Updated: Oct 5, 2024
concept
v3 (Flux Dev)
Woah. Flux rips hard when training.
Best model was at 3.5k steps.
Keywords (Optional): eyes crossed, tongue out, ahegao face
Strength: 0.8 - 1.0
Edit: Apparently there was a bug in the training code that produces slight artifacts in the resulting images. I will probably update this model as soon as I get the time to train it again. Sorry about that.
v2 (Pony)
It's back. The myth. The meme. The ahegao.
This was once again a lot more work than I anticipated, but I got to try out some new techniques on this one, so I hope you like it.
New Pony Version
Much improved Dataset
New Captioning
DoRA Training
Embedded embeddings (for hopefully more flexibility in models)
v1 (SDXL)
The model is super flexible. Trained at multiple resolution around 1024px.
Trained on over 300 high quality images, so I had to leave my computer running for two days straight, but I think the results are worth it.
The keywords can be used pretty independently, except maybe "drool" since there weren't as many training images for it.
After a few requests I've added the training data for anyone who wants to create an update or a model for SD1.5 since I don't really have the time/motivation anymore :)
Also, I think I may have started a trend with those supermarket images 😅