MiniMax H3 All-in-One Reference Automatic Prompt Workflow

Instructions

1. This workflow is modified based on the official MiniMax H3 All-in-One Reference workflow, with added automatic prompt nodes and prompt templates.

2. The automatic prompt node can only analyze a single image. Therefore, when using multi-image reference, it is recommended to use the prompt template on web platforms such as GPT and Gemini, or with agents like Codex and Claude.

   (Language model download link: https://huggingface.co/AVert888/Qwen3.5-9B-mmproj-q8_0-GGUF)

3. Resolution is adjusted by modifying numerical values, with detailed explanations provided within the workflow.

4. H3 model download link: https://huggingface.co/Comfy-Org/MiniMax-H3

   Available model variants and VRAM requirements:

   • minimax_h3_ref2va_pruned_int8_convrot.safetensors: 12GB VRAM

   • minimax_h3_ref2va_int8_convrot.safetensors: 16GB VRAM

   • minimax_h3_ref2va_bf16.safetensors: 16GB+ VRAM

   Please select according to your hardware configuration.

   Runtime reference (based on RTX 5060 Ti 16GB + 32GB RAM):

   Generating a 10-second 480p video with minimax_h3_ref2va_int8_convrot.safetensors takes approximately 950 seconds.

5. It is recommended to try the free online version first. If you are satisfied with the results, you may deploy it locally.

   Free online workflow: https://www.runninghub.ai/post/2079133532412309506/?inviteCode=rh-v1443