Exploration of Stable Diffusion spaces (prompt embedding and diffusion noise spaces) via pytorch, python code and examples.
By Giuseppe Insana, December 2022
torch, transfomers, diffusers, safetensors, PIL, matplotlib
Clone the repository and run the provided jupyter notebook.
Code and usage examples are all in the provided jupyter notebook. All the main functions have documentation on their usage and arguments.
Note: if the version on github is not rendering properly, try this one, via nbviewer.jupyter.org to preview the notebook before you run it on your environment (note: it jumps to usage examples).
- Simple text2img and repeated sampling
prompt2img([prompt1, prompt2], seed=seed, n_samples=3) - Interpolation between text prompts
interpolate_prompts([prompt1, prompt2], interpolate_steps=5)
- Walking in embedding space beyond the correct point produced by a prompt
images = beyond_prompt(prompt, walk_steps=8, walk_stepsize=0.01)
- Circular or spiral spherical walks through the diffusion noise space with 2 or 3 seeds
revolve_prompt(prompt, walk_steps=walk_steps, seed=seed, seed2=seed2)
Click to play - Multiple variations over a prompt
prompt_variations(prompt, variations=8) - Gradual interpolation between variations
variate_prompt(prompt=prompt, seed=seed, seed2=seed2, var_steps=6)
