Does it work with SDXL?

@v0xie The implementation you linked does not use conv_dim and conv_alpha arguments?

I haven't Implemented the glora for convolution since it has more than one way to implement the W' = A + WB. (For A and B are both low-rank matrix, in Conv layer. WB have more than one option)

I think I will directly impl B as linear layer between channels directly when I have time

@oliverban Yes, I've only tested it with SDXL

@KohakuBlueleaf Thank you! Even without conv layer, GLora is making way better images than any normal Lora I've trained.

@v0xie That's great!
I'm investigate the GLoKr things (mentioned in our paper)
I think I will release it in lycoris 2.0.0

Thanks for your work, @v0xie!

GLoRA looks promising, but I can't make the inference work with SD1.5. I only get seemingly random noise like the image below. The training with bmaltais' kohya_ss goes well, the samples generated during the training are looking good, your new module is correctly called, and yet…

Do you have any idea what could cause the issue?

Thanks for your work, @v0xie!

GLoRA looks promising, but I can't make the inference work with SD1.5. I only get seemingly random noise like the image below. The training with bmaltais' kohya_ss goes well, the samples generated during the training are looking good, your new module is correctly called, and yet…

Do you have any idea what could cause the issue?

That's super strange that's it's doing that. I hadn't tested SD1.5 with GLora before so I just trained one with latest dev branch of sd-scripts to see and it appears work correctly.

Can you try this LoRA and see if you're able to generate images with it? https://huggingface.co/v0xie/sd15-glora-monster_toy/blob/main/monster_toy_sd15_glora-000010.safetensors

Both the picture and safetensors have the metadata embedded with training/inference settings.

Yeah, your LoRA works great. Thanks to the embedded metadata, I've been able to track down the problem, and it seems to be network_alpha. I'm used to keeping mine at 1, which caused the issue; everything's fine when it's the same value as network_rank (or close enough).

It's strange, though, because as I said, the samples generated during the training were looking good, so I don't think there is a problem with the training; and yet, I don't see how network_alpha can have any impact during inference…

For me, GLoRA looks totally fine during training previews as previously mentioned, but for inference, anything above 0.3 weight produces fried images or noise nonsense. I have tested multiple configurations of lr, dim and alphas, but the problem doesn't go away

I don't see how network_alpha can have any impact during inference…

It should be scaled in inference too, but this implementation misses scaling, so it should work correctly only with alpha==network_dim. Everything else will be broken.