Developer Xingyu-Zheng has introduced the MrFlow (Multi-Resolution Flow Matching) method, which allows for a radical acceleration of modern diffusion models by shifting the primary computational load to low resolution. Thanks to integration via custom nodes from user RealRebelAI, this approach has become available in ComfyUI for Krea-2 and Z-Image Turbo models.
What Happened
Developer Xingyu-Zheng introduced the MrFlow (Multi-Resolution Flow Matching) method, designed to accelerate image generation in diffusion models without the need for retraining. Simultaneously, user RealRebelAI released a specialized set of nodes for ComfyUI that adapts this method to the Krea-2 and Z-Image Turbo architectures. The technical process involves generating a base image at low resolution (512x512), followed by an upscale via Real-ESRGAN in pixel space, and a final high-resolution refinement stage.
Context
The traditional process of high-resolution image generation in SOTA models requires expensive and slow high-resolution sampling. The MrFlow method changes this paradigm by replacing heavy computations with cheap low-resolution modeling followed by high-quality detail restoration.
Why It Matters for the Industry
For the AI industry, this method offers a way to achieve manifold increases in inference speed (up to 10–25x), which could lead to a paradigm shift from direct high-resolution sampling to hybrid multi-resolution pipelines. This potentially reduces cloud inference costs for startups and allows multi-scale approaches to be standardized as a baseline optimization method for diffusion models in production.
Why It Matters for Users
ComfyUI users can generate high-quality high-resolution images significantly faster with fewer computational resources. Using ready-made nodes for Krea-2 and ZIT allows for the instant implementation of ultra-fast pipelines into current workflows, substantially reducing latency during local generation.
Sources
Author
Look at AI, Editorial Team