Diffusers-FastAPI

The project is intended to provide Stable-Diffusion web service as simple as possible.

Features

• extra-long prompt overcoming clip model's 75-word limit
• • simply seprate prompt into slices under 75 words, and combine multiple prompt condition in diffusion
• txt2img
• img2img
• txt2img with prompt interpolation
• • this generate a sequence of image (video), smoothly alter from original prompt to target prompt.

Client

Reading codes in examples folder is enough for understanding APIs, maybe, I think.

"Ticket"

The idea is to memorize each image generation requests in specific IDs.

This may help future features like reproducing result from previous image-generation; change noise/latent by little scale each frame for image-sequence (video) generation. etc.

Additional scripts

model_interpolate_illustrator.py

usage: model_interpolate_illustrator.py [-h] --base-model BASE_MODEL --target-model TARGET_MODEL [--base-vae BASE_VAE] [--target-vae TARGET_VAE]
                                        (--match-start MATCH_START | --match-regex MATCH_REGEX) [--list-weights] [--dry] [--prompt PROMPT] [--width WIDTH]
                                        [--height HEIGHT] [--seed SEED] [--interp_steps INTERP_STEPS] [--infer_steps INFER_STEPS] [--output OUTPUT]

requires gifski executable Interpolates between base model and target model, generates an gif with model-mixing ratio from 0 to 1 example: