MiniMax H3 Fastest Workflow | 6GB VRAM 16GB RAM — v3.0 — v6.0 (Extreme Fast) 💀

Use this ComfyUI MiniMax H3 workflow to generate video from prompt text and a reference image, transform existing video, or upscale video; it accepts text, image, or video input and saves video output.

The graph includes reference-to-video and RTX video super-resolution nodes, along with audio decoding and video saving.

The workflow includes image, audio, and multiline text input controls, plus a video-saving output.

The release notes recommend using the latest versions of the required models and nodes.

Listed model files: minimax_h3_audio_vae_fp32.safetensors, minimax_h3_ref2va_pruned_int8_convrot.safetensors, minimax_h3_video_vae_fp16.safetensors, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, and taeh3.safetensors.

Listed node packs: comfyui_nvidia_rtx_nodes, comfyui-kjnodes, comfyui-spectrum-minimax-h3, and plaguekind-nodes.

For a speed-focused choice, the notes suggest 4 steps; they describe 6 steps as a balance of quality and consistency, and 8 steps as the highest-quality option with stronger face consistency.

Suggestion · not verified

The update notes recommend the Step 600 Lighting LoRA and say Tomato Lighting LoRA can be enabled for stronger lighting in NSFW generations.

The listing includes notes for several revisions: it says v6 runs at 5 steps, while later guidance gives 4-, 6-, and 8-step settings and says Spectrum was removed by default; the v6 graph still lists the Spectrum node pack. Check the step and Spectrum settings in your downloaded

Suggestion · not verified

A note suggests disabling LoRA if you encounter audio issues; check that advice against the version you are using.

Suggestion · not verified