Probably the cutest Flux.1-Dev checkpoint ever made. So far!

This particular model is a result of fine-tuning on just 4% of my full custom dataset, which I have been incrementally building and curating for the past 10-14 months or so, while researching and accumulating knowledge about Machine Learning, in an ongoing effort to achieve the highest possible quality model, within this genre.

As of the time of writing, this dataset consists of around 50,000 rare, diverse, (semi-) hand-picked images of a very high aesthetic quality.

In the project's development up to this point, I have written >200,000 lines of Python code, tested countless configurations, techniques, model architectures and pipelines, implemented code based on several scientific papers, ported optimization algorithms, contributed to several high-profile open-source projects, spent thousands of dollars worth of my own money on GPU rental, storage, LLM services and hardware upgrades.

Furthermore, I have built several large datasets for, and trained, both Classification, Instance Segmentation, Image Quality Assessment, Image Aesthetic Assessment, Age and Gender Estimation as well as Super Resolution models, (not to mention the absolute myriad of experiments with numerous image generation models not specifically related to this dataset), all of which have served to optimize the process and quality of the data enrichment and dataset curation, distilling a collection of roughly 600,000 images so far, into its current size.

Despite not being a true full fine-tune, but rather a careful merge of several low-rank models trained on different parts of the dataset with multiple different base models, I find this checkpoint to be performing incredibly well, especially considering that only 4% of my data was used!

I have now reached a point in my journey with this project, where I feel confident that training on the current full-size dataset, will yield absolutely stunning results, creating a truly State Of The Art model, that is both flexible, easy to use and produces outputs of a quality that is completely unmatched in this scene. I simply lack the resources to carry this out in practice.

If you can be of any support in the facilitation of this, please reach out to me on
Session :

  0596f864ff539378ce36b8073e8cd87fd1dbc970c6cfa33dc217f5fb209d9aec65

Aestheticute - BETA