There are over 10,000 images tagged "bad anatomy" on various online boorus. These images often have related tags like "bad hands" and "bad feet" as well.
Similar to my Boring SD XL Negative LORA (https://civarchive.com/models/421889?modelVersionId=475811), this LORA was trained on all of those images tagged "bad anatomy" or similar. Therefore putting it in the Negative Prompt should tell the AI to avoid these things and thus improve anatomy.
Pros:
-Does seem to improve anatomy in my testing, though the AI still isn't perfect
Could be Pro or Con depending on what you want:
-Tends to generate more detailed backgrounds
-Tends towards a more detailed or more realistic look for Sea Art XL. For the Autism Mix version it tends towards a more anime look.
Cons:
-Definitely doesn't fix all anatomy issues
-Putting it in the positives can be detrimental
First version named "badanatomy_SDXL_negative_LORA_SeaArtXL_v1" was trained with Sea Art XL and is intended to be used with that model (https://civarchive.com/models/391781/seaart-furry-xl-10). Version for AutismMix posted that's intended to be used with that model (https://civarchive.com/models/288584?modelVersionId=324524). Feel free to request if there are any other models anyone would like this trained for.
Description
First version trained for AutismMix, filename badanatomy_SDXL_negative_LORA_AutismMix_v1
FAQ
Comments (13)
how is this supposed to be used in comfy? I have yet to figure it out for any of the negative lora's I've come across. no trigger words, no settings posted. im sure its great, I just need to know how to use it in my workflow.
From what I can tell, load the LoRA, and put things like 'bad anatomy' and 'bad hands' in the negative.
I think the idea is that this LoRA gives the model something definitive to associate with "bad anatomy," it has the tags "bad anatomy, bad hands," baked into it. So when you then put those tags in the negative prompt, the model has more context for them, as opposed to without it, most models aren't trained with adversarial datasets from what I know, or at least not a lot of them.
Thanks to Up_punchington for the good response here! Also on the other negative LORA I trained https://civitai.com/models/421889 someone else had asked the same thing and here's the response I gave there:
"searching for an answer the best I could find was: https://www.reddit.com/r/comfyui/comments/17fgxhr/comment/k6a77nf/
Specifically, the comments that said:
"... that's my question. a "normal" lora like this would go in the negative prompt with a positive weight (won't work in positive with negative weight).... so how do we do that in comfyui? CLIP is generic and doesn't specify negative or positive prompt"
"But it does.
All CLIPS go thru all PROMPTS.
It might just say generic PROMPT. But whatever flows into Negative for the Sampler is the NEGATIVE.
CLIP lora -> positive prompt -> sampler
CLIP lora -> neg prompt -> same sampler"
"oohhhh!!! d'oh! the LORA loader goes BEFORE the cliptextencoder lol. i was putting it AFTER and never tying it to the cliptext encoder."
I hope that helps! Best of luck with all your gens!"
@Up_punchington Thanks a lot for helping people out, really appreciate it!
I don't believe putting a LoRA in the negative prompt actually negates it (at least in automatic1111 versions of SD). You can, however, use a negative strength like
<lora:badanatomy_AutismMix_negative_LORA:-1>In fact, in the code, negative prompts aren't even parsed for Lora:
a for furry also work when yes how use?
This might be unintended effect, but putting it on positive prompt with minus weight(about -0.75) gives some random light effects to image. Quite fun to use
so is that LoRa for negatives? never used that
in the positive with negative weight
Hi, can you do it for the model(T-ponynai3)? I tried the AutismMix version but it will messed up the style. Thanks!
I want to thank you for making this Lora. This helped me make the characters I generated more real like how the artists drew them. Thank you so much.
You cannot use a LoRA in the negative prompt in a1111 webui. Instead, use it in the positive prompt with a negative weight (e.g. try -0.2 for a start, and try higher and lower: this LoRA is very strong!). Try positive weights too, I'm not too sure what's going on with this LoRA. It definitely has some strong effects, though!
i love you, thanks
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.