SD3.5-Flash: few-step distillation brings quality image generation to consumer devices
ducha_aiki · x · 2026-09-12
The ECCV 2026 paper SD3.5-Flash introduces an efficient few-step distillation framework that brings high-quality Stable Diffusion 3.5 image generation to consumer devices.
Key methods:
- Reformulated distribution-matching objective tailored for few-step generation of rectified flow models
- "Timestep sharing" reduces gradient noise; "split-timestep fine-tuning" improves prompt alignment
- Pipeline optimizations (text encoder restructuring, specialized quantization) enable fast, memory-efficient deployment
Large-scale user studies show it consistently outperforms existing few-step methods, spanning phones to desktops and making generative AI practical on-device.
More from Multimodal
- AI-generated figure skating sequence glides across a frozen palace courtyard — Icy-Description-9806 · 2026-09-12
- MiniMax H3 ref2vid demo shows reference-image-to-video generation — nazgut · 2026-09-12
- Chained GPT Image, MiniMax video and Astra to turn images into Blender 3D — Hailuo_AI · 2026-09-12
- Teaching image generation models to understand hex color codes, an ECCV 2026 paper — ducha_aiki · 2026-09-12
- Consistent multi-view 3D generation: using a multi-angle reference matrix — LCD-9 · 2026-09-12
- JD Cloud debuts Lingjing AIGC platform with infinite canvas; student made award-winning short film in 4 days — 量子位 · 2026-09-12