Model fails on high-variance pose transfer (Sitting -> Standing)

#2
by Repaltoofficial - opened

Hi @linoyts , loving the Qwen-AnyPose integration!

I’ve been testing the model with some structural pose changes and found a consistent failure mode.

The Issue:
When trying to apply a Standing Pose (Reference) to a Seated Subject (Target), the model completely rejects the prompt. Instead of hallucinating the new legs/posture, it defaults to "Pixel Inertia" and outputs the original seated image unchanged (see attached result).

The Diagnosis:
It seems the model hits a "hard stop" when the skeletal difference between the source and target is too drastic (e.g., >45 degrees).

The Solution:
I run a data ops team (Repalto) and we are building a "High Variance" evaluation set (50 pairs of Sitting-to-Standing, Lying-to-Jumping) to test these exact edge cases.

I’d be happy to contribute this dataset to the repo if you want to use it to benchmark or fine-tune the next version. Let me know if you'd like me to send the JSON/images over.

Best,
Wahaj Barlas

image

Sign up or log in to comment