Training-free staged sampling method that speeds up pretrained flow-matching text-to-image diffusion models.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
MrFlow (Multi-Resolution Flow Matching) is an open-source implementation of a training-free acceleration method for pretrained flow-matching text-to-image diffusion models. It samples a low-resolution image, upsamples the decoded result in pixel space with Real-ESRGAN (x2), re-encodes it, injects low-strength scheduler-consistent noise, and runs a short high-resolution refinement. This moves most denoising cost from high-resolution to low-resolution sampling.
The README reports end-to-end speedups (including text encoding, VAE, super-resolution and diffusion passes) of about 8.25x on FLUX.1-dev, 10.3x on Qwen-Image, 8.79x on FLUX.2 Klein Base 9B and 21x on Z-Image-Turbo, with up to 25x when combined with Pi-Flow on Qwen-Image. Two core settings are provided: 12plus1 for aggressive acceleration and 20plus1 as a higher-quality operating point.
It is a Python library built on standard PyTorch and Diffusers, with no custom kernels, finetuning or learned upsampler. The repository includes reference scripts, model-family demos in examples/, a community contributions area, and a ComfyUI plugin. Users must supply their own base model checkpoints and Real-ESRGAN x2 weights, and the scripts require editing placeholder checkpoint paths.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Training-free, model-agnostic method that generates at low resolution, upscales, then does one high-res step. It speeds Z-Image Turbo about 21x, Flux Klein about 9x and Qwen Image over 10x; code and ComfyUI plugin released.
Closest neighbours by meaning, not just by category.
Model
Text-to-human-motion flow-matching model that generates full-resolution motion directly, with training-free joint constraints.
Model
Rank-asymmetric flow method for pixel-space image generation, incl. AsymFLUX.2 klein finetuned from FLUX.2 klein 9B.
Model · Meta
Meta and HKUST research method for fast artist-like 3D mesh generation using a MeshVAE and flow-based diffusion transformer.
Model · Microsoft
Microsoft's open-weights 4B model for image generation and editing, with a 4-step turbo variant.
Research · ByteDance
ByteDance research method for post-training image diffusion models using on-policy self-distillation from reward signals.
Model
One- and two-step image generation model distilled from Z-Image-Turbo using text-stream adversarial distillation.
