aroma

Simple diffuser daemon with utilities and web UI.

This is for users who

use huggingface/diffusers library & models
need txt2img loops running on local (especially with MPS) or colab need (some naive but maybe useful) features
- prompt weights as WebUI syntax using ()[](:weight)
- random prompt using {a; b: 1.5}
- random size image
- clip skip
- multistep highres fix
- textual inversions
- Load lora safetensors
and need simple UI
- bootstrap5
- gallery

Usage (Colab)

See aroma.ipynb. It'll configurate environment and run WebUI.

Note that you need ngrok to export web ui. Consider to set basic auth username & password & daemon password not to other person steal your GPU token and images!

Usage (Local)

Requirements

Bash
Micromamba/Mamba/Conda (mamba is recommended)
(For web UI) nodejs & npm
Models converted for diffusers library

Preparation

Create ./archives, ./models, ./outputs, ./state directories
- touch ./state/value.json
Link or put SD models in models directory
Modify config.json for init configuration
If you want to launch only daemon, use ./daemon/run.sh
- Ìt'll create conda environment and install required packages
- You can communicate with files in state directory
  - current_job.json: Current job values & info
  - last_job.json: Last job values & info
  - state.json: Current state of daemon
  - values.json: If you want to change some values, change this file. (JSON format, see config.json init values)
- All outputs are placed in outputs directory
To launch UI, prepare npm and run
- `pushd ui; npm install; popd; node main.js``
- You can turn on/off daemon in Web UI

Features

Daemon

Daemon will automatically run txt2img repeatedly without any rests. You can communicate with daemon using files in state directory.

WebUI

You can

Turn on/off daemon & see outputs
Change configuration for next generation
See images in simple masonry? gallery
Delete uneccesary images and archive them
Download some models from huggingface.co
Add archive ddownload page

Configurations

All default configurations are placed in default_config.json. You can overwrite the configuration by writing in config.json.

password: Daemon password
models_root, state_root, outputs_root, archives_root: Daemon directories. Please create them before run the daemon
image_format: image format. png/jpeg/webp are tested.
save_raw: If true, daemon will create not encoded image file and job infromation in output directory. In this case, someone can access the image via webui without password.
init_values
- model.path: model path related from models_root
- params
  - width, height: image size in integer it may be the multiple of 8
  - sampling_method: sampling method, e.g. DPM++ 2M Karras
  - sampling_steps
  - cfg_scale
  - prompt
  - negative_prompt
  - highres_fix: Array of highres fix config
    - scale: Optional, scale of width/height. One of this or width & height must be set.
    - width, height: Optional, New size of image, ONe of this or scale must be set.
    - stregth: Strengh of img2img
webui
- host: hostname
- port: port. If it is 0, use random port
- model_download_presets: An object of arrays. Key is huggingface repo id, and it's elements are subdirectories. For example, you want to add HF repo "asdf/test-model", put "asdf/test-model": [""]. If you want to add HF repo "asdf/zxc" and its subdirectory "model-1", "asdf/zxc": ["model-1"]

Notes

To use original SD checkpoint, run python ./convert_original_stable_diffusion_to_diffusers.py --checkpoint_path ./***.safetensors --from_safetensors --to_safetensors --dump_path ./extracted/path --half with https://github.com/huggingface/diffusers/blob/36f43ea75ab7cdf9b04f72bced0b1ab22036c21c/scripts/convert_original_stable_diffusion_to_diffusers.py

TODO

Non-safetensor LoRa Support
Optimize Lora Weight Change
Optimize LoRa Download
Support multi-lora and UI
Support Highres Fix UI

lumiknit/aroma