This LoRa creates platform shoes/heels that are often called "pleasers" or "stripper heels". I have a fetish for these heels and have done a lora for them in SD1.5 and SDXL and now for QWEN haha. Hopefully the toe issues are now finally gone with QWEN being such a powerful model.
Although QWEN has an understanding of these heels, it struggles to produce them consistently with prompting. Its very random and rare you see these type of heels with prompting for them. This LoRa should help with consistency.
NOTE: You should set the LoRa weight at a very low value, the QWEN model already understands the shoe/heel, it just needs the LoRa to gently guide it to it. A weight below 0.5 is likely fine.
If you prompt for poses with feet up lifted up vertically you are going to get deformities. Even without this lora you will get deformities with heels up. The base QWEN model was not trained on feet up with heels on, or very little. I've seen people make perfect pictures with bare feet up, so likely the base QWEN model was trained on bare feet up but not with heels in feet up, so lacks the understanding to generate it. If you want feet up you will have to stick to bare feet for now. I even had about 20% of my training data on feet up, but it likely needs a lora focusing on the concept all on its own to correct the issue, even then Im not sure that would be enough. We need a base QWEN model trained on NSFW stuff really to train NSFW loras correctly.
Description
2100 step version, this one is less overcooked, but slightly less detailed than the 3000 step version. You can use a weight between 0.4 to 1.0 with this 2100 step version and not overcook your image. I myself prefer the 3000 step version as its higher detailed. I think 2100 steps version is the more stable version though. 3000 version is overcooked. Take your pick...
FAQ
Comments (3)
damn! need a sdxl version
there is one, but sdxl is very bad with toes https://civitai.com/models/774728/platform-mules-sdxl
I have made a couple of heel loras for qwen, they do really well. I will upload them at some point, and my loras work good with yours. I also have some other loras I've made as well that work quite well. Not many people are making quality loras for qwen, which is odd to me because qwen is the best open source image model out there. It beats the brakes off of flux, sd, sdxl, illustrious, pony, none of them can even come close. Even with the 4 step lora which limits output variance its still miles ahead of everybody else with better prompt adhereance. WITH THAT SAID, though, you are right, qwen struggles with specific heel brands, but its worth noting, qwen can generate heels from any angle, no other model can seem to consistently do that.