Paper-Conference

We introduce ViDiT-Q, a quantization method specialized for diffusion transformers. For popular large-scale models (e.g., open-sora, Latte, Pixart-α, Pixart-Σ) for the video and image generation task, ViDiT-Q could achieve W8A8 quantization without metric degradation, and W4A8 without notable visual quality degradation.

We design MixDQ, a mixed-precision quantization framework that successfully tackles the challenging few-step text-to-image diffusion model quantization. With negligible visual quality degradation and content change, MixDQ could achieve W4A8, with equivalent 3.4x memory compression and 1.5x latency speedup.

We propose an efficient diffusion model evaluation method that condense the textual evaluation dataset, achieving evaluation quality with 5x larger data size.

We propose an adaptive inference method for 3D perception that could reduce 60% 3D voxels, 80% 2D pixels, achieving 5x FLOPs/Memory and 1.4x Latency.

We propose CodedVTR (Codebook-based Voxel TRansformer), which improves data efficiency and generalization ability for 3D sparse voxel transformers.

We propose DSA(differentiable sparsity allocation), which enalbes the differntial allocation for budgted pruning, acclearting the pruning process by 1.5x.