Segmind has released the weights and training code for two compressed Stable Diffusion models that cut parameter counts dramatically without wrecking image quality. SD-Small and SD-Tiny drop 35% and 55% of the base model’s parameters, respectively, using a block-removal knowledge distillation technique. The pretrained checkpoints are live on Hugging Face, and the distillation code sits in a public repo.
The approach borrows heavily from a paper on architectural compression of text-to-image diffusion models. Instead of training from scratch, a larger teacher model — here, Realistic-Vision 4.0 — guides the smaller student through the denoising process. The loss function has three parts: the standard latent loss between target and generated images, a loss matching the teacher’s generated latents to the student’s, and a feature-level loss that compares outputs at every block of the U-Nets. That last piece is what preserves fidelity, since the student learns not just the final answer but the intermediate steps the teacher takes to get there.
Segmind trained both models on the LAION Art Aesthetic dataset, filtering for image scores above 7.5 to ensure high-quality captions. SD-Small ran for 100K steps, SD-Tiny for 125K steps, both on 1 million images. The company claims inference latency improves by up to 100% — effectively doubling speed — which builds on their earlier voltaML acceleration library that already delivered 4-6X speedups through AITemplate and TensorRT.
This matters because the past year has been a race toward bigger models: SDXL, LLaMa-2 70B, Falcon 40B. Segmind is pushing in the opposite direction, betting that smaller, faster, cheaper models will matter more for actual deployment. The open-sourcing angle is the real hook here — anyone can grab the checkpoints, run them through the standard diffusers pipeline, and get near-base-model quality at a fraction of the compute cost. Whether the quality holds up across diverse prompts and styles is the open question, but for developers building on consumer hardware or paying per inference, a 55% parameter reduction with comparable fidelity is the kind of tradeoff worth testing.
