Alexandria Ocasio-Cortez, commonly known as AOC, is a well-known US politician. Here's my attempt at a SDXL version.
Description
FAQ
Comments (8)
Fantastic Lora, would you mind sharing or showing your data-set? How many images, similary? Very high quality.
I added a zip with the training data, it's 39 images with captions.
@echo_cipher Thanks, it really does go to show that the more high quality press photos someone has the FAR better the model will fair. Great data-set, very clear. Good expressions.
@echo_cipher Did you use Ai for captioning if so which tool, or did you do them all by hand?
@becausereasons Generally I use an automated tool for the first pass, then quickly look through them and fix any obvious mistakes. Eg. https://sd-caption-helper.vercel.app/ if you have Groq or OpenAI access, or https://github.com/jhc13/taggui to run a local model.
Another question for you, which repo/settings did you train with? This is genuinely the best likeness lora I've ever seen. I attempted a new Justin Trudeau version with more data, but it really turned out poorly (maybe due to the various years of his photos)
@becausereasons I used SimpleTuner and followed the instructions here: https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md
I used rank 8 (rank 4 also seems to work decently) for this one I believe, and also cropped the images to 1024x1024 which was a bit of a hassle. They claim image binning doesn't work very well for Flux, but 1024x1024 transfers well to other aspect ratios. The version I uploaded here was 4,000 steps, which is maybe a bit overtrained. I noticed that it does not do quite as well with complex poses and composition, but does very well with your typical "1girl medium shot" type stuff.
I trained another one to 2,500 steps with rank 4, and the likeness isn't quite as good, but it does have fewer of those issues.
This IS the best likeness lora I've used on Flux so far. Not only is the visual output great, the true glory is in the speed at which this works.
I genuinely have no idea how you did it, but please teach others if possible.
FYI: Sub 30 second generations with another lora also being used on FP8 Dev. Other likenesses take ~650 seconds at best with the exact same other Lora.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.