Two pose directors, one model: how MiniMax H3 Image turns a photo into a controlled 3D pose
If you have ever tried to change a character's pose with an image model, you know the failure mode: the face drifts, the clothes change, and the pose you wanted is roughly what you got the third time you rolled the dice. This episode is about a workflow that removes most of that luck, and it runs on RunningHub as an online workflow rather than a local ComfyUI install.
Who this is for
- People making character images who need a specific pose, not "something dynamic"
- Anyone who has a reference photo and wants that exact body position reused on their own character
- Experimenters who want to try H3 Image without building a local stack
The two branches, in plain terms
A. Manual posing (VNCCS). You pose a skeleton by hand in the 3D pose director. The skeleton becomes the conditioning for H3 Image, and the model renders your character in that pose in 6 steps at high resolution. This is the branch you use when the pose exists only in your head.
Workflow page: https://www.runninghub.cn/post/2122731136920145921/?inviteCode=rh-v1111
B. Photo reverse (SAM3D). You feed a photo; SAM3D extracts the 3D pose from it, and that pose drives the same generator. This is the branch for "make my character stand exactly like that reference".
Workflow page: https://www.runninghub.cn/post/2122731130955845634/?inviteCode=rh-v1111
How to use it
- Open the manual-posing workflow on RunningHub (link above). It runs in the browser; no local install.
- Load your character reference. The workflow keeps identity from the reference and takes pose from the skeleton.
- Pose the skeleton, or switch to branch B and give it a photo instead.
- Run. The chain samples in 6 steps and decodes at high resolution.
- Compare against the reference: keep the run where identity and pose both hold; discard the ones where the face drifted.
What is actually inside the JSON we publish
The exported workflow is 58 nodes across 25 node types: two MiniMaxH3ImageConditioning nodes (one per branch), two KSamplers, two VAEDecode/SaveImage paths, eight ComfySwitchNode entries that switch between the manual and reverse branches, MarkdownNote nodes for the in-canvas instructions, and two GetImageSize helpers. Model dependencies, named exactly: minimax_h3_image_pruned_int8_convrot.safetensors, minimax_h3_image_vae_fp16.safetensors, qwen3.5_9b_w4a8_gptq.safetensors, qwen3vl_32b_minimax_h3_int8_convrot.safetensors. No model weights are bundled; the JSON carries wiring and settings only.
Known limits, honestly stated
- Identity is good but not absolute. Extreme poses (heavy occlusion, profile + crossed limbs) still drift. Keep the reference simple and well lit.
- One embedded capture image ships with the JSON as a canvas reference. Treat it as an illustration, not as training data.
- Branch B needs your own character photo; the SAM3D path is about extracting pose, not about giving you a character.
- SAM3D itself is an online experience in this episode. There is no downloadable JSON for the SAM3D photo-reverse workflow here, and we are not going to pretend otherwise: use the RunningHub page.
- 6 steps is fast, not magic. Complex results may need more than one run.
Episode and resources
- Video: https://youtu.be/Ncg4_CFSaaM
- Chinese original: https://www.bilibili.com/video/BV1nSsr67EHA/
- Manual pose workflow (online): https://www.runninghub.cn/post/2122731136920145921/?inviteCode=rh-v1111
- SAM3D photo reverse (online only): https://www.runninghub.cn/post/2122731130955845634/?inviteCode=rh-v1111
If you build something with either branch, the interesting failure cases are the ones worth sharing: which poses break identity first, and how much reference quality matters.
Top comments (0)