MiniMax inference with ComfyUI

Run MiniMax H3 video and native-audio generation in ComfyUI on RunPod. The template provisions the models, VAEs, custom nodes and example workflows needed for text-to-video, image-to-video, first/last-frame video and reference-to-video generation.

What to expect

These templates are intended for users who want to run ComfyUI on RunPod and are comfortable following technical setup steps. No Linux expertise is required for normal use, but basic familiarity with RunPod pods, logs, tokens and file management is helpful.

The pod is an experimental workspace for video creators who want to combine generation models, editing models, LoRAs and custom nodes in open and flexible creative pipelines. It supports video generation, identity-aware video generation, prompt enhancement and long video generation.

The approach is aligned with the open-model discussion in Eric Hartford's Uncensored Models: local and open AI systems give advanced users more ownership, control and composability. The pod is a technical workspace for responsible use rather than a content platform or safety policy.

Included features

• Automatic model provisioning through environment variables.
• Model downloads selected for the available VRAM and GPU architecture (Ada Lovelace or Blackwell).
• GPU-accelerated video encoding with FFmpeg and NVENC to speed up video creation on slower vCPUs in available on host (auto check).
• CUDA 13.0 runtime with compiled attention acceleration.
• Authentication for ComfyUI, Code Server, Hugging Face and CivitAI.
• Uncensored Heretic Qwen-VL text encoders for inference and prompt enhancement.
• LoRA Manager, installed custom nodes and accelerators.
• Ready-to-use example workflows.
• 8-step turbo LoRAs, including LightX2V and Hyperflow.
• Native llama.cpp and CUDA-enabled llama-cpp-python support.
• Upload/download omichar and Refmod files.

RunPod deployment

Start here: Deploy and start the MiniMax template on RunPod.

Once the pod is running, use the workflow examples below to choose what you want to create. Follow the MiniMax tutorial for the steps from opening ComfyUI to running your first workflow.

Choose a workflow

Start with the generation task, then choose a standard, Turbo, LightX2V, PDD Acc or Hyperflow variant. The JSON filenames below identify the workflows in pod/minimax; the guide links lead to the examples further down this page. Availability on your pod depends on the selected template and provisioned models.

Video generation

Goal and input	Workflow JSON	Guide
Generate video from a text prompt	[Standard T2V] t2v-pod.json, [Turbo T2V] turbo-t2v-pod.json, [LightX2V T2V] lightx2v-t2v-pod.json, [Hyperflow T2V] Hyperflow-t2v-pod.json	Text-to-video
Animate an image or use first/last frames with a prompt	[Standard FL2V] fl2v-pod.json, [Turbo FL2V] turbo-fl2v-pod.json, [LightX2V FL2V] lightx2v-fl2v-pod.json, [Hyperflow FL2V] Hyperflow-fl2v-pod.json	Image-to-video
Guide generation with reference images, video or audio	[Standard R2V] r2v-pod.json, [LightX2V R2V] lightx2v-r2v-pod.json, [Hyperflow R2V] Hyperflow-r2v-pod.json	Reference-to-video
Continue a reference-guided scene across several shots	[LightX2V multi-shot] lightx2v-r2v-multi-shot.json	Multi-shot
Generate connected shots with context pinning	[LightX2V context pin] lightx2v-r2v-multi-shot-context-pin.json, [Hyperflow context pin] Hyperflow-r2v-multi-shot-context-pin.json	Multi-shot
Manage generation stages in one graph	[Director] director-pod.json	Director

Prompt generation and assisted workflows

Choose Qwen-VL for local prompt generation and enhancement. The integrated variants combine it with video generation; the prompt-builder variants provide tools for composing prompts and organizing reference media.

Output examples

• The clips below demonstrate output from the standard and advanced workflows available on the pod.
• Video and audio are generated directly by the model.
• Person reference images were generated using Krea-2 or Qwen Image 2.1.
• Animation reference characters (Elsa, Anna, Ken, Barbie) used for workflow demonstration purpose only.

Ref2va examples

Standard reference-to-video

Enhanced reference-to-video

Hyperflow reference-to-video

Multi-shot continuation

Director motion continuation

fl2va examples

Image-to-video

Text-to-video

Director example

ref2va + QWEN VL prompt enhancer

• Original prompt: "Person from image 1 kisses person of image 2 on a beach while carressing each other. Imtimate scene full of sensuality in a provoking pose".
• Final prompt

Standard workflows

Reference-to-video (Ref2va)

Use the enhanced Ref2va workflow to generate video and audio from a reference image while preserving the subject or visual direction.

View the enhanced Ref2va workflow

Multi-shot reference-to-video

Use the LightX2V multi-shot workflow for connected shots, or choose a LightX2V or PDD Acc context-pin variant. The example below shows three connected 10-second shots from one reference image; actual duration depends on workflow settings and hardware.

View the multi-shot workflow

Image-to-video (i2v)

Use the enhanced fl2va image-to-video workflow with or without a turbo LoRA.

View the image-to-video workflow

Text-to-video (t2v)

Use the enhanced fl2va text-to-video workflow to generate video and native audio from a prompt.

View the text-to-video workflow

Advanced workflows and tools

Director: all-in-one

Use Director to manage generation stages in a single workflow.

2 shot workflow

View the Director workflow shots

Motion continuity workflow

View the Director workflow continuity

Prompt tools

Choose a prompt workflow based on the model and interface you want to use.

Tool	Best suited for
Qwen3.8-27B-Uncensored with llama.cpp	Local prompt generation and enhancement
Fantastic prompt builder and media manager	Building prompts while organizing reference media

View the optional Heretic MiniMax-H2 Qwen-VL generation-tail workflow

[Workflow JSON] `prompt-generator-tail-pod.json`

View the Qwen3.8-27B prompt enhancer

View the fantastic prompt builder and media manager

Video and sound preview

Preview generated video and audio inside the workflow before exporting the result.

View the video and sound preview workflow

GPU accelerated video encoding with nvenc

10 times faster encoding when runpod host supports option, successful tested on RTX 6000 PRO, L40S, RTX 6000 MiG and RTX 6000 Ada.

Custom node in workflow

MiniMax H3: Omnichar vs. RefMod

Both ComfyUI extensions provide reusable references to improve character consistency in MiniMax H3. However, they differ significantly in how references are prepared, stored, and processed.

The main difference: Omnichar offers a simpler, character-oriented workflow that can be assisted by a VLM (Vision Language Model), while RefMod provides a more advanced approach based on pre-encoded latent references, with optional LLM (Large Language Model) assistance for prompt generation.

Feature Comparison

Feature	Omnichar	MiniMax H3 RefMod
GitHub	ComfyUI-Omnichar	ComfyUI-MiniMaxH3Mod
Primary purpose	Create reusable characters with consistent identity, appearance, clothing, and voice.	Create reusable, pre-encoded image, video, and audio references for MiniMax H3.
AI assistance	VLM: Can analyze reference images and help generate character descriptions.	LLM: Can generate structured prompts using stored reference mappings.
VLM/LLM required?	No. Character descriptions can also be entered manually.	No. Prompts can be written manually without an LLM.
Preparation	Select face, body, and clothing reference images and provide a character description.	Encode images, video, or audio into latent references using the corresponding H3 VAE.
Workflow	Create character → Save .char → Load character → Generate.	Extract references → Save .safetensors → Load RefMods → Apply conditioning → Generate.
File format	.char	.safetensors
Stored content	Reference images, locked character description, optional voice, and potentially a trained LoRA.	Full or compressed latent image, video, and audio references with metadata.
Reference compression	Stores character reference images; processing is performed by the target workflow.	Supports Full Reference and Compressed Reference modes with configurable token budgets.
Prompt handling	Decode Character prepares the character description and numbered references.	H3 RefMod Text Encode supports numbered references such as <Picture 1>, <Video 1>, and <Audio 1>.
Main nodes	Encode Character, Save Character, Load Character, Decode Character, Character References Split.	Create H3 RefMod, Create H3 RefMod Master, Load H3 RefMods, Apply H3 RefMod, H3 RefMod Text Encode.
Multiple characters	Multiple character files and reference sets can be used within a workflow.	Multiple independent RefMods can be combined with individual reference strengths.
Reference control	Primarily reference selection, ordering, roles, and image count.	Advanced control over extraction, resolution, masks, tokens, strength, copies, and step curves.
Video and motion references	Primarily focused on character consistency rather than storing motion latents.	Explicitly supports extracting and storing temporal video references.
Audio references	Optional voice sample included in the character file for MiniMax H3.	Audio references encoded with the H3 audio VAE; reliable speaker identity transfer is not yet validated.
LoRA support	Can include a trained LoRA and apply it through Apply Character LoRA.	Does not train LoRAs; uses stored latent references instead of trained model weights.
Other AI models	Supports workflows for other architectures, including FLUX.2 and Krea 2.	Specifically designed for MiniMax H3.
Ease of use	Easy: Create, save, select, and reuse a character.	Advanced: Requires more configuration of reference extraction and conditioning.
Main advantage	A single portable character file that can be reused across supported workflows and models.	Fine-grained control over the reference information supplied to MiniMax H3.
Main limitation	Character consistency still depends on reference quality and the generation model.	Compression may discard details, while larger reference token budgets increase processing costs.

Which Method Should You Choose?

Omnichar — Recommended for ease of use

Omnichar is particularly suitable for users who want to create and reuse consistent characters without complex configuration. A VLM can simplify the preparation of character descriptions by analyzing reference images.

The resulting .char file can be reused across different supported generation models and workflows. This makes Omnichar a convenient solution for character-based image and video generation.

RefMod — Recommended for advanced MiniMax H3 workflows

RefMod is designed for users who need greater control over how reference information is encoded and applied.

An LLM can assist in generating prompts that reference stored images, video, and audio using the correct numbering and mapping.

Because reference extraction is performed in advance, the resulting latent files can be reused without repeatedly encoding the original source material.

RefMod also provides extensive controls for reference strengths, compression, token budgets, masks, and conditioning behavior.

Important Technical Distinction

Although Omnichar can be used with a VLM and RefMod with an LLM, neither AI component is mandatory.

The fundamental architectural difference is:

• Omnichar stores reusable character information, including source reference images and descriptions, in a portable .char file.
• RefMod stores pre-extracted MiniMax H3 latent references in .safetensors files, which can be loaded directly into the H3 reference-conditioning pipeline.

RefMod's experimental H3 RefMod Text Encode also supports presenting visual references to the H3 vision/text encoder, so its functionality is not limited to text-only LLM prompting.

References

• ComfyUI-Omnichar — GitHub
• Omnichar Studio — GitHub
• ComfyUI-MiniMaxH3Mod — GitHub

Character creator workflow

Hardware output comparison

These examples illustrate the quality, speed and generation limits of several GPU configurations.

GPU	Characteristic
L40S	Good output quality
RTX 5090	Fast, with resolution and duration restrictions
RTX PRO 6000	Fast, without those restrictions
RTX 3090/4090	Slower, with resolution and duration restrictions

View the hardware comparison screenshots

L40S — good quality, stable

RTX 6000 Ada — good quality acceptable speed

RTX 5090 — fast, with resolution and duration restrictions

RTX PRO 6000 — fast, without restrictions

RTX 3090/4090 — slower, with resolution and duration restrictions

More information

• Environment configuration
• Hardware guidance
• Image setup
• Custom nodes
• Resources
• Updates