Reproduction difference about FLUX

[yz413]

Great work, and thank you for sharing this project. While trying to reproduce the results, I noticed that when using TaylorSeer versus not using it, the generated images (FLUX.1-dev) show noticeable differences. This is not quite as consistent as shown in Figure 4 of the paper, where the results appear almost identical.

Similarly, in WAN 2.1, I observed that the generated videos differ quite a lot depending on whether TaylorSeer acceleration is enabled or not.

To test this, I made a simple modification in the diffusers_taylorseer_flux.py file to disable TaylorSeer. Specifically, I commented out the code following the TaylorSeer settings and added a function to ensure randomness. All other parameters are default.

def seed_everything(seed):
    random.seed(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

I was wondering if the use of TaylorSeer might introduce a certain degree of acceptable quality variation. Could you please share your insights on this?
Thank you again for your excellent work, and I really appreciate your clarification on this point.

[Shenyi-Z]

#30 Just refer to this issue. The significant differences mainly stem from the fact that we did not fine-tune hyper-parameters of TaylorSeer for WAN 2.1. If you intend to apply it, further adjustments to relevant parameters are necessary. For example, setting first enhance to a larger value here can help achieve greater structural consistency in the model. The same applies to FLUX.

[yz413]

Thank you for your prompt reply! It does seem like there’s still a trade-off between quality and speed to some extent—for example, in cache_init.py, parameters like cache_dic['fresh_threshold'] and cache_dic['first_enhance'] may need adjustment.

That said, I do believe TaylorSeer is quite successful overall. Without directly comparing to the non-accelerated baseline, the speed improvement is significant, and it still maintains good output quality. A completely lossless acceleration would indeed be challenging for cache-based approaches and likely requires further exploration.

Thank you again for your efforts and for sharing this valuable work!