The wildest thing about textual inversion is that it's done without finetuning stable diffusion, meaning that all the concepts already exist in it, they just need to be discovered.

We're going to be doing archaeology on these models for decades.

The Stable Diffusion innovation just doesn't stop!

Here's a new open-source model from the Mona that produces seamless tiling images: replicate.com/tommoore515/ma…

We just added textual-inversion training in 🧨diffusers.

Textual Inversion lets you personalize #stablediffusion model on your own images with just 3-5 samples.

GitHub: github.com/huggingface/di…
Training: colab.research.google.com/github/hugging…
Inference: colab.research.google.com/github/hugging…

Interestingly, I think a language model can interact with the physical world indirectly through interacting with users. In principle, it can update its internal states based on such feedback, just like actual embodied agents. Its update rules can be reward or even survival based.

The seasons..
#stablediffusion

Released my 'Image Variations' version of Stable Diffusion. Get the code and models, along with some basic instruction, in my GitHub repo: github.com/justinpinkney/…

The unreasonable effectiveness of transformers is rather astonishing. SOTA performance in almost every subfield of AI (natural language, image modeling, RL, ...) using a single architectural idea.

Also, this is a pretty big step in terms of sample efficient RL 👏🔥

Magic 🧙‍♂️🌊