets. AYF generalizes previous and concurrent methods. (4/n)

🔸With an Euler flow map parametrization ("average velocity"), one of AYF's objectives corresponds to the concurrent MeanFlow Models ( a great paper! We focus on distillation, whereas MeanFlow - complementarily - focuses on training from scratch. (5/n)

🔸Further, we improve the teacher during distillation with autoguidance. Optional post-training adversarial fine-tuning can further boost quality with negligible loss in diversity. AYF achieves SOTA few-step generation on ImageNet benchmarks. ImageNet512 samples below. (6/n)

🔸Thanks to autoguidance, which is distilled into our flow maps, we can rely on parameter-efficient architectures (EDM2-S). Thereby, in terms of wall clock time, AYF's few-step generation remains faster than previous large single-step generators. Comparison to sCM below. (7/n)

🔸We also distilled FLUX.1-dev (credit to @robrombach, @andi_blatt, @timudk et al.) into a fast text-to-image flow map. This took only few hours on 8x A100s. We achieve SOTA few-step text-to-image generation among methods without adversarial objectives. See 4-step samples. (8/n)

🔸Credit to the brilliant @amsabour, who did all implementation and heavy lifting! Please see our project page and paper for details. Code coming soon! 🔥Project page: 📜Paper: More text-to-image AYF samples below. (9/9)

No jumping, No Running. Workouts at home at any time.🕒🏠 BEST 15 min Beginner Home Workout for Weight Loss 🧘‍♀️🔥

@CSProfKGD @amsabour @FidlerSanja Congrats @karsten_kreis! Inspiring work!

@amsabour @FidlerSanja Stellar work! CT flow-map distillation feels like the propulsion system generative models needed. We’re experimenting with similar few-step flows to simulate canopy microclimates & optimize autonomous greenhouses. Excited to dive into the paper—thanks for sharing!