Generate video with native audio from a written scene, a starting image or reference media. MiniMax H3's reference mode accepts images, video and audio to guide the result.
ComfyUI and JupyterLab with H3 model downloads
Text-to-video, image-to-video and reference-to-video graphs
Manual prompts or OpenRouter-assisted Auto Prompt variants
11 workflows included by default
See the workflow breakdown
MiniMax H3 6 workflows · Default
3 manual-prompt graphs plus 3 Auto Prompt variants.
Latent upscaling 5 workflows · Default
Image-to-video and text-to-video with Auto Prompt or custom prompts, plus reference-to-video with Auto Prompt. All eleven install together.
Before you deploy
MiniMax H3 is already enabled in this template. No model-selection change is needed. In Edit Template, open Environment Variables to adjust precision or add optional download tokens before deploying.
Model selection
download_minimax_h3Default: true
MiniMax H3 is already enabled. Leave this value unchanged to download its models and workflows.
minimax_quantDefault: int8
Choose int8, fp8 or nvfp4; use false for full bf16. Leave int8 unchanged unless you need another precision. Full bf16 needs more storage.
CivitAI downloads
civitai_tokenDefault: empty
Your CivitAI API token. Add it if you want to download your own LoRAs or checkpoints.
CIVITAI_LORASDefault: empty
Comma-separated LoRA model-version IDs to download into models/loras. Replace any replace_with_ids placeholder with your IDs.
CIVITAI_CHECKPOINTSDefault: empty
Comma-separated checkpoint model-version IDs to download into models/checkpoints. Leave empty if you do not need any.
Optional access
LLM_KEYDefault: empty
Your OpenRouter API key for Auto Prompt workflows. Manual-prompt workflows do not need it.
HF_TOKENDefault: empty
Optional Hugging Face read token for authenticated model downloads.
