This is a Lora trained in densepose images.
recommended to use densepose images at 1024x1024px resolution.
Description
Retrained on 1024x1024px. Must fix stretching vertically.
Its recomended to use 1024x1024px densepose image..
FAQ
Comments (14)
It works (much better than without it). The only thing is that it messes with people's proportions, stretching them vertically. I'll try to fix that in Photoshop.
The same.
It happens because it was trained with different resolutions. Ill try to fix it by retraining in 1024x1024
How do you generate those pose diagrams?
in comfyui with controlnetaux nodes
yeah something like this could be intensely powerful, if you package this up a little people will love it. Youre skipping controlnet entirely
@nucelar Does it handle... uh... Two people?
@nucelar We definitely need links to ressources, this is promising, thanks 🙌
@brnfd24434343d i didnt try it... it was trained on single subjects
@GlowingGuardianGirl what resources do you mean?
Guys you're gonna love this! Through trial and error I found out that by using this Qwen LoRa:
https://civitai.com/models/2191265/mannequins-pose-dummies-qwen-image-edit-2509
With this exact same prompt:
Generate a picture of abstract mannequins in the exact same pose as the people in the image. The mannequins should have Color-code body regions with artificial flat colors (yellow head, purple torso, green arms, blue legs, cyan lower limbs).
Uniform black background.. The mannequins do not have any hair. The background of the image is a plain featureless black void. Do not include any objects the people in the image may or may not be interacting with.,
You can create a sort of Densepose map. I will post my results in the gallery, but try this if you can.
How many image pairs did you use in the training?
good job, densepose is underused and underrated stuff. In case you are going to retrain, I'd like to ask you to diversify dataset from purely irl humans to cg/anime/etc stuff included in meaningful proportion, to mitigate current lean turning characters into humans (when generating instead of editing)
Ill try to expand dataset with objects, anime, etc... it was an expermient this model is trained basically in woman poses.