Super-lightweight and easy-to-use terminal shell for Flux.1 image generation model, created for people who like light-weight solutions with full control over everything happening in the code.
Current state of the project: everything is working as expected; tested on my own machine, NVIDA RTX 3080 Ti mGPU (16GB VRAM), 32GB system RAM.
- Image generation on demand (obviously)
- Everythng is customizeable
- Fully-automatic file name generation system & file saving
- Fully-automatic call history logging (can be disabled during first launch or inside settings)
- Can be used on machines with borked / outdated GPU driver and on machines with low VRAM (don't expect any miracles though)
New pipeline configuration optionsLive image generator parameters change (without program restarts)
Open powershell / bash terminal inside project folder, execute following command, it should install most packages automatically:
pip install -r requirement.txtNote that I used Python 3.10.6 to run this project.
Before you get to play with image generator, the model needs to downloaded on your local hard drive. The download process is semi-automatic / guided.
Both 'schnell' and 'dev' models use 30~40 GB of free disk space by default (each).
- Open Flux1.dev webpage in your browser
- Log in or Sign Up, agree to model access terms
- Click on your avatar, open Settings page
- Select Access Tokens and press "Create New Token"
- In "token type" bar select "Read", add token name and click "Create token"
- Keep page open for now, you'll need to copy access token later
- Back in the project: open config catalog
- Open file "user_secrets.md", copy everything inside
- Create a new file "user_secrets.py" and paste text from a previous step
- Copy your access token from huggingface webpage and replace value inside marked quotes '...', save and close the file
- Open project folder with VSCode or whichever code editor you're using
- Run main.py script (I'm using Python 3.10.6 for package compatibility)
- Follow the instructions on screen (if any)
- After successful load of the Flux.1 model you'll be asked to "enter your prompt" - type your desired request. Empty and whitespace-only requests will be automatically handled by the program as invalid.
- If you wish to exit the program type "exit" instead of your request - or simply kill the process, should free-up system memory either way.
- If you wish to change flux call parameters (aspect ratio, image resolution, guidance scale, number of steps) - you can do this on startup, or in the main menu - type "menu" when the program asking for a prompt.
This option is available on program launch and in 'menu', right inside the terminal shell.
If you wish to change the DEFAULT parameters (instead of session-only settings), here's one of the methods on how you can do so:
- Open FluxConfigurationProvider.py
- Look for a line
def __init__ ...., in brackets you'll find call parameters. - Change paraments to your liking, description for each parameter is provided above, in the same file
Open settings and set value of "initilized" variable to False, i.e. initilized = False and then launch / restart the program.
Note that old model will not be deleted this way, in order to switch back you'll need to complete this process again.
If you want to add custom images to the Flux.1.dev model (i.e. fine-tune it) AND if you don't trust corporations with your data - there is a way to expand model locally, but you're gonna need at least 24 GB of VRAM to do so.
See: ostris/ai-toolkit & Youtube Guide
Quick guide on how to solve known issues.
If your issue is not listed in this section, please create "New Issue" on official GitHub page.
By default huggingface library caches all models inside C:\Users\%your_username%\.cache\huggingface\ catalog.
If you're using separate hard drives for storing files and OS and / or you're low on free hard drive space - you can change the default cache location by creating 'HF_HOME' user / environment variable.
Here's how to do this on Windows OS:
- Open 'This Computer' or 'My Computer'
- Right-click on empty space inside opened window, select 'Properties'
- On your left click 'Advanced system settings'
- In newly opened window look for 'User variables for %your_username%'
- Click 'New...'
- Inside 'Variable name' field input
HF_HOME - Inside 'Variable value' field input path to desired cache location, for example:
D:\NeuralNetworks\huggingface(note that 'huggingface' catalog has to be created manually) - Click 'OK', 'OK', 'OK', close the 'System' window
- Restart your VSCode to apply changes (or whichever terminal or code editor you're using)
If you already have something downloaded inside the default cache location - simply move all contents of 'huggingface' catalog into inside newly created 'huggingface' folder at your desired cache location.
By default pip should download package 'torch==2.0.1+cu117', where 'cu117' - CUDA driver version supported by MY graphics card, '11.7'.
I have no way to verify package / CUDA driver back-compatibility, so just in case, here's the instruction:
-
To find CUDA driver version of your graphics card open Command line terminal (cmd.exe) and type: 'nvidia-smi' (NVIDIA graphics cards).
-
Depending on that info you might want to downgrade or upgrade project packages to more suitable version. List of PyTorch ('torch') package versions can be found here (latest) and here (public archive).
-
If you're using older PyTorch versions like me, you might experience some problems with 'cuda' element 'missing' on program launch - this means you've installed package that was compiled without CUDA driver support, so you'll need to uninstall it and look for a package with CUDA driver enabled.
-
In my case I had to 'build' my install script from scraps:
## taken from the public archive page - this is how we find latest package version for supported CUDA driver
# CUDA 11.7
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1
## taken from the public archive page - this is how we find how to get package compiled with CUDA driver enabled
# CUDA 11.7
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
## Taken from the "get latest version" page - this is how we get correct command arguments syntax
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118will result in:
pip install torch==2.0.0+cu117 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu117This might be difficult to understand at first, but this is how things are with public PyTorch archive, unfortunately.