Mr_CadillacV8 - MiniMax H3 Character Reference Director — Consistent Characters

MiniMax H3 Character Reference Director — Consistent Characters, Audio References, Start/End Frames & More

After several weeks of testing, rebuilding and refining my MiniMax H3 workflow, I’m finally releasing the first public version of my H3 Character Reference Director Workflow.

The goal was simple:

Make H3 easier to control and significantly improve character consistency across video generations.

Instead of manually wiring and describing every reference again and again, the Character Reference Director manages the references, roles and prompts for you.

It supports character reference images for:

Face identity and different face angles

Facial expressions

Full body proportions

Clothing

Hairstyle

Relative height and body scale

Exact environments and locations

Multi-shot character references

The workflow also supports:

Multiple characters

Character presets / saved character projects

Audio references for voice timbre

Exact dialogue control

Two-person dialogue

Off-screen / POV speakers

Video motion references

Start frame

Optional end frame

Reference-based environments

Target scene presets

Community LoRAs

Spectrum acceleration

ComfyUI Kitchen Attention

Turbo LoRA support

Generated final-frame preview and saving

One of the biggest improvements during development was reference prompting.

H3 often treats weak image descriptions as inspiration and creates a similar character instead of the exact person. The Director now uses stronger reference-role prompts designed to tell H3 exactly what information each image should preserve.

For example, body references can explicitly preserve:

body proportions, height impression, chest proportions, waist, hips, glute proportions, silhouette and relative scale.

The same principle is used for faces, clothing, hairstyles and environments.

I have also spent a lot of time improving dialogue handling. H3 has a tendency to improvise, repeat lines or continue speaking after the written dialogue ends. The included dialogue presets are designed to reduce this behaviour and clearly separate:

voice reference → speaker → exact dialogue → scene audio.

Performance

The supplied workflows also support ComfyUI Kitchen Attention.

On my RTX 3090, switching from standard PyTorch Attention to ComfyUI Kitchen Attention reduced one 5-second test from approximately:

577 seconds → 350 seconds

with no visible reduction in character quality in my tests.

A 10-second generation that previously took close to 12 minutes completed in approximately 8½ minutes.

Two workflows are included

Reference-to-Video WorkflowBest when the character references define the generated shot.

Start / End Frame + Character References WorkflowCombines first-frame / optional last-frame control with the Character Reference Director.

This second workflow has become one of my favourite ways to work with H3 because the start frame can define the scene while the character references preserve identity separately.

Who is this for?

This workflow is aimed at people who already use MiniMax H3 in ComfyUI and want:

better character consistency

faster reference setup

reusable characters

more structured prompting

stronger control over dialogue and audio references

easier multi-character workflows

less repetitive manual work

This is not a magical “one-click perfect video” system. H3 still has its own limitations, and difficult scenes can still require experimentation.

But compared with using H3 references manually, this has made my own workflow much faster, cleaner and more predictable.

The full workflow package, Character Reference Director custom node and setup instructions are available for paid Patreon members.

Join the membership to download the complete release.