Shrinking diffusion sampling from dozens of steps to four

What it does

Flash Diffusion is a distillation framework that trains a student denoiser to predict, in a single forward pass, what a standard teacher model would produce over multiple denoising steps. The README shows it applied to several popular backbones—SD 1.5, SDXL, SD3, and PixArt-α—and tasks beyond text-to-image, including inpainting, super-resolution, and face-swapping. The resulting models run in as few as four function evaluations (NFEs), and the authors claim state-of-the-art FID and CLIP scores on COCO benchmarks for few-step generation.

The interesting bit

Instead of fixing the timestep distribution during distillation, the training shifts it dynamically to force the student to handle specific noise levels. The student also learns to predict the teacher’s denoised output directly, not just the noise, which appears to be the trick behind the speed. Perhaps more practically, the authors emphasize that the method is LoRA-compatible and can accelerate existing community LoRAs without retraining them.

Key highlights

• Supports both UNet (SD 1.5, SDXL) and DiT (PixArt-α, SD3) backbones.
• Claims training takes only “several GPU hours” with fewer trainable parameters than comparable methods.
• Provides ready-made checkpoints for Flash SD, Flash SDXL, Flash SD3, and Flash PixArt on Hugging Face.
• Includes wrappers for custom conditional diffusion models and T2I adapters, not just off-the-shelf text-to-image pipelines.
• Can accelerate existing LoRAs in a training-free manner.

Verdict

Worth a look if you are building image-generation pipelines where latency matters and you do not want to abandon the Stable Diffusion ecosystem. If you are looking for a plug-and-play consumer GUI, this is still a research repository with training scripts.