AsymFlow drops VAEs to generate sharper images directly in pixel space
Researchers just introduced AsymFlow, a generative framework that ditches latent spaces and VAEs to synthesize images directly in pixel space. By restricting noise prediction to a low-rank subspace, it boosts speed by 40% and achieves 1.57 FID on ImageNet. Why it matters: it eliminates the artificial plastic sheen of latent diffusion to yield crisp, photorealistic textures.