Could you provide a model and parameters to reproduce the davidwu effect in the paper?

[XiaoluDing]

Amazing effect!! could you provide a model and parameters to reproduce the davidwu effect in the paper, plz

[drboog]

I didn't save weights of that fine-tuned model, but you can try to fine-tune the model by your self. You can try slightly larger batch size and more iteration steps to ensure good details, then tune the hyper-parameters of fusion sampling at testing.

[ywdong]

I didn't save weights of that fine-tuned model, but you can try to fine-tune the model by your self. You can try slightly larger batch size and more iteration steps to ensure good details, then tune the hyper-parameters of fusion sampling at testing.

Which hyper-parameters of fusion is important? I test my own pic, with bs=2, iteration=100, other hyper-parameters is the same with your test code. But the image is not so good.

[drboog]

I didn't save weights of that fine-tuned model, but you can try to fine-tune the model by your self. You can try slightly larger batch size and more iteration steps to ensure good details, then tune the hyper-parameters of fusion sampling at testing.

Which hyper-parameters of fusion is important? I test my own pic, with bs=2, iteration=100, other hyper-parameters is the same with your test code. But the image is not so good.

First of all, you need to make sure that your details are well-preserved after fine-tuning, i.e. "a photo of S*" can generate expected details. Otherwise, you may want to change batch size, learning rate, and iteration number. I think you should try more iterations, if your batch size is 2.

After fine-tuning, please read the comments in the example, different hyper-parameters control different things. Basically, tune cfg, ref_cfg and refine_emb_scale. How to tune them is based on what your current generation looks like. You may want to use a loop to generate multiple images for your current hyper-parameter combination, to better understand and tune it.