Baseten distills FLUX.2 for 2.5x faster images
Baseten has cut FLUX.2's sampling from 20 denoising steps to 8 using timestep distillation, generating images 2.5x faster without noticeable quality loss.
Inference provider Baseten has published a distilled version of the FLUX.2 image model on Hugging Face. The model needs 8 denoising steps instead of the standard 20, which makes generation about 2.5x faster.
The work builds on Distribution Matching Distillation (DMD) and its successor, DMD2. Rather than copying the teacher model step by step, the student model learns to match the teacher's overall output distribution. An auxiliary "fake" score model tracks what the student is currently producing, which stops the student from collapsing into one narrow style of output. DMD2 also adds a GAN discriminator, which helps the student recover fine detail and texture.
Baseten reports that updating the fake score model five times for each student update worked best. Updating both at the same rate made training unstable. A 10:1 ratio was stable but converged more slowly.
Baseten's evidence for quality is a side-by-side comparison on a single prompt, "a cat sitting on a chair." The distilled model at 8 steps looks comparable to the original at 20 steps, while plain FLUX.2 at 8 steps shows visible artifacts. The post does not report quantitative benchmarks.
Baseten describes the approach as more effective than DiT caching, which it says gives only about a 1.5x speedup. The two methods cannot be combined. Caching relies on adjacent timesteps producing similar outputs, and that stops being true once distillation spreads the steps far apart. The company says the broader technique can cut sampling to 4–8 steps for 2–3x speedups.