Important to Read:
Qwen-Image-2512:
Check on "about this model" to find the links of the models recommended to achieve this quality, unfortunately any type of quantization downgrades the quality and worsen the prompt adherence, I recommend using my workflow, or adding the upscaling step to your own workflow. All images were upscaled and I've used the seedvariance node. On an RTX 4060 ti it takes around 2m15s every image, including the upscaling.
Recommended weights: (0.80 - 1.00)
Trigger words "erect penis exposed", "bare ass".
A recommendation:
It would be better to explore and give different instructions to the model to represent directions, further from my suggestions.
"erect penis exposed (orientation_back)" to show the penis from the back perspective, it may not work perfectly, and sometimes it works better only using "bare ass". There are also (orientation_front/side/up/down) but it doesn't seem to be followed at all, but you are free to test it.
It took around 21 hours of training (Onetrainer), 512 resolution, 152 images.
It can cause deformation and unwanted results. And I know the two 2D outputs aren't ideal, but it has a little flexibility for 2.5D styles (as much as the base does). And it does not perform sex or sucking, it can do sucking sometimes but mostly can't.
Zimage-Turbo:
To get the best results, you need to use this LoRA in order for it to work properly.
Recommended strength (1.00 - 1.20)
For optimal results, use: "erect circumcised penis exposed" but "erect penis exposed" should work. I recommend to use the workflow on the images to achieve high quality outputs.
The samples were generated with the "Qwen3 4b" text econder but for optimal results I recommend "Qwen3 instruct 2507 4b" since it was trained with it.
It took less than 6 hours of training (Onetrainer), 512 resolution, 94 images.
As always, remember that the model can produce some unpleasant results, struggle with some prompts, and in weird cases, cause deformation.
Description
FAQ
Comments (6)
thank you, where is Qwen3 instruct 2507 4b?
Hi, you can get the gguf version from Huggingface (https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/tree/main) I personally use the bf16 but I had to merge the files myself.
Thank you, I enjoyed this lora, seems to work best with solo images
Sorry but why do we need a distill patch lora? I test it but only get picture blurry. And I find on huggingface says it is for lora training.
Hi, I think I saw something about that, but some people started using it on Comfyui with their LoRAs, so I did it too with other LoRAs and it seems to work for me, but try my worflow (delete the seedvr2 if you don't want to use it) and see if there's a difference, just in case I do not use the "ModelSamplingAuraFlow" node.
@JKf_05 I confused it with the lora on Comfy HF. It worked after downloading the lora from your link.