Unifying Distributional Training for One-Step Visual Generation
Paper • 2609.35763 • Published • 2
Official resources for MGFlow: Unifying Distributional Training for One-Step Visual Generation.
Note Paper introducing the distributional training framework and MGFlow. Code: https://github.com/shihaoyang0423/MGFlow Project page: https://shihaoyang0423.github.io/MGFlow-website/
Note ImageNet and text-to-image checkpoints, fitted Gaussian references, frozen encoders, and evaluation assets.
Note Reconstructed COCO and GenEval reference images and prompts used for text-to-image post-training.
One-step text-to-image with MGFlow on FLUX.2 [klein] 4B
Note Try one-step text-to-image generation with MGFlow, powered by Hugging Face ZeroGPU.