About general ability

[lucasjinreal]

[encounter1997]

Hi, thanks for your interest.

We want to clarify that this is an inference-only demo, that only supports our trained styles (denoted as 102, 103, 106, .etc, and the corresponding style images are presented in the examples). As can be seen, the selected style is None, which means this is an inference with the vanilla MUSE-Pytorch model, without our trained StyleDrop. We will update the gradio demo to better clarify its usage.

The gradio demo does not yet support training & inference with custom images, but you can still refer to the instructions here in our readme to train your custom StyleDrop weight.

Welcome further feedback on training your custom StyleDrop weights!

[lucasjinreal]

@encounter1997 I see. When will a training support at least on my local machine?

[encounter1997]

Could you provide information about the specifications of your local machine? Our code has been tested on Ubuntu 22.04 with a GPU that has 24GB of memory.

[lucasjinreal]

@encounter1997 Will 16GB mem work?

[2blackbar]

how long it trains on 3090 , whats the max resolution?

[zideliu]

how long it trains on 3090 , whats the max resolution?

If you set config.sample_interval = False here, 10 minutes enough, the max resolution is 256x256

[zideliu]

@encounter1997 Will 16GB mem work?

Not very clear, u can use fp16 clip by set prompt_model,_,_ = open_clip.create_model_and_transforms('ViT-bigG-14', 'laion2b_s39b_b160k',device='cpu') here. Have a try