• : Standard Approach
• : Augmentation Approach (More Prefer)
• Open In Replicate : Augmentation Approach

In this repo, you can control image's style with text prompts, such as (PIL.Image("doge.jpg"), "cheese cake").
I apply CLIP Loss on style transfer model's style code, then we can cost about 1 min(with gpu) to get folllowing results:

Image is in Exploring the structure of a real-time, arbitrary neural artistic stylization network

In this repo, we edit the style code S with CLIP Loss : CLIP_model(T(c, S), text_prompt).

For Augmentation Approach, we augment T(c, S)$ before compute CLIP loss.

I didn't upload some bad result implementation such as

• Multiple Content image c
• Noise Style Code S_noise = S * gaussian(mean=1, std=eplison)

I borrow

• Some codes from StyleCLIP, style-augmentation
• Base model : magenta arbitrary style transfer, CLIP
• Idea of augmentation after generate an image : CLIPDraw