MrFlow: 10x Image Generation Acceleration Without Distillation
量子位 · wechat · 2026-07-08
A research team from Beihang University, NTU, and ETH has proposed MrFlow (Multi-Resolution Flow Matching). Utilizing a three-stage pipeline—"low-resolution drafting → pixel-space super-resolution → single-step high-res refinement"—it achieves an end-to-end 10.35x acceleration on diffusion models like Qwen-Image (compressing 49.32s to 4.77s) while keeping quantitative metric errors within approximately 1%.
Requiring no specific hardware coordination or distillation fine-tuning, the core principle shifts the computational focus from high-resolution sampling to the low-resolution phase, using only a single step of high-resolution inference to finalize details. Compared to solutions like feature caching, MrFlow demonstrates greater stability in scenarios requiring speedups of 4x or more.
The paper topped the HuggingFace Daily Papers on its release day and garnered over 200 GitHub stars within three days, marking it as a representative training-free method in the recent image generation acceleration landscape.
More from Multimodal
- A fine-tuned Krea 2 raw model produced a rainy-night driving scene — darlens13 · 2026-07-27
- Users ask whether Video2X can load custom OpenModelDB models — Used-Profit2355 · 2026-07-27
- A builder wants AI to reverse-engineer viral video effects into ComfyUI workflows — stale2000 · 2026-07-27
- Midjourney V8.2 adds personalization and shows off stylized image outputs — Mr_AllenT · 2026-07-27
- Midjourney’s image variety draws a Krea 2 comparison and asks how to reproduce it — diffusion_throwaway · 2026-07-27
- AI short film sets a 1985 dystopia to music and leans into cinema — ProfessorKey98 · 2026-07-27