OrbitQuant has been introduced—an innovative post-training quantization (PTQ) method for Diffusion Transformers (DiT) working with images and video. By using RPBH (Randomized Permuted Block-Hadamard) rotation, the technology becomes data-agnostic, allowing for the stabilization of activation distributions without the need for calibration data.
What Happened
Developers have presented OrbitQuant, which sets a new state-of-the-art (SOTA) performance level for models such as FLUX.1, Wan 2.1, and CogVideoX. The technology enables efficient use of the W2A4 mode (2 bits for weights and 4 bits for activations), ensuring high generation quality even under extreme compression.
Context
Traditional quantization methods for diffusion models often face the problem of activation instability across different generation timesteps. This usually requires expensive recalibration for specific conditions, which complicates the model deployment process.
Why It Matters for the Industry
For the industry, OrbitQuant solves the problem of activation instability by allowing a single codebook to be used for all conditions without recalibration. This simplifies and accelerates the deployment of heavy DiT architectures on various hardware and reduces inference costs.
Why It Matters for Users
Users will be able to run ultra-powerful video and image generators, such as FLUX or Wan, on consumer hardware with less video memory (VRAM). This significantly lowers the barrier to entry for the local use of high-quality AI-generated content.
What Is Not Yet Known / Limitations
At the moment, the open-source code and model weights have not been published, which limits the ability to practically verify performance and integrate it into production environments.
Sources
Author
Look at AI, Editorial Staff