All showcase images were created in ComfyUI using an upscaling + inpainting workflow.
For best results, use adetailer and/or hiresfix.
V2 is out
What V2 offers
Better quality fishnets
Less artifacts (they're still present, just not as often)
Better solo focus
What V2 struggles with
multiple girls or non-solo focus
Torn fishnets
Bras
Prompt flexibility
Hands (sometimes)
Overall I think the quality of V2 is better than V1, although it has its downsides. V1 offers more flexibility in your prompts, while V2 gives a slightly more accurate look of fishnet material being pulled. In my opinion, the average fishnet bodysuit that V2 generates just looks better than those in V1. Choose either version depending on what you're using it for :)
V2 also less likely to add extra nips lmao
First lora pls go easy xd
Basically just made it because I was trying to get a character to pull open their own fishnet bodysuit but it wasn't looking right (kept adding a new layer of sheer cloth that she was pulling when there was only fishnet material visible on her body, kinda making it look like she was pulling off a sheer tubetop).
Tested it out on 4 models and it seems fairly consistent, although I suggest using the model's recommended settings. It's not perfect but imo it looks better than default on some models. It will occasionally add extra nips, likely due to the data it was trained on. I don't have enough buzz to re-train this model at the moment and haven't learned how to train locally yet, so hopefully for anybody using this it's good enough! :D
PS: I have not tested this with other loras so I don't know how compatible it is with others.
Description
Recommended lora weights:
For multiple characters:
0.4-0.6 OR using "side-by-side" in positive prompt, 0.6-1.0
For single character:
0.5-1.0
This version slightly changes the art style of the base model, likely due to the type of data it was trained on and my poor tagging/captioning of the dataset. The lower weights you use, the more of the original model's art style will be preserved, however the fishnet pulling will become less detailed.
This version seems to have a bit of a bias towards typical open fishnets rather than bodystocking-material fishnets. I don't really know how to tag for a specific type of fishnets, so if anyone has suggestions/recommendations please let me know.
This version also likes to slightly change the art style and give a slightly fuller body type sometimes. I'm aware of this and will re-train it. For now, either use V1 or drop the lora weight when using V2 if these changes aren't desirable. V1 also seems to allow more prompting flexibility than V2, at the cost of fishnet quality. Neither versions are perfect so pick your poison o7
I removed the underscores in the trigger word, don't need them anymore. Use tags such as 'breasts out', 'nipples', 'pasties', 'heart pasties', 'no panties', 'pussy', etc. for different results. Experiment :)
FAQ
Comments (2)
I think it's generally useful to lean towards "mesh top, mesh bodysuit or mesh stocking". Fishnet usually overemphasizes thigh stockings while "mesh" is a general pattern, technically you could use both to improve the understanding.
Actually I kinda think the limitations are more due to "pulling clothes /(material/)" as a concept since it's not that well internalized without using "torn clothing" alongside it. I'd suggest adding more pulled bodysuit variations into the dataset to improve the concept and understanding of pulling
layers ^(definitely recommend grabbing some Wilhelmina pantyhose pull gifs/screencaps since there is no better data for that concept)
I'm not gonna lie I've been neglecting trying to use "mesh" as a tag since Google typically shows more of a sheer material look, but the few images I just generated with it - even without the lora - looked basically spot on. Probably should've searched up mesh on danbooru.domai lol, great advice ty :)
I never played BD2 but have seen that gif everywhere and didn't know her name, will definitely use her in the next set of training data. I think training it on a better understanding of pulling layers is probably a good idea alongside training it how fishnet/mesh material should look while it's pulled. The main reason I made the lora in the first place was because checkpoints were typically good at making characters pull clothing, but when it came to fishnets it just didn't want to pull the fishnet material itself and instead decided to add a new layer (without additional prompting at least). Training it on other pulled bodysuit types is probably a good idea!