V2.0
Completely new training method and expanded dataset.
A LoRA trained on the pose/scene of a woman upright with a face in her ass!
All of the preview pictures can be imported into ComfyUI to see exactly how they were generated. No extra embeddings, upscalers or detailers! Just a hi-res fix to get to a reasonable resolution.
Example images from Automatic1111 generations are also included as the output differs slightly.
**Some models do not seem to generate the guy behind the girl, while others have no problem.
PerfectDeliberate-Anime and Dreamshaper seem to generate images well
epicRealism does not.
*The prompt from the example images matters a lot. Use the negative and follow the format of the positive prompt to get an idea of how to describe your scene.
Trigger word: face in ass
Add other terms as well such as "man behind her with his face in her ass" or "man's face in her ass".
Clip Skip: 2
Optimal Strength: 0.5-0.85
*more varied scenes start to generate as the strength lowers, but a decent output may take longer to achieve
**Realistic models may require a greater strength
***768x768 is the best resolution for generation in my testing, but definitely experiment!
I highly recommend extra embeddings, upscalers, and detailers to get the most out of your composition!
Description
Initial Release
FAQ
Comments (8)
WOW~ YOU r an Aladdin's lamp~dude~
Dude, check your file,PLZ. maybe you post a wrong one. It didn't work.
Would it be possible to post your workflow?
I did some tests with the first image in Automatic1111 and some things I noticed:
The prompt is very important, I would follow the settings from one of the pictures and edit the scene from there.
Results also seem to be model specific, if you try with perfectdeliberate anime, or dreamshaper 8, the images seem to be decent. For some reason epicRealism just doesn't want to play nice.
This LoRA was trained on a dataset of strictly 768x768 images, so that resolution may be the best as well. I am preparing another dataset for the portrait aspect ratio and want to try merging the two to see if it improves.
@Questionable_Models After many tests, the summary is as follows:
1,SDXL version of the lora, generate a very low percentage of valid images, weighted at 0.8 or less, it is almost impossible to generate valid images. Only when the weight in 1 or more may have a 10% chance of generating success.
2,1.5 version of lora, in SD weibu can not generate any image, the image is a pink color, no image. The chance of generating a valid image on ComfyUI is close to SDXL, the success rate is very low.
3, the prompts used your prompts exactly as they are, no modifications. The SDXL test model uses animagineXLV3_v30, aamXLAnimeMix_v10, ponyDiffusionV6XL_v6, dreamshaper8.0, except for dreamshaper, which has a higher success rate, and the others need to increase the lora weight to 1.0 or above to have 10~20% chance of success. If you use your recommended 0.5~0.85, the success rate is extremely low.
4, using your picture directly into ComfyUI, using the same workflow as you, the effect is similar to my own workflow effect. The result is still not ideal.
@BIG_A Thank you for taking the time to document your tests. It definitely sounds like something may not be right with the generation, especially if the SD1.5 version only generates a pink color instead of an image in WebUI, and the images imported into ComfyUI do not generate the same image.
Using WebUI on my computer, I was able to generate the pictures I posted in the gallery with the same model downloaded from this page, and I re-downloaded it to make sure it was the same one and compared the hashes of the files.
I've been trying to replicate the failures on my end for a few hours, and I can definitely see epicrealism does not want to generate proper outputs, but the other models I tested, and the XL models (aamXLAnimeMix_v10, JuggernautXL) I tested seemed to have a fairly good success rate.
I did take notes from my epoch tests, and epoch 12 (the one chosen) had about a 40% success rate on multiple test runs at about 0.7 strength on my computer using perfectdeliberate_v20 (the anime version). I did a few hundred generations with each seed at multiple lora strengths (0.5, 0.6, 0.7, and 0.8).
From what's described, unfortunately it sounds like something else might be going on.
@Questionable_Models maybe you are right, so it's fine.I will test more to fine out what exactly is going on with me. THX ,dude.
This new SD1.5 version should be more consistent :)