Version 2
Finetuned for over 9000 more steps with recaptioned images and some new images to help keep its head on straight. This is my best LoRA for Qwen Image Edit 2511. It also works with Qwen Image 2512, but I have a separate LoRA trained just on that model here.
Qwen Image Edit model is easy to produce bad skin and body hair. I think it is related to not only sampler/scheduler, but also the quantized versions of the model and the lightning 4/8 steps applied. I'm convinced that is what does it, but there are a handful of combinations that work well to get good images with 4/8 steps. Qwen Image 2512 doesn't have this issue for me and many more schedulers and samplers produce great results by comparison.
See my showcase samples for prompt help. Use strength 1.0 unless mixing with other LoRAs then 0.6-0.8 should work, and that's in general it seems.
Please share your results and feedback is appreciated.
Original post
Here's another LoRA for use with Qwen Image or Qwen Image Edit created with my 150 image training dataset
I'm getting great results with both models using this LoRA, perhaps better than the ZImage Base LoRA. This lora is getting most consistent results for having the head of the penis on the right way for any angle!
Description
Finetuned for 3000 more steps with recaptioned images and some new images to help keep it's head on straight. It's still frustrating in general but I think this one with Qwen Image 2512 gives the best results. If you want to edit existing image to show or fix their penis Qwen Image Edit 2511 works great with this lora as well. Use strength 1.0 unless mixing with other LoRAs.














