MiniMax inference with ComfyUI¶
A streamlined and automated environment for running ComfyUI with MiniMax H3 video and native-audio generation, optimized for RunPod.
What to expect¶
The template provisions the official ComfyUI MiniMax H3 model repacks, VAEs and workflows. Basic familiarity with RunPod pods, logs, secrets and file management is useful; normal use does not require Linux administration experience.
When to use this template¶
Use this template for MiniMax H3 text-to-video, image-to-video, first/last-frame video and reference-to-video workflows with generated audio.
🔧 Features¶
- Automatic model provisioning through environment variables.
- Models downloads depending on VRAM and architecture (Ada Lovelace / Blackwell).
- CUDA 12.8 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.
- Example workflows.
- 4-step , 8-step Turbo-loras (turbo and lightx2v) included
- llama-cpp native and llama-cpp-python with CUDA 12.8 support.
📦 Deployment on RunPod¶
📘 Tutorial¶
Ref2va¶
Enhanced standard workflow Reference to video/audio (Ref2va)¶
Multi-Shot/Motion-Context Workflow with 3 x 10 seconds starting from a reference image¶
fl2va¶
Enhanced standard workflow image to video/audio¶
with or without turbo
Enhanced standard workflow text to video/audio¶
Custom workflows¶
Director , all in one¶
Advanced REF workflow with 2 samplers (warmup) and PDD Acc turbo lora¶
Prompt generator using uncensored heretic Minimax-H2 QWEN-VL with generation tail¶
Prompt generator using Qwen3.8-27B-Uncensored with llama-cpp¶
Fantastic prompt builder & media manager¶
Video & sound preview¶
Ref2va¶
Standard¶
Enhanced¶
Multi-shot continuation¶
fl2va¶
i2v¶
t2v¶
Example video with director¶
Pod running on L40S (good quality)¶
Pod running on RTX 5090 (fast but restricted in resolution and duration)¶
Pod running on RTX PRO 6000 (fast and no restrictions)¶
Pod running on RTX 3090/4090 (slow and restricted in resolution and duration)¶
