AI BriefingMrFlow Accelerates Diffusion Inference Speed by Up to 10x
Key point
MrFlow has been proposed as a technology that dramatically speeds up the inference of Diffusion models through step-wise resolution adjustment without additional training.
Details
Proposed to solve the blurring and artifact problems caused by existing multi-resolution generation methods, MrFlow is a training-free acceleration strategy.
The working principle is as follows:
- Low-resolution stage: Quickly generates the main structure.
- Pixel-space upsampling: Uses a lightweight pretrained GAN-based model to increase resolution.
- High-frequency resampling: Reinforces details by injecting low-intensity noise.
- High-resolution refinement: Finally polishes the details.
In tests on FLUX.1-dev and Qwen-Image, it achieved up to a 10x end-to-end acceleration compared to existing methods, while keeping image quality degradation (OneIG) within 1%. Furthermore, when combined with existing Timestep Distillation techniques, acceleration of up to 25x is possible.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.