Rank-asymmetric flow method for pixel-space image generation, incl. AsymFLUX.2 klein finetuned from FLUX.2 klein 9B.
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.
AsymFlow (Asymmetric Flow Models) is a research method for flow-based image generation directly in pixel space, without a VAE or latent space. It uses a rank-asymmetric velocity parameterization: noise is predicted only in a low-rank, patch-wise subspace while data is predicted at full dimension, and the full velocity is recovered analytically. The network architecture and the training and sampling procedures are unchanged. The paper is by Hansheng Chen, Jan Ackermann, Minseo Kim, Gordon Wetzstein and Leonidas Guibas (arXiv 2605.12964).
On ImageNet 256×256, AsymFlow-H/16 reports 1.76 FID with a JiT-H/16 network and 1.57 FID with an added REPA loss, ahead of the pixel diffusion models in the project's comparison table. The method also offers a way to finetune pretrained latent flow models into pixel-space models. The authors finetuned FLUX.2 klein 9B into AsymFLUX.2 klein for text-to-image generation. They report that it beats the latent base on HPSv3 (10.66 vs 9.50), DPG-Bench (86.8 vs 85.2) and GenEval (0.82 vs 0.80), with more realistic textures. Evaluations use 1024×1024 images.
The project page links an interactive demo hosted on a Hugging Face Space. A YouTube creator reports that code and training code have been released, but the provided pages give no repository, weights or license URLs, so these are not confirmed here. The results are the authors' own benchmark numbers.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Image generator that works directly in pixel space without a VAE or latent space. Architecture tweaks made it 40% faster and it gives more photorealistic results; code and training code released.
Closest neighbours by meaning, not just by category.
Library
Training-free staged sampling method that speeds up pretrained flow-matching text-to-image diffusion models.
Model · Meta
Meta and HKUST research method for fast artist-like 3D mesh generation using a MeshVAE and flow-based diffusion transformer.
Model
Pixel-space text-to-image diffusion model transferred from latent diffusion models, without a VAE, supporting 4K output.
Model
Text-to-human-motion flow-matching model that generates full-resolution motion directly, with training-free joint constraints.
Model
One- and two-step image generation model distilled from Z-Image-Turbo using text-stream adversarial distillation.
Model
Text-to-image foundation model family (1B, 2B, 5B) using Semantic-First Diffusion with separate semantic and texture streams.
