OrbitQuant introduces a data-agnostic post-training quantization method for image and video diffusion transformers. It uses normalization and a randomized permuted block-Hadamard rotation to produce a fixed marginal distribution, allowing one Lloyd-Max codebook to serve different timesteps, prompts, guidance branches, and layers with the same input dimension. Weight rows absorb the rotation offline, leaving only an activation-side rotation at runtime. The paper evaluates FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX, and claims state-of-the-art PTQ results across several low-bit settings, including usable image-DiT generation at W2A4.
No heat snapshots are available in the last 24 hours.