MrFlow: Training-Free Flow Matching Acceleration Hits 10x Speedup
pmttyji · reddit · 2026-07-05
The paper "Multi-Resolution Flow Matching" introduces MrFlow, a training-free, multi-resolution acceleration strategy for pre-trained flow-matching models. It rapidly generates main structures at low resolutions, uses a lightweight pre-trained GAN for pixel-space super-resolution, injects low-intensity noise for high-frequency resampling, and finally refines details at high resolution.
Achieving approximately 10x end-to-end acceleration on FLUX.1-dev and Qwen-Image, its OneIG metric stays within a 1% margin compared to pre-acceleration. It significantly outperforms other training-free methods and requires no custom kernels or additional training.
Related event: MrFlow: Training-Free Acceleration for Flow-Matching Models(2 posts)→
More from Multimodal
- Wan2.2 image-to-video Space trends on Hugging Face — JCisHERE · 2026-10-27
- PlayCanvas' open-source splat-transform tool hits ~35K weekly NPM downloads — willeastcott · 2026-10-27
- Creator builds Snorricam video series with Seedance 2.5, sharing prompts — LudovicCreator · 2026-10-27
- Gradium launches free webpage TTS Chrome extension powered by its 40M-parameter Phonon model — mattturck · 2026-10-27
- Vortex of Infinite Recursion prompt goes viral: images that spiral endlessly into themselves — azed_ai · 2026-10-27
- 170+ AI Creators Made One Collective Film, Premiering November 6 — Loo_Atreides · 2026-10-27