OrbitQuant: Efficient Quantization for Diffusion Transformers

🤖 OrbitQuant: Efficient Quantization for Diffusion Transformers

OrbitQuant has been introduced for the post-training quantization (PTQ) of Diffusion Transformers (DiT) used for images and video. Thanks to RPBH-rotation, the method requires no calibration data and sets SOTA on FLUX.1, Wan 2.1, and CogVideoX models, enabling W2A4 mode usage without loss of quality.

🌍 The technology addresses the problem of activation instability across different generation steps, simplifying the deployment of heavy models on consumer hardware.

👤 Powerful video and image generators will be able to run faster and require significantly less VRAM while maintaining high detail.

Source 1: https://arxiv.org/abs/2607.02461