This is tests of training a flux model, and all licensing conditions for the Flux Dev model apply. You can find the details here: https://huggingface.co/black-forest-labs/FLUX.1-dev. It's a great model!
A huge thank you to Datacrunch.io for generously providing their resources for training and support. I highly recommend them.
This is Checkpoint. CLIP and VAE is already on
This is my model trained on Flux, and it aims to create more diverse and realistic faces. The results show improvements in anatomy, facial features, and overall realism. However, a side effect is that noise appears in the images. This can be corrected with any upscaler - just find one that works for you. The model was trained on approximately 7,000 images over the course of several hours using powerful H100 cards. If the test results are mostly positive, I will train it on a larger dataset.
The number of training steps was limited due to budget constraints. Additionally, I used a small portion of my dataset, only 7,000 images, to reduce training time. When I get the opportunity, I plan to train a more refined model. You can use this model as you like, but please remember to adhere to the Black Forest Lab license terms.
I welcome any feedback in the comments - feel free to share any errors or shortcomings of the model. This is just an experiment.
Description
This is the gguf version. Be careful! The examples compare the full model and NF4. Note the captions on the bottom of the images
FAQ
Comments (8)
Thank you! This is by far one of the best, when not the very best, Flux checkpoints that I've tried to date. Looking forward to further developments from you. By the way, I'm using Forge and the 16-step Hyper Lora with it.
Thanks
Please try version 2. I'd like to hear your opinion. Is it worth continuing? I am now training a new model (version 3) for 15,000 images.
@Sa_May I'm just starting to have a look at it, took a while to download on my connection (would have really preferred a UNET version) the when I tried to load it in Forge (without clip vae and t5) it complained - worked after I loaded clipl, ae and t5. First impressions are a bit mixed, but I tend to prefer the GGUF V1 at this stage. V2 seems to have a habit of distorting hands, although the skin seems a bit better. Haven't looked at feet and lying down poses yet, I'll try to look soon and report back :-)
In what folder do you put this model?
This is exceptionally good - but I'm not switching to it because of the noise issues. Excited to see how this works when you sort that out, and you're doing very good work.
Please Post Wokflow
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.