Verified setup sequence

Run MiniMax H3 in ComfyUI

Use ComfyUI 0.30.0 or newer, choose one diffusion partition, install the shared text encoder and two VAEs, then validate a baseline before adding Turbo, Sage or cache nodes.

Evidence checked August 11, 2026

Baseline setup sequence

Complete these 5 steps before adding optional acceleration.

Update ComfyUI

The reviewed official tutorial requires version 0.30.0+. Record the exact commit, PyTorch and CUDA/ROCm versions when troubleshooting.

Choose one partition

Use FL2VA for T2V, I2V and first/last-frame work. Use Ref2VA for mixed reference images, video or audio.

Install the four-file set

One diffusion checkpoint plus the Qwen3-VL encoder, video VAE and audio VAE totals about 42.47 GB decimal.

Load an official template

Confirm that every model loader resolves. Generate a short baseline and validate both the visual stream and stereo audio.

Add one accelerator at a time

Turbo and Sage are optional. EasyCache is not defaulted here because an open report describes audio degradation.

The exact core files

The diffusion checkpoint changes by partition; the remaining three files are shared.

minimax_h3_fl2va_pruned_int8_convrot.safetensorsComfyUI/models/diffusion_models/ · 20.97 GB
minimax_h3_ref2va_pruned_int8_convrot.safetensorsComfyUI/models/diffusion_models/ · 20.97 GB
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsComfyUI/models/text_encoders/ · 15.69 GB
minimax_h3_video_vae_fp16.safetensorsComfyUI/models/vae/ · 5.21 GB
minimax_h3_audio_vae_fp32.safetensorsComfyUI/models/vae/ · 0.61 GB