DSAQuant proposes stage-aware quantization for video diffusion

The authors of the paper on arXiv showed that conventional quantization-aware training degrades precisely the details and sharpness of video models, while preserving overall composition and motion. The reason is that quantization does not distinguish between denoising stages: early steps build structure, while late steps are responsible for textures. DSAQuant aligns training across these stages and disables CFG on the final steps so that quantization errors do not turn into high-frequency artifacts. The method was validated on the Wan and CogVideoX families in W4A4 and W3 modes, and the code has been released as open source.

What it means

This work concerns two generators from our reference — Wan and CogVideoX. Both are open source, both solve the text-to-video task, both have code under the Apache-2.0 license, and both run through the diffusers library. Wan from Alibaba is a large model: the largest weight file is 10.6 GB, and all files together total 64.3 GB. At the same time, the authors claim support for consumer GPUs: the smaller T2V-1.3B version requires 8.19 GB of VRAM and generates a five-second 480P clip on an RTX 4090 in about four minutes without quantization. CogVideoX from Zhipu AI is noticeably lighter in terms of weights — 9.2 GB for the largest file, 20 GB in total, but as described by the author, it targets devices at the level of NVIDIA H100 and above, with CUDA 12.4. For owners of consumer graphics cards, DSAQuant is another way to reduce the memory requirements of Wan and CogVideoX without changing hardware. The authors do not specify in the abstract exactly how much consumption decreases or how generation speed changes after quantization.