DEV Community

AI-KSK
AI-KSK

Posted on

MiniMax H3 Image: Two Pose Directors (Manual VNCCS + SAM3D Photo Reverse) - ComfyUI Workflow Guide

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

  1. Open the manual-posing workflow on RunningHub (link above). It runs in the browser; no local install.
  2. Load your character reference. The workflow keeps identity from the reference and takes pose from the skeleton.
  3. Pose the skeleton, or switch to branch B and give it a photo instead.
  4. Run. The chain samples in 6 steps and decodes at high resolution.
  5. 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

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)