RT-Lynx: Putting GEMM Sparsity in the Right Way for Diffusion Models

RT-Lynx accelerates diffusion transformers by sparsifying activations instead of weights, targeting a more hardware-friendly path for inference.

The method adds a compensation mechanism and a LoRA branch to preserve visual quality while keeping image generation fast.

Custom fused CUDA kernels and 2:4 semi-structured sparsity help unlock Tensor Core efficiency and reduce GEMM cost.

Experiments on models such as Qwen-Image, FLUX.1-dev, and Z-Image show speedups with little or no noticeable quality loss.