diffusion models

torch.load runs attacker code before the first denoising step

Moebius is a 226M-parameter inpainting model its authors say matches the 11.9B FLUX.1-Fill-Dev at 26ms per step via linear attention and latent distillation.