Nude males, 18-30 years old.
"Qwen Edit Anal v2" for "Qwen Image Edit 2511":
The same training data set as for the version v1 but more training.
Maybe worse for some use then v1.
Preview images has been made (converting them from 2d to 3d) from very good images from civitai. Thanks to all creators! Inpaint of a penis works also very good.
"Qwen Edit Anal v1" for "Qwen Image Edit 2511":
Preview images have used (not the best or fastest) workflow embedded
The training data was mostly anal penetration between guys and additionally some shemales and hetero sex. You can replace a pussy with a penis very easy. You can make a realistic version of an anal sex draw/sd image, you may need two editings: the first one to make a realistic version and the second one to make realistic penis, anal intercourse. You can make persons with breasts and penis. Because the focus was anal intercourse, complex poses without anal intercourse wont work immediately, you will probably (I am no expert) have to create an image and then edit relevant parts with this lora.
You can get vagina, but to get vagina you would have to describe a vagina in finest details.
Qwen Image V0.5:
61 training epochs with the same 30 000 images. I am afraid, the LoRa is overtrained now. I always get males even if I ask for females.
. I got good results with workflow from https://civarchive.com/models/1841581/qwen-image-wf
But I changed the node "ModelSamplingAuraFlow" value from 4 to 2-3. This workflow uses CFG = 2.5
The more details your prompt has, the better the result. I used Gemini 2.5 Pro for caption, so Gemini style prompts would be the best. If you use this LoRa without other LoRas, then you will only get males even if the prompt says woman
Qwen Image V0.3:
33 training epochs with the same 30 000 images
Qwen Image V0.1:
15 training epochs with 30 000 images
Description
FAQ
Comments (12)
this lora is so underrated, it is the ONLY USABLE penis lora for qwen edit so far. Thank you and looking forward your next update on qwen edit
I try to train a z-image turbo lora right now. Depending on result I will try z-image base or qwen edit. Could you post some examples of "before editing" and "after editing" images?
It would help me to understand the needs for a penis, testicles editing lora better. Thank you!
What is the trigger word? I cannot get it to work with Qwen Edit (both 2509 and 2511)
@focker123 I am sorry, I made this Lora for Qwen Image and never tested it with Qwen Edit.
I can't test it now because I train a new lora. Sorry
Works pretty well with Qwen-Edit-2511 at ~0.8 strength. Thanks! 🍆
It doesn't work for me with Qwen-Edit-2511. What is your setting?
@focker123 Using the RES4LYF "ClownsharKSampler" for the Sampler node
• eta: 0.50
• Sampler: linear/euler
• Scheduler: simple
• Steps: 8
• CFG: 1.0
• "ModelSamplingAuraFlow" node: Shift set to 3.00
• "CFGNorm" node: Strength set to 1.00
• LoRAs: penis_LoRA (strength 0.80) + Qwen-Image-Edit-2511-Lightning-8steps-V1.0-fp32.safetensors (strength 1.00)
@razr112 Did you use the fp8 version? I can't get it to work with "qwen_image_edit_2511_fp8mixed" and "qwen_image_edit_2509_fp8_e4m3fn" using your suggestions. My prompt is "remove the man's underwear to make his penis visible". All I got was deformed penises like the lora was not even applied.
@focker123 Nope, I use "qwen-image-edit-2511-Q8_0.gguf". Haven't tried it with 2509. How are you loading the LoRA model? Like which node are you using to load the LoRA.
@razr112 I get it to work with CFG 2.5 and lora strength 1.5
Here is my workflow: https://imgur.com/5l3F1Vm
Could you share your workflow? I wonder how does it work with strength 0.8 and cfg 1.0
@focker123 The core difference between my workflow and yours is how the images are being fed into the sampler. I'm not exactly sure how much that affects LoRAs, but the method I use definitely helps with avoiding other issues, such as pixel shift, so I recommend using it regardless. If you use the method I use, just make sure to scale the image to 1.00 total megapixels (and divisible by a multiple of 16) before VAE Encoding.
Here’s a diagram of the node connections I use for image input: https://imgur.com/XuSv3Tj.
Note that both “TextEncodeQwenImageEditPlus” nodes only receive a “clip” and “vae” input. No images are fed into those nodes with this workflow. I use the same method with Flux.2-klein (minus the Qwen Text Encode nodes).
You may also want to utilize the “Power Lora Loader” node for loading the LoRAs since it incorporates the clip model when loading LoRAs. I’m not sure how much that affects output, but I’m assuming it does affect it in some way. Plus it just makes using LoRAs way more convenient.
Also, I highly recommend using the “ClownsharKSampler” node that I mentioned previously for the sampler. It can noticeably improve output quality, especially when dealing with image editing models. It’s part of the “RES4LYF” custom node set, which can be downloaded via Comfy Manager.
Super underrated 10/10