search

p1atdev / LECO

Low-rank adaptation for Erasing COncepts from diffusion models.
https://arxiv.org/abs/2303.07345
Apache License 2.0
307 stars 23 forks source link

[SDXL test]IndexError: tuple index out of range #15

Closed sdbds closed 1 year ago

sdbds commented 1 year ago

image

p1atdev commented 1 year ago

It's maybe fixed. Please try it out on the sdxl branch.

sdbds commented 1 year ago

It's maybe fixed. Please try it out on the sdxl branch.

thank u for replying！it fixed but there is another error

image

p1atdev commented 1 year ago

I fixed it at 4b19877cadf058825163576c7d6c0cf9bd8a8ffd. (Sorry for the inconvenience because I don't have enough GPU resources for testing...)

sdbds commented 1 year ago

I fixed it at 4b19877. (Sorry for the inconvenience because I don't have enough GPU resources for testing...)

i can make a pr with Paged optimizer which can save 75% VRAM if you need. it needs gradit checkpoint and bitsandbytes over 0.39.1.

i tested it only using 8GB in 1024X1024 SDXL with rank32 alpha 16

sdbds commented 1 year ago

I fixed it at 4b19877. (Sorry for the inconvenience because I don't have enough GPU resources for testing...)

OK,it works well now. only it suddenly uses bigger VRAM in dyn res... i had to use batch size 4 for almost 40G VRAM.

i think maybe dyn res is not use resulation for middle squre. for example resulation: 1024 it still use 512X576 to finetune

sdbds commented 1 year ago

image SDXL result is nothing changed.... SD 1.5 with same prompt works well

p1atdev commented 1 year ago

In previous test (https://huggingface.co/p1atdev/leco/blob/main/elon_musk_sdxl_09.safetensors), it only worked when applied at very high weight (such as 10.0). Please try higher weight if you haven't tried.

I don't know if it will work, but anyway it seems that its effect is less on SDXL than on SDv1.x or SDv2.x. (https://twitter.com/gcem156/status/1682020714579365891)

sdbds commented 1 year ago

In previous test (https://huggingface.co/p1atdev/leco/blob/main/elon_musk_sdxl_09.safetensors), it only worked when applied at very high weight (such as 10.0). Please try higher weight if you haven't tried.

I don't know if it will work, but anyway it seems that its effect is less on SDXL than on SDv1.x or SDv2.x. (https://twitter.com/gcem156/status/1682020714579365891)

i use weight 10 in comfyui(in comfy_ui,it max weight is 10) but it changed a little almost cant see Should i use higher guidance_scale?

i will test for more result.