mflux_steps.md

Setup the repo

git clone [email protected]:filipstrand/mflux.git
cd mflux && pip install -r requirements.txt

Make a run script

Name this anything, maybe flux.py. Make sure to update the two paths marked below.

import sys
sys.path.append("/path/to/mflux/repo/src")  # <- Update

from flux_1_schnell.config.config import Config
from flux_1_schnell.models.flux import Flux1Schnell

flux = Flux1Schnell("black-forest-labs/FLUX.1-schnell")

image = flux.generate_image(
    seed=3,
    prompt="A rocketship exactly in the style of the cartoon rocket emoji. The rocket ship is flying in the sky. The logo 'MLX' is written on the side of the ship in black letters.",
    config=Config(num_inference_steps=2)
)

image.save("/path/to/save/image.png")  # <- Update

Run the script

python flux.py

NB The model weights will be downloaded on the first run. This can take a long time.

FYI if like me you're using the Apple XCode-provided git, and don't use HomeBrew or MacPorts, you can just download git-lfs Mac - Intel Silicon zip. Unzip it, then do e.g.

cd ~/Downloads/git-lfs-*/
sudo ./install.sh

You should now have /usr/local/bin/git-lfs, and can now start the instructions above with git lfs install.

A few notes:

At first step make sure to install git-lfs or you'll get error "git: 'lfs' is not a git command. See 'git --help'.". The following works on Mac OS with homebrew installed:

brew update
brew install git-lfs
git lfs install
git clone https://huggingface.co/black-forest-labs/FLUX.1-schnell

This took about 2-3 hours for me. If it only takes a few seconds, git lfs above didn't work properly. You should see du -h FLUX.1-schnell/ report ~108GB of space use.

In the provided flux.py file you will have to change 3 paths total (sys append, flux location, and image save location).

Generation takes ~20 seconds on my Macbook M3 Max. Super cool!

@awni Thanks for providing the installation guide. I should have added something similar to the readme, but was so eager to get it out that I forgot :) Will add something similar.

@karpathy Awesome that you have tried it out and thanks for the kind words, it means a lot! You have definitely inspired the idea of writing this kind of project from scratch and keeping it lightweight and bare-bones. Also nice to see the speedup from more powerful machines!

In terms of the size of the weights, we can get away with only using ~34GB of space since it only loads the following tensors:

https://github.com/filipstrand/mflux/blob/5023437e89b0a54232e0ecd05da6447ed2ea738d/src/flux_1_schnell/weights/weight_handler.py#L10

In terms of the size of the weights, we can get away with only using ~34GB of space since it only loads the following tensors:

It might be better to use snapshot_download directly to pull only the files you need.

That will also save needing to install and setup git-lfs and the entire first step of cloning the repo. It will also save needing specify the path to the model in the generation script.

Here's an example of how that works.

@filipstrand also please feel free to borrow any useful setup notes from this gist to put in your README directly.

@filipstrand if you do follow awni's breadcrumbs & simplify the requirements, pls come mention that here, as well.

FWIW, my small variation on @awni's runner:

'''
Make sure you set `MFLUX_PATH` and `FLUX_SCHNELL_PATH` according to where you cloned these components.

Sample run: `time python flux.py fluxtest.png "Cartoon scene of the Boulder flatirons in winter"`
'''
import sys
from pathlib import Path

MFLUX_PATH = Path.home() / Path('src/mflux/src')
FLUX_SCHNELL_PATH = Path.home() / Path('src/FLUX.1-schnell')

sys.path.append(str(MFLUX_PATH))
output = sys.argv[1]
prompt = ' '.join(sys.argv[2:])

print('Output file (PNG):', output, '\n', 'Prompt:', prompt)

from flux_1_schnell.config.config import Config
from flux_1_schnell.models.flux import Flux1Schnell

flux = Flux1Schnell(str(FLUX_SCHNELL_PATH) + '/')

image = flux.generate_image(
    seed=3,
    prompt=prompt,
    config=Config(num_inference_steps=2)
)

image.save(output)

@awni Thanks a lot for the tip! This has now been updated in main.

@uogbuji Thanks for your provided example. I ended up going with what @awni suggested and used the snapshot_download feature. This also mirrors how a typical Huggingface Diffusers setup work, so I tried to stay as close to that as possible. If you want to have a look at the changes, they can be found in this commit: filipstrand/mflux@e3a85df

@awni Thanks a lot for the tip! This has now been updated in main.

@uogbuji Thanks for your provided example. I ended up going with what @awni suggested and used the snapshot_download feature. This also mirrors how a typical Huggingface Diffusers setup work, so I tried to stay as close to that as possible. If you want to have a look at the changes, they can be found in this commit: filipstrand/mflux@e3a85df

Love it!!

I'm adding M4 Max results: 18.7 seconds!

mflux-generate --prompt "Luxury food photograph" --model schnell --steps 2 2 1.75s user 10.71s system 66% cpu 18.721 total

@ivanfioravanti Added to the performance table!