I tried to improve in any possible way my Anime Kisses lora but all the efforts were in vain, so I decided to make the LoCon version. You can consider this the second version for that lora.
Pretty much superior to the lora version in any way:
Works with a much wider area of CFG and other settings.
More consistent.
Lower impact on art style.
Seems to be working decently also with realistic models.
I suggest high rateo (0.9-1.0) for more NSFW kisses, otherwise it's good arround 7-8.
Pictures are done with EasyNegative, kl-f8-anime2 VAE and various models.
Using LoCons is very simple, just install the extension on the webui and you can use it just like any other lora, so don't be afraid.
Description
FAQ
Comments (8)
Never usedd Locons, gonna try right away
💖💖💖💖💖
2 questions.
1. Is the lower impact on style because you diversified the dataset or because of the LoCon?
2. What optimizer did you use for training?
1 It's true I enlarged the dataset, but that wasn't helping since i tried it for updating the lora, with bad results. It either did nothing or deformed faces. I think being locon improved things, it is has lighter training then the lora but gets results, so this is why has lower impact on artstyle.
2 addam 8 bit, I didn't experiment much with that, I would like to know which one is the best for concepts
@JollyIm Thanks a lot for the answer, I'm working on a concept and trying to deal with the style problems, I'll give it a try.
Yeah, I get you, because at discord they are advocating on Adafactor for characters, so it made me curious about what you used. Thanks again.
@JollyIm I'm sorry, I tried recreating your training (from the metadata) but I'm doing something wrong, I'm getting really bad images and no concept at all. Would you be able to tell me if you had similar experience and what your convolution values were?
Im getting an error in the console
oRA model animeKissesLocon_v1(6a9f2d73d4e3) loaded: IncompatibleKeys(missingkeys=['lora_unet_input_blocks_1_0_in_layers_2.lora_down.weight', 'lora_unet_input_blocks_1_0_out_layers_3.lora_down.weight', 'lora_unet_input_blocks_2_0_in_layers_2.lora_down.weight', 'lora_unet_input_blocks_2_0_out_layers_3.lora_down.weight', 'lora_unet_input_blocks_3_0_op.lora_down.weight', 'lora_unet_input_blocks_4_0_in_layers_2.lora_down.weight', 'lora_unet_input_blocks_4_0_out_layers_3.lora_down.weight', 'lora_unet_input_blocks_5_0_in_layers_2.lora_down.weight', ...
How many images in the training set? What was the base model?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.