Possible difficulty in adding X-Adapter/LoRA-X/DeltaSampling?

[TomLucidor]

Could SimpleTuner include LoRA knowledge being able to migrate from one model to the next (assuming we have base models)? https://github.com/showlab/X-Adapter https://ar5iv.labs.arxiv.org/html/2501.16559 https://github.com/Zhidong-Gao/DeltaSampling

[bghira]

DeltaSampling is an inference-time thing, not training.

X-Adapter never worked too well, and they already distribute the code in that github link that can be used for trying to train your own.

Overall, my view of X-Adapter was that it's a highly complicated piece of machinery that is more easily accomplished by retraining the LoRA. i think it's intended for adjusting models that the individual/organisation did not create and or no longer have access to the initial dataset for.

[TomLucidor]

it's intended for adjusting models that the individual/organisation did not create and or no longer have access to the initial dataset for

This is kinda something that would be really useful as time goes on for concepts with missing datasets. CivitAI situations and the like

[bghira]

i'm personally not a fan of trying to maintain models that are removed or you didn't have the original dataset for. you have the whole world at your fingertips - train a new LoRA! make a new dataset. but leaning on these X-Adapter stuff that a) don't work well, b) introduce maintenance burden to training tools like simpletuner, isn't going to happen (at least, not in this project / timeline)