MiniMax H3 (Comfy) / referenceToVideo
Characters stay, scenes change
- images
- diffusion_model
- loras
- api_config
- video
- draft_cache_audio
- draft_cache_video
- workflow_id
- raw_json
This is the MiniMax H3 node you reach for when you have a character, a style, or a subject you refuse to lose between shots. referenceToVideo takes one or more images as required input, holds onto them as the identity anchors, and generates a video that keeps them consistent while the scene moves. It's the character-consistency move in the H3 (Comfy) menu - and it's the simplest of the three H3 nodes to use, because the references are a plain required IMAGE socket rather than the cloud variant's JSON fields.
Part of civitai-comfy-nodes, Civitai's official ~160-node pack generated from its Orchestration API spec. Cloud job: submit a videoGen, long-poll, get a native VIDEO tensor back. Billed in Buzz, nothing local.
Inputs that matter
- images (required, IMAGE socket) - the reference(s). A clean character sheet or a single strong subject still beats three cluttered shots.
- prompt (required) - what happens to that subject.
- duration - default 5.
- width (864) / height (480) - Comfy-scale defaults; small and fast, not 2K.
- seed, steps (20) - standard.
- fast - EasyCache on; speed over fidelity for iteration.
- turbo - the distilled fast path.
- diffusion_model (CIVITAI_AIR) / loras (CIVITAI_LORAS) - checkpoint override and community LoRA stack via the Civitai selectors.
Outputs: video (VIDEO), draft_cache_audio (AUDIO), draft_cache_video (VIDEO), workflow_id and raw_json (STRING).
Install and auth
Standard for the pack - official, Comfy Registry:
# ComfyUI Manager: search "Civitai Comfy Nodes"
cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt
Restart, authenticate via a Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth.
Common issues
Reference quality is the whole game. H3 is strong at holding identity, but only as good as the reference you feed it - inconsistent angles, busy backgrounds, or multiple subjects in one image split the model's attention. One clean reference of the thing you care about beats a collage. fast/turbo blur identity on hard cases - they're for draft iteration; promote to full quality before judging consistency. Cost accumulates across shots: this node is made for multi-shot workflows, and per-output-second billing means a 5-shot scene is five jobs. Your references leave your machine - don't feed it anything you wouldn't send to MiniMax directly. And the pack is early preview, so behavior can change without notice.
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| prompt | STRING | — | |
| images | IMAGE | — | |
| seedopt | INT | 00–4294967295 | — |
| durationopt | INT | 54–15 | — |
| widthopt | INT | 86432–1344 | — |
| heightopt | INT | 48032–1344 | — |
| stepsopt | INT | 201–60 | — |
| fastopt | BOOLEAN | false | Enables EasyCache with its default configuration. |
| diffusion_modelopt | CIVITAI_AIR | — | |
| lorasopt | CIVITAI_LORAS | — | |
| turboopt | BOOLEAN | false | — |
| api_configopt | CIVITAI_CONFIG | Optional Civitai Auth connection; defaults to CIVITAI_API_TOKEN or stored OAuth login. |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| video | VIDEO | — |
| draft_cache_audio | AUDIO | — |
| draft_cache_video | VIDEO | — |
| workflow_id | STRING | — |
| raw_json | STRING | — |