Creates pose dummy mannequins from reference images.
The intended use case is to use the resulting images as pose references for image generation with e.g. Nano Banana Pro or for LoRA training.
Sample prompt:
Generate a picture of abstract CGI mannequins in the exact same pose as the people in the image. The mannequins should have fingers, toes, and an abstract representation of facial expression. The mannequins do not have any hair. They have perfectly clean gray skin with a black wireframe overlaid on top without any discolorations. The background of the image is a plain featureless middle gray void. Do not include any objects the people in the image may or may not be interacting with.You can get somewhat similar results out of base Qwen-Image-Edit with appropriate prompting, but it is a lot less consistent in my experience.
It seems to work okay with cfg = 1.0 and steps = 20. Higher cfg values can produce better results, but may affect how faithful the result is.
The 8-step Lightning LoRA can also be very viable and often produces cleaner results.
Version 2.0 notes
Version 2.0 for Qwen-Image-2.1 was trained on a slightly different dataset using the same control images as Version 1.0 but with target images generated by Version 1.0. These new target images are more stylistically consistent and have better spatial alignment with the control images. While I consider it an improvement overall, the Version 2.0 LoRA may behave a little differently than Version 1.0.
I trained the version 2.0 LoRA for Qwen-Image-2.1 for 100 epochs (1600 steps), but the final epoch had poor prompt adherence (the wireframe overlay was nearly impossible to disable, even with negative prompting) which is why I chose to release epoch 60 instead.
Description
Trained on the exact same dataset as V1 for Qwen-Image-Edit 2509.
I wouldn't necessarily recommend this version if you can run Qwen-Image-Edit because it seems Flux.2 Klein doesn't understand anatomy quite as well as Qwen-Image-Edit does. But it's probably close enough for many purposes.
Sample prompt:
Generate a picture of CGI mannequins in the exact same pose as the people in the image. The mannequins should have fingers, toes, eyes, and a detailed facial expression. The mannequins do not have any hair. They have perfectly clean white skin without any discolorations. The background of the image is a plain featureless shiny black void. Do not include any objects the people in the image may or may not be interacting with. Do not change the people's pose in any way.FAQ
Comments (10)
Thanks for the wonderful model. If you don't mind, I'll leave My Workflow here (made it during a moment of inspiration). Maybe someone will find it useful. Thanks for your work! ❤️ 👇
https://civitai.com/models/2471200
Thanks for sharing! For full disclosure: Version 1.0 of this model was trained exclusively on female subjects and unfortunately has a tendency to feminize male subjects (sometimes more, sometimes less). It also wasn't trained to preserve faces and facial expression very accurately. The latter was intentional since I mainly intended to represent the pose of the body and not the exact likeness of the person.
If you think there's interest in a version that addresses these weaknesses, I might have another go at it. The training data is pretty easy to produce and I threw this one together in just a bit over an hour without much planning.
@VV24 I wasn't aiming for absolute accuracy in the facial rendering (just a rough outline was enough). The model handles this well and your work on maintaining the pose identical to the reference image is simply superb! Thank you! 👍
会更新qwen image 2.1的吗？
I am planning to do a training run on Qwen-Image 2.1, hopefully soon. But I'll probably wait until there's a decent 3rd party image-to-image trainer available. Hopefully, fal.ai gets on that soon...
@VV24 Looking forward to it! Hope fal.ai adds support soon
@VV24 I’ve checked https://modelscope.cn/aigc/model-training. This is the official community site for Alibaba’s Qwen team in China. It already supports LoRA training for Qwen Image 2.1 image generation and image editing. I’m not sure if there’s an international version of this site; without it, translation might be tricky.
@liao942287300dog If you need it urgently for Qwen-Image 2.1, feel free to distill my LoRA yourself. Just run a handful of your own reference photos through it and that's your training dataset. This LoRA was trained on just 15 pairs of images so it shouldn't be too much work.
@VV24 It is incredible that such good results were achieved with only 15 pairs of datasets. I will try to train it.
@liao942287300dog I think it helps that this type of edit is pretty close to what a typical base model can already do. Training a custom LoRA mainly improves the model's understanding of what exactly you want it to do, so it produces more accurate and consistent results. And from my experience you don't need a large dataset to accomplish that.
Most of the control images in my dataset were AI-generated themselves by the way. Only 4 of them are actual photos.