StyleGAN-nada
StyleGAN-nada copied to clipboard
rinongal
Migration with a specific content image
If I want to specify a specific content image for style migration, which parameter should I set? You don't seem to have provided a parameter for changing the content image.
I have tried a number of operations that only allow me to make style changes to these four images. In addition, I did not find a path to these four images, nor was I able to specify an image local to my device for the transformation
These 4 images are GAN generated. This isn't really a style-transfer method, so there's no real notion of content and style images. It's a zero-shot finetuning method for GANs. If you want to use it for Style Transfer, what you can do is use an inversion encoder to map your image into the GAN's latent space, then re-generate it using the fine-tuned GAN. One of our Colab notebooks should have example code for doing this, and we had a HuggingFace space that lets you do the same operation.
Sorry, I tried to run the colab you provided, but I found the following error when executing "Step 3: Train the model".
In addition, when I executed "step 4: convert the image to the new domain" in "step 4: generate samples with the new model", the following error occurred.
On the first error: I see that torchvision was updated to use "value_range" instead of "range", you'll have to either downgrade to an older version of torchvision from before this change, or update the save_images function (https://github.com/rinongal/StyleGAN-nada/blob/dc8406ae2173ad186f8f03f3cadf65e613ac9364/ZSSGAN/utils/file_utils.py#L33) to use the new API.
On the second error: You need to run the 3rd cell before you can run the 4th. The net object is created in the 3rd cell.