Wuli-Art's MiniMax H3 Turbo offers fixed-step audio-video checkpoints with a dedicated ComfyUI node pack. Its four-, six- and eight-step options do not all use the same packaging. Choosing the matching loader and checkpoint is the first task; lowering a step widget in an unrelated workflow is not an equivalent installation.
This guide follows the published model and node documentation reviewed October 17, 2026. H3Vid has not benchmarked these checkpoints. The cover is generated concept art about model components, not a Wuli output or installation screenshot.
Understand the three packages
The model repository separates turbo-4step, turbo-6step and turbo-8step. Six and eight steps use complete Diffusers-format transformers. The four-step option supplies a compact adapter plus fused video and audio output projections, while retaining a dependency on the official transformer.
The node implementation's guide explains why a generic LoRA loader is incomplete for the four-step file. Its step-dependent projections also need to be applied. A familiar safetensors extension does not make two adapters interchangeable, and the compact file does not remove the base model's memory requirements.
Treat these folders as execution choices. Keep the selected variant, local path and repository revision in a project note. If you switch from eight to four steps while pointing at a local eight-step directory, update that path too. Otherwise the visible step count and actual model source can disagree.
Install in the right Python environment
The node pack expects a local ComfyUI installation with CUDA-enabled PyTorch. Its documented installation is:
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/wuli-art/MiniMax-H3-Turbo-ComfyUI.git
cd MiniMax-H3-Turbo-ComfyUI
python -m pip install -r requirements.txt
Use the interpreter that actually starts ComfyUI. For Windows Portable, that is the embedded Python referenced in the project's instructions, rather than an unrelated system installation. Restart the ComfyUI server after installing dependencies. Refreshing a browser tab alone does not reload Python nodes.
The requirements file is the authority for this node pack's versions. Reinstall it when updating the checkout. Missing PEFT or missing Qwen3-VL support should be resolved as dependency problems before changing a prompt or downloading another model.
Reserve space for both the selected Wuli package and the original H3 components. The first request can spend substantial time downloading and loading. Measure a warm subsequent request separately if you want to understand repeated-use latency.
Start from the supplied graph
Open one of the repository's example workflows. Use the four-step graph for the compact package. The six-step setup starts from the eight-step example and changes the selector to six, with the corresponding model source.
The key connection is MiniMax H3 Turbo Generate to MiniMax H3 Turbo Save MP4, carrying both frames and audio. Start with automatic downloads if you have not prepared an offline model store: the documented default leaves model_root empty and uses MiniMaxAI/MiniMax-H3 as components_root.
For text-to-video, leave first and last image inputs disconnected. For image conditioning, attach a selected Load Image output to first_frame; attach a second image to last_frame only when you want an ending constraint. Make sure the loaded images are yours, rather than leftover filenames from someone else's graph.
A simple illustrative scene is:
A paper airplane glides slowly through an empty sunlit classroom. The camera stays still. Soft paper rustling and quiet room ambience; no speech or music.
This example reduces moving parts for an initial check. It is not a tested recipe or a promise that negative sound instructions always succeed. Listen to the exported track rather than deciding that a silent-looking scene generated silence.
Check the output before optimizing
The supplied workflow defaults include 124 frames and a 1344×768 canvas; export uses 24 fps. That frame count is about 5.17 seconds, so avoid describing it as an exact five-second measurement. The save route creates one H.264/AAC MP4 rather than a separate WAV deliverable.
Verify that the output plays from beginning to end and includes the expected audio stream. If an endpoint image is important, compare the opening and ending frames to the inputs. Inspect motion and sound together: a plausible final still does not establish that the intermediate transition worked.
For comparison, hold the prompt, images, seed, canvas and frame count constant. Change the checkpoint selection deliberately and record loading time separately. The author's interactive comparison provides selected prompts and configurations for inspection, not a guarantee of equivalent quality on your own scene.
Keep names and access routes precise
“Turbo” appears in several H3 products. Wuli-Art's local checkpoints are distinct from fal's hosted H3 Max Turbo and from unrelated LightX2V adapters. H3Vid's Turbo page addresses its online generation route; it is not a claim that this Wuli node pack runs behind that form.
Current primary documentation describes the three fixed-step packages above. Community references to an “F2” fusion should be checked against an actual released file and supported workflow before treating that label as an additional downloadable variant. Do not substitute a similarly named file to fill the gap.
For an offline setup, follow the model-root forms in the node guide and preserve both the selected package and required official components. Review the model's linked license separately from the node code. The useful next step is one complete, saved audio-video render with known inputs; broader speed or quality conclusions require a controlled comparison on your hardware.