Nerogar
[Feat]: use lora for sampling?
Describe your use-case.
Sometimes I want to train a lora that will be used with other loras, and I want to see the results from sampling. So I don't need to open ComfyUI or A1111 to validate it.
What would you like to see as a solution?
Maybe an additional tab like additional embeddings, or attached to every sampling definition.
Have you considered alternatives? List them here.
No response
I tend to make a full checkpoint + lora or checkpoint + lycoris merge and then train that, but I'm having trouble merging the LOHA from OneTrainer currently. I don't know what's up with them, but I can't seem to merge them into base SDXL models using Kohya.
… to merge them into base SDXL models using Kohya.
This is off topic. Please use #396
Describe your use-case.
Sometimes I want to train a lora that will be used with other loras, and I want to see the results from sampling. So I don't need to open ComfyUI or A1111 to validate it.
What would you like to see as a solution?
Maybe an additional tab like additional embeddings, or attached to every sampling definition.
Have you considered alternatives? List them here.
No response
Unfortunately, the sampling in OTs is unlikely to ever match the quality of Comfy, A111, or other dedicated inference systems. This has been discussed previously, and from what I recall, we intentionally use a simplified inferencing process. I'm tentatively closing this issue as I'm fairly confident it's a "won't fix" situation. Even if we add support for multi-LoRA usage, the results still won't match the visual quality of other systems necessitating you still use them.
However, if matching the quality of A111/Comfy isn't your primary concern, feel free to reopen this. Just keep in mind that this specific issue is unlikely to be resolved.