This was phenomenally tricky to produce (at least for me) for a number of reasons:

1. The model (and any fine-tuned version I tried, both based on SD1.x and SD2.x) confuses the appearance of the subjects in #Schiele paintings with the appearance of the artist himself https://t.co/Lb4Hlkt1AA

2. Differently from the other images in my series, here the model tries to generate an image of Francesca Woodman, too. So the resulting subject is a mix of different faces/aspects.

3. Egon Schiele painted a lot of nudes and Francesca Woodman photographed a lot of nudes. And so the generated images tend to be NSFW. There are a lot of boundaries that must be set via prompt engineering.

4. Obviously, this subject does NOT look like him. That’s why I’ve been sitting on this one for weeks before publishing.

If anybody can do better and share the technique, that would be great. Otherwise, I’ll wait for #stablediffusion3 and try again.

“thus far no other organisation has followed” https://t.co/md5w421NOX

I’m referring to the amazing work of @toyxyz3
https://t.co/i0p3LLdpW1

One day we’ll laugh at the weird things we had to do on a lazy Sunday morning to generate images of people with anatomically correct hands.

#stablediffusion

@AdityaKanu_ You have provided the tweet below, and it was involuntary because you didn’t know that it would become useful material for my newsletter. https://t.co/XhIg6bvwFN

@TomLikesRobots Now do all the models in TV commercials and celebrities in tabloids.

@camenduru Finally, I had time for a proper review of @camenduru’s free adaptation of #stablediffusion and #A1111 for GDrive+Colab. It works really well, even in an inpainting scenario with Multi #ControlNet and custom models.

Obvs, I run some more weird tests for https://t.co/sKxnbqStfY https://t.co/z7CfElMBnF