SDXL Breeder

This is a tool for breeding Noken from SDXL. It's a work in progress.

We can generate many images per second -- this is one idea for exploring the latent spaces between prompts.

The idea

Explore sending nokens instead of tokens to CLIP for embedding.

Capabilities:

users can generate images from a prompt (seeding the universe)
users can generate images by selecting two existing images, which then mixes their embeddings (latent space of token embeddings sent to CLIP for guidance) to generate a new image

If you don't like the image, you can try again and get a different mix!

What?

Let's desconstruct what happens when you send a prompt to SDXL (Turbo):

images = pipeline(prompt)

SDXL converts the prompt into prompt_embeds and pooled_prompt_embeds - and diffusers exposes a method encode_prompt that does this:

prompt_embeds, _, pooled_prompt_embeds, _ = pipeline.encode_prompt(prompt)

Note: For this experiment we are ignoring negative prompts

While it is fun to play with this latent space (try replacing prompt_embeds with torch.randn_like(prompt_embeds)), what I want to do is deeper still.

Those prompt_embeds are 77-dimensional vectors generated by CLIP. SDXL utilizes two text_encoder/tokenizer - but the logic is very similar - tokenize the input, then call the text_encoder with those token ids:

input_ids = tokenizer.encode(
    prompt,
    return_tensors="pt",
    padding="max_length",
    max_length=tokenizer.model_max_length,
)

prompt_embeds = text_encoder(input_ids=input_ids, output_hidden_states=True)
pooled_prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds.hidden_states[-2].to("cpu")
return prompt_embeds, pooled_prompt_embeds

While we can easily generate the token embeddings using CLIP, we cannot send these embeddings (or modifications of them) into the CLIP Text Encoder Model.

def get_tok_embed(prompt, tokenizer, text_encoder):
    input_ids = tokenizer.encode(
        prompt,
        return_tensors="pt",
        padding="max_length",
        max_length=tokenizer.model_max_length,
    ).to(DEVICE)
    return text_encoder.get_input_embeddings()(input_ids)

With a small patch to HF Transformers we can send any token embeddings (input_embeds) we wish:

def encode_tokens(inputs_embeds, text_encoder):
    prompt_embeds = text_encoder(inputs_embeds=inputs_embeds, output_hidden_states=True)
    pooled_prompt_embeds = prompt_embeds[0]
    prompt_embeds = prompt_embeds.hidden_states[-2].to("cpu")
    return prompt_embeds, pooled_prompt_embeds

Remembering that SDXL uses two different text_encoders, we can send nokens to both to generate the prompt_embeds and pooled_prompt_embeds that SDXL needs:

def gen_prompt_embeds(tok_embeds):
    [lil_clip, big_clip] = tok_embeds

    pe1, _ = encode_tokens(lil_clip, pipeline.text_encoder)
    pe2, ppe2 = encode_tokens(big_clip, pipeline.text_encoder_2)

    prompt_embeds = torch.concat([pe1, pe2], dim=-1)
    pooled_prompt_embeds = ppe2  # SDXL only returns pooled prompt embeds text_encoder_2

    return prompt_embeds, pooled_prompt_embeds

And finally putting it all together:

prompt_embeds, pooled_prompt_embeds = gen_prompt_embeds(token_embeds)

images = pipeline(
    prompt_embeds=prompt_embeds,
    pooled_prompt_embeds=pooled_prompt_embeds,
    guidance_scale=0.0,
    num_inference_steps=1,
    num_images_per_prompt=num_images,
).images

This then allows us to modify the token embeddings we send in! SLERP? Genetic algorithms where users decide what survive to seed the next generate?

In backend/engine.py you can see a start to exploring this in the mix function.

Status

I've modified HF transformers to send token embeddings to CLIP instead of tokens - this opens up sending nokens!
Using SDXL Turbo makes this fast (enough) on MPS to start exploring
The engine currently works on the entire space of 77 tokens - playing with the current UI suggests working on individual Nokens could allowed controlled madness.
- instead of breeding the entire token embedding, breed noken(s) to use in prompts: "A drawing in the style of [noken] of a whippet sleeping"
the svelte app was created as a quick test - I've not spent much time on it

Installation

Backend

I highly recommend installing in a virtualenv as this requires a patched version of transformers.

cd backend
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Then to start the backend:

uvicorn main:app --reload --port 4444

Frontend

For the draft svelte frontend:

cd frontend
npm install
npm run dev

Then visit http://localhost:5173 to see the frontend.

click "new prompt" to generate an image from a prompt
click "breed" to generate an image by mixing the token embeddings of two existing images
drag and drop images to breed them
shift click an image to delete it

jakedahn/sdxl-breeder