My first version of a working lora based on my Medabots Checkpoint. Failed at least 10 times trying to get something that works and i believe that this is it. Captioned on 369 images as well. Works with a cfg of 7-7.5, clip skip 2, and around 25-50 sampling steps.
Need testers to try it out and see if it works. Previous attempts produced a black and pink image or non coherent images.
Currently this version works for me but if it works for others, i dont know so testers needed. please comment and post images of your results please and thank you.
Description
Initial Release
FAQ
Comments (5)
Wew, these example generated Medabots are amazing looking!
It works extremely well and enhances the checkpoint version when they are both used together. I have attempted to use other checkpoints with varying degree of styles showing up with most ending up as mecha girls and not looking like medabots.
For the first version (thanks for being first), definitely does not work on other Checkpoints without distortion, maybe the checkpoint can be turned into a LoRA directly (KohyaSS) for better ease of use?
Yeah it seems to not want to work well with other things. The Checkpoint and the lora pretty much need each other to work and im not sure how to make it more flexible to work with other things
@allelujahnightingale direct training into LoRA with an established anime checkpoint. If it works on one anime checkpoint directly it will work on most anime checkpoints. https://civitai.com/articles/1939/an-attempt-at-classification-of-models-sd-v15 https://civitai.com/articles/2724/come-to-the-conclusion-that-using-different-models-in-the-lora-trainer-doesnt-do-anything
atm I tried to extract the Medabots LoRA from the checkpointto see how it turns out. AnyLoRA works as well as RevAnimated, Darksushi and OnlyAnime works as well, while MeinaMix, Counterfeit, and Anythingv5 seems to fail. (don't quote me on this, needs more testing, but the "similarity hypothesis" stands)
32, 64, and 128 dimensions all yield good results (not 16 tho and nobody does 256), and I default to either FP32 (or for smaller file size BF16 and never FP16) https://civitai.com/articles/397/rfktrs-in-depth-guide-to-training-high-quality-models https://civitai.com/articles/1676/what-is-the-difference-between-fp16-and-bf16-here-a-good-explanation-for-you
Will need to blend 2 LoRAs (the experiement and your LoRA) together next since that works as well just to check. hopefully things can get better in the future with dynamic LoRAs
P.S check this out, vibes weird https://civitai.com/models/26415/a-certain-theory-for-lora-transfer
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.