This is a low vram workflow for video generation with the MiniMax H3 model. This model supports audio along with video. A 5s video takes about 5 minutes on a 3060 12G GPU.

REQUIRED MODELS

• minimax_h3_fl2va_pruned_int8_convrot.safetensors (base, int8)

• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder)

• minimax_h3_video_vae_fp16.safetensors (video VAE)

• minimax_h3_audio_vae_fp32.safetensors (audio VAE)

• minimax_h3_turbo_v4_step600_ema.safetensors (Turbo LoRA)

HOW TO USE

Drop your start image into the Load Start Image node and write your prompt. The resolution chain solves H3's canvas formula automatically, so you don't need to work out dimensions for a given aspect ratio.

For audio, add a separate line to the prompt, for example:

Audio: she says clearly, "Six steps. That's all it took to make this."

Underneath, a low arctic wind. No music.

Scheduler is set to simple at 6 steps. Push to 8 if a specific prompt looks soft - speech held up fine at 6 in testing.

NOTES

- Output muxes at 24 fps

- Video and audio decode on separate paths, then combine

- Duration is set in seconds and converted to a frame count automatically

- Two notes at the top of the graph list the models and a short overview

The cover clip was generated with this workflow at 6 steps, from a single still image. The voice in it is the model's, not dubbed.

Full video walkthrough of every node: