If you like it when women are dressed and it barely covers anything and if you like curves and gloss, this model is exactly for you. Use the body descriptors according to your needs, sometimes personally i think the curves become a bit too much and it depends a bit on the poses you ask for. Its works also with character loras, the face and hair shouldn't be affected by the lora. Enjoy!
Description
I released the first model thinking i couldn't do better and that it was still nice to have for the few concepts it could do, i entirely reworked the dataset, im still learning how to make loras but i genuinely feel this version is state of the art, atleast its the best i ever made for responsiveness to prompts and adherence to concepts, sorry for the bait and switch, enjoy!
FAQ
Comments (4)
What is the difference between the first and second versions, and why did you remove v1?
The first version dataset had more images but believe it or not it was actually making the model worse, it made the model not very responsive to prompting by the user for no gain in accuracy. Many images of the first dataset were bunched in subconcept sets, i expected this would help the model but it didn't.
For this version, i kept the images in the dataset to a minimum, i deleted images and kept only what i felt was more in spirit of what i wanted to see for each subconcept that i wanted to be a part of the lora to be represented and reworked some of the training prompts. I removed the first lora because the new one makes it irrelevant now that i managed to make one that work as i wanted instead of what i could manage.
The first was only there as a ''its better than nothing''. I feel about my loras that way until/if i manage to remake them better. I feel wasting someones time trying to make it work instead of going straight to something that is responsive would be a bad experience so i delete the first, im still learning to make loras so when i put out new models they might not be quite responsive.
@maximalex213 I understand. But versioning isn't a matter of perfectionism. Here’s a real-life example: I downloaded the previous version of your LoRA—Version 1. If I hadn't noticed that you swapped the files, I would have been convinced I had the single, latest version installed. Even if my generations didn't match up, I would have looked for the problem on my end, never suspecting I was using a completely different version. In my opinion, when you make improvements, you should keep the previous versions available—or, if you consider them completely broken, simply add a note to that effect. 99% of people won't download anything other than the latest version (nobody has the time to check everything). If someone downloads a version you consider imperfect despite the warning, that’s no longer your problem—since you did warn them.
@Unicom Thats a valid take, i feel like however nobody reads the version notes because they are buried low in a corner of the page. Ill consider it for future models that get adjustements, thanks for sharing your thoughts on that!