OrbitQuant enables data-agnostic quantization for image and video diffusion transformers

Researchers present OrbitQuant, a data-agnostic weight-activation quantizer designed to address the high inference costs of diffusion transformers (DiTs). The method bypasses range estimation by quantizing in a normalized, rotated basis using a randomized permuted block-Hadamard rotation.

• A single Lloyd-Max codebook serves all timesteps, prompts, and layers for a given input dimension.
• Weight rows are extended offline to absorb the rotation, canceling it inside linear layers so only forward activation rotation remains at runtime.
• The approach transfers from image to video with no per-modality tuning required.
• It sets the state of the art for post-training quantization across FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX.

The technique pushes PTQ of image diffusion transformers to W2A4 while maintaining usable generation quality.