This checkpoint is based on RealismEngine v3 and merged with the 200 LoRAs I've posted so far. All LoRAs were trained on https://oneshotlora.com
Onedude is a checkpoint that excels at bringing out the characteristics of any given character LoRA while retaining the realism and artistic strengths of Realism Engine.
As is tradition, the images in this post feature my next LoRA that was not included in the training data set.
Description
This version is once again based on Realism Engine v3.0, but this time it's merged with the 200 LoRAs I've uploaded so far. The goal here is a checkpoint that excels at bringing out the characteristics of any given character LoRA while retaining the realism and artistic strengths of Realism Engine. All LoRA models were trained by https://oneshotlora.com
As always, any feedback is most appreciated!
FAQ
Comments (2)
I'm about to endeavor on the process of merging a metric shit-ton of lora's into a model. Could you offer some insight on your process? Did you merge all the lora's together first, then merge them into the model? Did you merge them in batches? One at a time? Any particular ratio you stuck with, like .3, .5. .7?
For this version I just used an even ratio across all the LoRAs (1.0 / 200 = 0.005). I used Kohya's sd-scripts (networks/sdxl_merge_lora.py) to do the merge, which merges each LoRA to the checkpoint in succession. Since all my LoRAs have triggers, I think much of the value gets diluted into weak/unused tokens in the resulting merge. I'm considering finetuning the result on generated images from each LoRA in the next merge. I can't seem to find much in the way of authoritative documentation on the process, so I'm making it up as I go along.