Mix of custom Jab lora I am tweaking and the Illumination style found here: https://civarchive.com/models/36945/illumination-style
Does petite and voluptuous women really well
I've mainly been using https://civarchive.com/models/58431/darksun lately, would like to see if the lora works well with other Models and other loras so I would really appreciate if yall posted your generated pics
Description
I've mainly been using https://civitai.com/models/58431/darksun lately, would like to see if the lora works well with other Models.
FAQ
Comments (14)
Thanks for making this! I couldnt get the other one to work.
Which model You`ve used to generate your examples ?
Darksun
I can't get this to work. I get this error: (Any help?)
2023-06-27 19:29:50,043 - sd_dynamic_prompts.dynamic_prompting - INFO - Prompt matrix will create 1 images in a total of 1 batches.
INFO:sd_dynamic_prompts.dynamic_prompting:Prompt matrix will create 1 images in a total of 1 batches.
locon load lora method
loading Lora D:\AI\LORA-OFFLOAD\Stylistic\IlluminatingJabMix.safetensors: AttributeError
Traceback (most recent call last):
File "D:\AI\stable-diffusion-webui\extensions\a1111-sd-webui-locon\scripts\..\..\..\extensions-builtin/Lora\lora.py", line 253, in load_loras
lora = load_lora(name, lora_on_disk)
File "D:\AI\stable-diffusion-webui\extensions\a1111-sd-webui-locon\scripts\main.py", line 374, in load_lora
is_sd2 = 'model_transformer_resblocks' in shared.sd_model.lora_layer_mapping
File "D:\AI\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1614, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'LatentDiffusion' object has no attribute 'lora_layer_mapping'
I've only gotten a similar error with other Loras when I hadn't updated SD, extensions, or I was using the wrong Model version. This is based on 1.5 SD
@Corruptedcupid That sounds right, I use 1.5 exclusively. I'll try updating it to the new one that just came out.
I don't see any other generated pix though so its possible others are having the same issue....(I really want to use this one!)
That was good advice, I did a simple "git pull" and was able to get the Lora to load finally.
However I got some very strange outputs, not like yours and the weights weren't working as your chart depicted. I was only able to get it to "toonify" using your prompt but turning the weight up to 1-1.4 which then distorted the output. (see my uploaded pix) Any clue why it might be behaving like that. I assume I can get it to do my bidding with some clever prompting, but I'm curious why its not behaving like yours. (I use clip skip 2 and darksun like you mentioned)
@SL0TWRX Possibly but won't know until someone else comments
@SL0TWRX you are using really strong weights 1.3 is too much
@Corruptedcupid I agree, but if you look at my samples theyre all using different weights, at your example I cant really see the effect at all.
@Corruptedcupid Resolution seems to be playing a factor in this. what resolution did you train at? what would you reccomend? I inspected your sample and saw it was at 411x617, I tried that and the results do seem slightly better, but still not the Jab feel I want.
@SL0TWRX You're never going to be able to replicate what I generated 100%, I have different hardware. The lora is a merge, raw settings for all my examples are 512 x 768 Hi res fix enabled and upscaled by 1.5 making the image 768 x 1152 at final generation
I finally got this mostly working.
I think I was getting poor results because my SD install needed to updated (automatic1111), and my CFG was too high, Its usually 7 default, but I think I was going as high as 9. This model really needs to be at 5.
I also noticed that the choice of upscaler can DRAMATICALLY alter the look of the model, which often can be nice, but loses the "jab" lineart feel.
I found these upscalers provided the best results, but I can't choose a favorite.
DPM++ SDE , DPM++ SDE KARRAS, DPM Adaptive, DPM++ 2s a Karras, UniPc, Euler a, Euler, LMS, Heun, DPM2, DPM2 a, DPM++ 2Sa,
You might want to include this info in the description to save Jab Fans like me frustration.
I also noticed you are using NKMD4x upscale, I don't use that, but had a decent upscale result with R-ESGRAN 4x+ Anime.
I also notice the "mature" token makes the breasts comically large, not my favorite, making the model younger solves this.
Thanks very much for this, I would like to see you push this more. If you do it will probably be my #1 favorite LORA style and I might use it nearly exclusively, I just need to see how versatile it is when mixing other loras and other checkpoints other than Darksun. (Which is quite good)
5 stars.