MiniMax H3 ComfyUI workflow templates

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    │   └── minimax_h3_ref2va_pruned_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors

The Comfy-Org quantized four-file set has a static download and storage footprint of approximately 42.47GB in decimal units. Keeping the Ref2VA checkpoint as well adds approximately 20.97GB. These figures describe storage footprint, not minimum VRAM or system RAM requirements.

Set the CLIP Loader type to minimax. The workflow requires ComfyUI 0.30.0 or newer.

Sources: ComfyUI MiniMax H3 tutorial and Comfy-Org MiniMax-H3 repository. Checked 2026-08-13.

• Choose T2V when you only have a prompt.
• Choose I2V when one image needs to control the composition or subject.
• Choose R2V when you need a reference person, action, shot or sound.
• Choose the api_ series when you do not have a 12GB+ graphics card.

The R2V template ran at 1344×768 for 124 frames on an RTX 3060 12GB in 2,581.8 seconds, or approximately 43 minutes. Peak VRAM was 11,649 MiB, 94.8% of 12,288 MiB, and peak system memory was 43,587 MiB.

One machine, one configuration, with audio and every acceleration option disabled. The full test card lists every version, hash and setting behind these figures. The system memory page shows why that RAM number barely moves when the workload does.