MiniMax H3 Subject Binding (Plan v2)

Most of the time, one reference image means one job: this is the person, this is the scene. But sometimes a single photo is doing triple duty - a frame that establishes the main character's face, her coat, and the lighting style all at once. H3 lets you split that into separate Subjects, but writing it by hand means carefully pointing each <Subject N> at the same picture without double-registering the media. Subject Binding (Plan v2) is the advanced node that handles exactly this.

Its one rule, stated in the node's own description: it never registers or routes the physical media a second time. The image or video already lives in the plan as a registered reference; this node just attaches additional Subject interpretations to the same media. You can't accidentally double-count against H3's 9-image / 3-video / 12-file limits by adding bindings, because the media isn't being re-added - only the roles are.

How it wires in

You start from an Image Reference or Video Reference that's already in the chain, and connect its reference_handle output. That handle is the pack's way of saying "this binding points at that media." Then you define the new Subject:

• subject_name - the human alias this binding creates (e.g. "the red coat").
• content_type - what this binding actually defines from the media: identity/appearance, object/prop, scene/environment, visual style, or pose/expression/action/motion.
• retention - one explicit relationship: fully preserved, partially preserved, attribute transfer, or weak reference.
• shot_scope - optional numeric scope; no shot prose belongs in this field.
• notes - visible facts that belong to this role only, so the coat gets coat facts and the face gets face facts.
• transfer_target_subject - required only when retention is attribute_transfer: the upstream Subject that receives the attribute.

The reference_handle output passes through, so you can chain another binding onto the same physical media.

The two real use cases

The pack's own docs frame it as: combine multiple references as evidence for one Subject, or split one media asset into multiple Subjects/roles. Both directions work through the same node, which is why this is the "advanced" one - beginners usually get by fine with one Image Reference per role and never touch it.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart, look under MiniMax H3 → Plan v2. No extra dependencies.

Gotchas

• You need the handle, not the plan alone. Connecting h3_plan but skipping reference_handle means the node can't tell which media to bind to, and you get a validation error. The handle is the whole point.
• Don't use it to register new media. If you want a new image, add another Image Reference. Binding an unregistered asset isn't what this node does, and the "never routes media a second time" guarantee only makes sense if the media already has a home.
• Attribute transfer needs an existing target. attribute_transfer with an empty transfer_target_subject is rejected; the target must already exist upstream in the chain.

Where this shines: a character-replacement workflow where one reference image defines both the replacement's face and their wardrobe, or a style-transfer shot where a single keyframe supplies both composition and a motion Subject. It's the node you reach for when your graph starts feeling like it's re-loading the same file just to say something different about it.