AsymFlow lifts latent models to pixel-space fidelity
AsymFlow introduces rank-asymmetric velocity parameterization to enable efficient, high-fidelity image generation directly in pixel space. By focusing noise prediction on a low-rank subspace, it allows developers to "lift" existing latent models like FLUX into sharper, more detailed pixel-space versions.
The "latent-to-pixel" bridge is a major win for visual fidelity, finally bypassing the blurring artifacts inherent to VAE compression.
- –Achieves a leading 1.57 FID on ImageNet, outperforming prior pixel-space diffusion models by a significant margin.
- –Low-rank noise modeling reduces the computational burden of pixel-space training, resulting in 40% faster convergence.
- –Successfully converts the 9B-parameter FLUX.2 klein into a pixel-space model that beats the base latent version in human preference evals.
- –Provides a scalable path for upgrading pretrained latent weights to high-resolution pixel outputs without requiring a full architectural redesign.
DISCOVERED
1h ago
2026-05-17
PUBLISHED
1h ago
2026-05-17
RELEVANCE
AUTHOR
AI Search