you need🎬 H3 Character RefMod Workflow to use this refmod
✨ What is a Character RefMod?
Think of it as a reusable character reference pack.
Instead of selecting the original reference images every time, you load a character’s .safetensors RefMod file and use it in your video workflow.
Creating a RefMod does not require training the H3 model’s weights. The file stores reference information that H3 can reuse. See the RefMod node documentation.
🏷️ Character Reference Prompt
All my character RefMods use this standard reference phrase:
<Subject 1> is the woman from <Video 1><Video X> matches the RefMod loading slot. Use <Video 1> for slot 1, <Video 2> for slot 2, and so on.
Then describe the scene using <Subject 1>:
<Subject 1> is the woman from <Video 1>.
<Subject 1> walks through a rainy city street at night, wearing a black coat.Want to reference only her face? Say so in the prompt, just as you would with normal reference mode:
<Subject 1> is the woman from <Video 1>.
Use <Video 1> only as a reference for her facial identity.
<Subject 1> wears a white spacesuit inside a futuristic spacecraft.This lets you request a different outfit, hairstyle or setting while using the character’s face as the reference.
🎥 What is it useful for?
Reusing a character across different scenes.
Making videos with new actions, camera angles and settings.
Switching characters by selecting another RefMod file.
Building your own library of reusable characters.
What results should you expect?
The aim is to keep the character recognizable across videos. Check the example videos to see the results. Appearance can still vary with the prompt, movement and camera angle; clothing and other details in the references can also influence the output.
🧩 Install the Required Nodes
You need an updated ComfyUI with MiniMax H3 support and the models used by this workflow.
Start here:
Drag the workflow
.jsoninto ComfyUI.Open ComfyUI Manager → Install Missing Custom Nodes.
Install the missing nodes, then restart ComfyUI.
This installation option is documented by ComfyUI Manager.
If the RefMod nodes are still missing:
Open the ComfyUI-MiniMaxH3Mod repository, click Code → Download ZIP, and extract it into:
ComfyUI/custom_nodes/ComfyUI-MiniMaxH3Mod/Make sure __init__.py is directly inside that folder. Follow the repository’s dependency installation instructions, then restart ComfyUI.
📁 Where Do Character Files Go?
Place your character RefMod .safetensors files here:
ComfyUI/
└── models/
└── refmods/
├── Character_A.safetensors
└── Character_B.safetensorsCreate the refmods folder if it does not exist. This is the standard location used by the RefMod nodes.
or
ComfyUI\custom_nodes\ComfyUI-MiniMaxH3Mod\mods❤️ Credits & Character Packs
RefMod nodes: Luisacaotica / ComfyUI-MiniMaxH3Mod.
Description
FAQ
Comments (3)
Damn this is great, thanks for all this and the explanations!
Could you share the prompt you used for the Harley Ark 2 example?
subject_definitions:
<Subject 1> is the adult woman from <Video 1>, including her AAA game-model identity, facial mesh, skin texture, hairstyle, high-poly body, and the exact original outfit worn in that video.
summary:
[reference generation] The target video generates an 8-second ultra-high-precision AAA game cinematic dance guided by <Subject 1> from <Video 1>, with professional studio lighting, photorealistic MetaHuman skin shaders, and a locked thigh-level framing. She starts in her original outfit and that outfit is removed piece by piece every 2 seconds on the beat until she is fully nude, with her pubic hair and vulva clearly visible in the final beat.
retention_analysis:
<Subject 1> (appears in [Shot 1]): partially_preserved - AAA facial geometry, micro-expressions, subsurface skin, hairstyle, and high-poly body stay consistent from <Video 1>, and the original outfit is present at the start, then removed piece by piece.
detailed_description:
The target video is in an ultra-crisp, photorealistic 3D AAA game cinematic style, Unreal Engine 5 render aesthetic, set in a high-end dark virtual photo studio with professional three-point studio lighting, sharp rim lights, and subtle volumetric haze. The frame is a locked, stable thigh-level medium shot from the top of the head down to mid-thigh, with the pelvis kept inside the frame. Both arms stay connected, both hands stay attached with all fingers visible, and no garment passes over the hands.
[Shot 1] <Subject 1> performs a seductive rhythmic body-isolation dance, facing the lens. Her face, hair, body, and complete original outfit match <Video 1>, rendered with high-poly topology, micro-grained material, and ray-traced cloth. From 00:00.000 to 00:02.000 she dances in that full original outfit, hands resting on her collarbones, hips swaying. At 00:02.000, on a heavy bass drop, she shrugs and the outermost original piece slides off her shoulders and drops behind her. The next original layer remains, with realistic specular highlights and cloth folds, as she rolls her shoulders. At 00:04.000, on the next snare hit, she pinches the center-front clasp and returns her hands to her upper chest. That layer springs open, falls off both shoulders, and lands behind her. Only the last original waist garment remains. At 00:06.000, on the final beat drop, the waist closure releases as she pops her hips and sets her feet apart. The last original garment drops straight down past her thighs. She is fully nude in the same locked frame. A natural patch of dark pubic hair is clearly visible on her pubic mound, and her vulva is unobstructed and readable between her slightly parted thighs, with the same high-poly skin detail as the rest of her body. Her hands stay on her waist and do not cover it. She finishes with a sharp hip lock and holds an alluring crystal-clear gaze into the lens through 00:08.000. Her facial mesh, skin, hair, and body still match <Video 1>. The removed original clothes lie on the floor, separate from her body.
overall_soundscape:
Heavy low-frequency sub-bass drops, razor-sharp electronic snare snaps aligned with the 2-second clothing drops, soft high-detail fabric swishes, and gentle controlled exhalations.
non_diegetic_music:
A high-energy, sensual 128-BPM modern cyberpunk club track with hard beat drops every 2 seconds, punchy percussion, and an intense driving synth bassline.