Diffusion Transformer (DiT)

VC-Attention is a new low-bit attention kernel that optimizes video diffusion transformer inference by addressing value quantization errors and softmax bottlenecks. It utilizes value smoothing via k-means clustering and a fused probability casting method to achieve high fidelity and significant hardware acceleration.

The researchers present Self Gradient Forcing (SGF) to solve the lack of gradient flow in historical KV caches during autoregressive video generation. This method enables much more stable and consistent long-form video extrapolation without the massive memory overhead of full backpropagation.

This paper proposes ReChannel, a novel architecture that transforms text-to-image models from RGB generators into efficient dense prediction engines. By treating transformer tokens as spatial carriers for task-specific data rather than RGB pixels, the method achieves new state-of-the-art performance with much higher computational efficiency.