Pose transfer is solved for a single image. The harder production problem is doing it many times with the same character and having the character stay recognisably the same person across every pose.
This episode builds that production flow: a per-character director that takes a pose reference and returns that character in that pose.
What the workflow actually does
There are three conditioning streams and they have to stay separate:
- Character identity — from a portrait reference. Must not be contaminated by the pose reference.
- Pose geometry — from a pose reference image or a skeleton. This is the only place body position comes from.
- Scene and lighting — optionally fixed, so the character can be placed into an established environment.
The common mistake is conditioning identity and pose from the same reference. That works for one generation and drifts immediately after — because the model cannot separate "this is who" from "this is how they are standing".
The Pose Studio part
Pose Studio is the interactive half: you position a skeleton rather than hunting for a reference photo. That matters for production because finding a reference photo for every needed pose is the actual bottleneck, not the generation.
Skeletons also give you exact control of things reference photos do not: camera height, hand rotation, and the exact degree of a twist.
Where it degrades
- Extreme poses — above roughly 120° of limb rotation, the character's proportions start to distort because the identity conditioning fights the geometry.
- Heavy occlusion — an arm crossing the face is handled; two limbs crossing each other is where it breaks.
- Unusual body types — the model has a strong prior toward a standard adult proportion. Stylised characters (chibi, three-head-tall) need their own reference set.
Why this is a production flow and not a trick
The value is repeatability. Once the character reference and the environment are fixed, a pose is a single variable. That is what lets you shoot a hundred poses of one character in an afternoon and have them all look like the same person.
Files
- HF: https://huggingface.co/AI-KSK/minimax-h3-character-pose-director
- ModelScope: https://modelscope.cn/models/Sunjian520/minimax-h3-character-pose-director
Video walkthrough (Chinese): https://youtu.be/eQep5Vq03RA
Bilibili: https://www.bilibili.com/video/BV14cht6NE8D/
Notes
- ComfyUI workflow export; Qwen Image 2.1 and Pose Studio nodes required.
- Tutorial material, not a product.
- Independent AI-KSK release; not affiliated with the model authors.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)