AUTOSIGNAL 车智信号

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

Research Frontier研究前沿

OrbitQuant: Data-Agnostic Quantization Achieves 2-Bit Weights in Diffusion TransformersOrbitQuant：图像视频扩散Transformer实现无校准2比特量化

S 1.7 T1 1 sources1 个来源 R7-research

OrbitQuant is a data-agnostic weight-activation quantizer for image and video diffusion transformers that requires no calibration data, using randomized permuted block-Hadamard rotation to normalize activations across all timesteps and prompts.
The method achieves state-of-the-art results across FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX at several low-bit settings, including producing usable images at 2-bit weights (W2A4) where prior approaches collapse.
The rotation is folded into weights offline and cancels inside each linear layer, leaving only a cheap forward rotation on activations at runtime, and transfers seamlessly from image to video without per-modality tuning.

OrbitQuant是一个用于图像和视频扩散Transformer的无数据权重-激活量化器，无需校准数据，使用随机排列块哈达玛旋转来归一化所有时间步和提示词中的激活。
该方法在FLUX.1、Z-Image-Turbo、Wan 2.1和CogVideoX模型上实现最先进效果，在多个低比特设置下表现最优，包括在2比特权重(W2A4)下生成可用图像，而以往方法在此设置下会失效。
旋转离线折叠到权重中并在每个线性层内部抵消，运行时仅需对激活进行廉价的前向旋转，且无需每个模态单独调优即可从图像转移到视频。

hf-daily-papers