Hugging Face 開源 SD-Small 與 SD-Tiny 知識蒸餾程式碼與模型權重
Original: Open-sourcing Knowledge Distillation Code and Weights of SD-Small and SD-Tiny
While Stable Diffusion (SD) 1.5 has demonstrated powerful image generation capabilities, its 860 million parameter count still presents…
Hugging Face 宣布開源基於 Stable Diffusion 蒸餾的輕量化模型 SD-Small（約 5.11 億參數）與 SD-Tiny（約 3.23 億參數）的權重與完整訓練程式碼。透過知識蒸餾技術減少 UNet 的層數，這些模型在保留高圖像質量的同時，實現了 1.5 到 2 倍的推理加速。此舉不僅降低了消費級顯卡與行動裝置運行 AI 繪圖的門檻，也讓開發者能自行蒸餾客製化的 SD 模型。
While Stable Diffusion (SD) 1.5 has demonstrated powerful image generation capabilities, its 860 million parameter count still presents challenges for edge devices, mobile platforms, and lower-end GPUs in terms of inference speed and memory footprint. To address this pain point, the Hugging Face team published research results on applying "knowledge distillation" to SD, and officially open-sourced the weights and training code for two lightweight models: SD-Small and SD-Tiny.
Free shows the 3-line summary; Pro unlocks the full deep summary (~300 words) so you never have to click through.
See Pro plans →Want the original English / full article?
Read on Hugging Face Blog →Summaries are AI-generated; the original article is authoritative.