About
This LoRA is intended to generate an undressed version of the subject (on the right) alongside a clothed version (on the left). Precede your prompt with the "onoff" keyword, and then describe your subject (including any clothes you wish to appear on the lefthand image).
Versions
The latest version 4.0 takes advantage of more recent optimization enhancements and seems more reliable than previous versions. However, it was trained at 640x640 pixels, and it also tends to oversaturate and increase the brightness of the image if you use it at a smaller resolution. Please let me know how it works for you.
v4.0 - Best at 640x640. Used the dynamic learning rate from the Prodigy optimization settings and a more controlled dataset to more reliably produce the desired configuration.
v3.26 - Best at 512x512. Increased the size and quality of the training set and captions compared to previous versions, but used a smaller network rank (still larger than version 1.x) to improve overall control
v2.39 - Larger training set than 1.x, larger network rank, can handle more subjects and styles. Might work better for more heavily tuned models like CyberRealistic.
v1.28 - Original version, best suited for individual female subjects
v1.50 - More heavily trained original version if you have trouble getting 1.28 to work with your prompt.
If you're looking for the right hand image to be wearing underwear instead of nude, try the dedicated On/Off (Underwear) LoRA.
Tips for best results
Prompt
Precede your description of the subject with "onoff," and mention the subject's clothing in the following sentence. Adding "side-by-side photo" or "split screen" can help steer your base model toward generating the right shapes for the LoRA to take hold.
You can try "onoff couple" and "onoff male" to work with subjects besides individual women.
Weight
In most cases the models should work at the default weight of 1. If the concept is not being applied on your image, and you've already tried adding "helper" phrases like "side-by-side photo" and "split screen," first try increasing the weight of the "onoff" keyword in the prompt rather than the weight of the LoRA; both methods can work, but an overcooked LoRA model is likely to warp faces.
Speaking of which, I recommend using high res fix for better faces.
Control Net
Check out Mastabe's image post for an example of what you can accomplish with Control Net for fine control over the side-by-side image placement: https://civarchive.com/posts/234028
Description
This version may be overbaked. It ran for 50 epochs vs the 28 epochs of the initial version.
FAQ
Comments (12)
Interesting idea to do this with a LoRA. I've been creating such images using 'X/Y/Z' plot. If I compare the two then X/Y/Z plot seems more reliable, though it is slower because it generates two images whereas the LoRA approach combines the on/off into one image.
Out of 60 images generated, using 3 different models and both lora versions, only 2 came out how it should be but even then the girls didn't look alike.
Thank you so much for creating this! Works very well. I am using a custom mixed model that is mostly Deliberate V2 but I did it with pure Deliberate and it looks similar. About half the time I get 3 panels instead of 2 but I like it. Sometimes the middle panel is 1/2 dressed? Haha. This is my new favorite Lora :-)
Note: I am NOT using the trigger word. I used 'side by side full body photo'
Prompt:
side by side full body photo <lora:onOffLora_v128:0.8>, of mature woman wearing a business dress, hourglass figure, saggy tits, huge nipples, simple background, sharp focus, 8k, studio lighting, photo realistic, highly detailed, canon R5, 20mm focal length. (highly detailed skin:1.1), masterpiece, best quality, photorealistic
Negative prompt: poorly drawn face, poorly drawn hands, poorly drawn feet, bad hands,deformed, young, (firm:1.1), obese, (bad-hands-5), anime, 3d, cgi, obese, beard, moustache, painting, plastic skin
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 5.5, Seed: Face restoration: CodeFormer, Size: 768x768, Model: Cutom or deliberate_v2
Can this handle Male subjects or primarily only Females?
curious, does it work with img2img?
This lora works great!
Only problem I've had is that it always seems to make bad/weird nipples
DAMN, this does indeed do a pretty good job. Thanks for making it.
I feel like this would nuke any other loras you might use with. Is that accurate?
Hi man, great work with this Lora!
Since you mention overbaking at 50 epochs, could you use a few more images, to have a 50 epochs model that's not overbaked?
I do have a few thousands onoff pairs, both photo and anime, and i'll happily send them if they could be useful.
is it possible to do this with img2img ?
Honestly impressed. This LORA actually has a decent rate of hits.
This is awesome! A good while back I trained a checkpoint with the same goal. This is so much better as a Lora, THANK YOU!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.