关于Taylor参数

[Horizon2021-Li]

我用qwen-image测试了加速框架，确实能够加速生图速度，但是伴随着的是作图崩坏。后续有计划开放Taylor关键步数的参数调整吗？让用户可以容易调整多少步是全量计算，多少步是Taylor预测

[Tammytcl]

Thank you for your valuable feedback. We have also tested TaylorSeer on Qwen-Image and found that, indeed, its performance is not as strong in this setting. However, we have developed several other methods that show excellent results on both Qwen-Image and Qwen-Image-Edit. We plan to release the corresponding code after the papers are accepted.

You can refer to our latest works on arXiv for more details:

FreqCa: Accelerating Diffusion Models via Frequency-Aware Caching
https://arxiv.org/abs/2510.08669

Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion Transformers
https://arxiv.org/abs/2510.04188