ChebBooster: A Training-Free Approach for Efficient Diffusion Transformer Inference via Chebyshev-Inspired Extrapolation
ChebBooster claims faster Diffusion Transformer sampling without retraining the model.
The paper proposes a Chebyshev-inspired extrapolation method for DiT inference, using barycentric evaluation for numerical stability. It separates weight precomputation from a lightweight online step to reduce overhead. The authors report tests on DiT-XL/2, PixArt-$\Sigma$, and FLUX.1-dev, with up to 3.68x latency speedup and 5.12x FLOPs reduction versus training-free baselines. Source: ArXiv · AI/CL/LG's note.
The paper proposes a Chebyshev-inspired extrapolation method for DiT inference, using barycentric evaluation for numerical stability. It separates weight precomputation from a lightweight online step to reduce overhead. The authors report tests on DiT-XL/2, PixArt-$\Sigma$, and FLUX.1-dev, with up to 3.68x latency speedup and 5.12x FLOPs reduction versus training-free baselines. Source: ArXiv · AI/CL/LG's note.
score 5