Yihuang Chen

Tingyan Wen, Haoyu Li, Yihuang Chen et al.

Diffusion models achieve remarkable generative quality, but computational overhead scales with step count, model depth, and sequence length. Feature caching is effective since adjacent timesteps yield highly similar features. However, an inherent trade-off remains: aggressive timestep reuse offers large speedups but can easily cross the critical line, hurting fidelity, while block- or token-level reuse is safer but yields limited computational savings. We present X-Slim (eXtreme-Slimming Caching), a training-free, cache-based accelerator that, to our knowledge, is the first unified framework to exploit cacheable redundancy across timesteps, structure (blocks), and space (tokens). Rather than simply mixing levels, X-Slim introduces a dual-threshold controller that turns caching into a push-then-polish process: it first pushes reuse at the timestep level up to an early-warning line, then switches to lightweight block- and token-level refresh to polish the remaining redundancy, and triggers full inference once the critical line is crossed to reset accumulated error. At each level, context-aware indicators decide when and where to cache. Across diverse tasks, X-Slim advances the speed-quality frontier. On FLUX.1-dev and HunyuanVideo, it reduces latency by up to 4.97x and 3.52x with minimal perceptual loss. On DiT-XL/2, it reaches 3.13x acceleration and improves FID by 2.42 over prior methods.

Feice Huang, Zuliang Han, Xing Zhou et al.

Diffusion transformers have demonstrated strong capabilities in generating high-quality images. However, as model size increases, the growing memory footprint and inference latency pose significant challenges for practical deployment. Recent studies in large language models (LLMs) show that rotation-based techniques can smooth outliers and enable 4-bit quantization, but these approaches often incur substantial overhead and struggle with row-wise outliers in diffusion transformers. To address these challenges, we propose ConvRot, a group-wise rotation-based quantization method that leverages regular Hadamard transform (RHT) to suppress both row-wise and column-wise outliers while reducing complexity from quadratic to linear. Building on this, we design ConvLinear4bit, a plug-and-play module that integrates rotation, quantization, GEMM, and dequantization, enabling W4A4 inference without retraining and preserving visual quality. Experiments on FLUX.1-dev demonstrate a 2.26$\times$ speedup and 4.05$\times$ memory reduction while maintaining image fidelity. To our knowledge, this is the first application of rotation-based quantization for plug-and-play W4A4 inference in diffusion transformers.