AsymFlow drops VAEs to generate sharper images directly in pixel space

If you have generated AI images lately, you have probably noticed that unmistakable airbrushed plastic sheen. Skin looks unnaturally smooth, and fine textures often blur out. That visual flaw comes straight from the dominant playbook: compressing images into an abstract latent space to save compute, then relying on a VAE to decode everything back into pixels.

Researchers introduced Asymmetric Flow Models (AsymFlow), a new framework that skips the latent shortcut entirely. It generates images directly in native pixel space without ever touching a VAE.

Taming high-dimensional noise

Generating images directly at the pixel level has always been the holy grail, but historically it has been brutally slow and computationally expensive.

In raw pixel space, standard models waste immense network capacity trying to predict high-dimensional noise alongside the actual image data. AsymFlow fixes this by introducing an asymmetric parameterization.

Basically, the model restricts noise prediction to a patch-wise low-rank subspace while keeping the actual image data prediction full-dimensional. It then analytically reconstructs the full velocity trajectory without altering core network architectures.

By offloading the heavy burden of modeling chaotic noise across every pixel, AsymFlow boosts generation speed by 40%, effectively removing the compute bottleneck that previously held pixel models back.

Beating FLUX.2 at its own game

On the standard ImageNet 256x256 benchmark, AsymFlow scored a 1.57 FID, comfortably outpacing prior pixel diffusion setups like PixelDiT and JiT.

Even better, you do not have to train these models from scratch. The researchers figured out how to align this low-rank pixel subspace directly with the latent space of existing architectures. That allows developers to seamlessly initialize a pixel model from a pretrained latent checkpoint.

To prove it, they finetuned the 9B-parameter FLUX.2 klein into AsymFLUX.2 klein. The model retained its base semantic reasoning while drastically sharpening fine textures. On benchmarks like HPSv3, DPG-Bench, and GenEval, the finetuned pixel version consistently beat its latent predecessor while producing noticeably more realistic textures.

Why it matters

We think AsymFlow marks a pivotal moment for generative vision architectures.

For creators and digital artists, ditching the VAE means saying goodbye to blurry edges and fake, plasticky skin tones. Details like strands of hair, fabric weave, and intricate lighting bounce stay crisp because the model paints directly on the final canvas rather than through a lossy compression pipeline.

For the wider AI industry, this challenges the multi-year assumption that latent diffusion is the only viable path for frontier image models. By demonstrating that latent models can be finetuned into pixel-space powerhouses without throwing away existing pretraining investments, AsymFlow hands open-source builders a practical blueprint to bridge the gap between efficiency and photographic fidelity.