msch-a2v
Legacy component. Install MSCH Nodes instead: github.com/mariobilly/msch-comfyui-nodes
Nodes (7)
The node where MiniMax H3 actually makes your music video
Prototype a music video tiny, then render it big — all on your own GPU
The node that turns a song into a beat-snapped shot list
Load your song as audio you can actually sync to
A one-input node that saves you from wasting a GPU render on the wrong cut
Where your sampled shots stop being latents and become the actual video
Where your prompt schedule becomes actual H3 shots — conditioning, references and timing in one place
Superseded by MSCH Nodes: one install for the complete collection.
This standalone repository is retained for historical links and existing installations. New nodes and updates belong to the unified pack. Follow the migration guide before switching to avoid duplicate nodes.
MSCH A2V
Beat-synced prompt sequencing, shot planning, MiniMax H3 sampling, assembly and pixel upscaling for music videos.
Node reference · Example workflows and results · Publishing guide
<!-- MSCH SHOWCASE -->
View the showcase: videos, images and workflows
<!-- /MSCH SHOWCASE -->Included nodes
| Node | What it does | |---|---| | MschA2V Beat Prompt Sequencer | Open the browser sequencer to author beat-snapped prompt blocks and compile the saved schedule. | | MschA2V Load Audio Path | Decode an audio path into ComfyUI AUDIO and report its duration. | | MschA2V Preview Audio | Write a temporary playable audio preview for a connected AUDIO value. | | MschA2V Shot Planner | Convert the prompt schedule into MiniMax H3 shot conditioning and starting latents using the CLIP encoder, video VAE, audio VAE and timeline audio. | | MschA2V Beat KSampler | Sample each planned MiniMax H3 shot using the connected model. | | MschA2V Shot Assembler | Decode sampled per-shot latents with the video VAE and assemble them into an IMAGE sequence using the plan's timing and overlap behavior. | | MschA2V Beat Pixel Upscale KSampler | Run an optional pixel-space refinement pass over sampled shots: decode, resize toward target_long_side, re-encode with the VAE and resample against the shot plan. |
Installation
Clone into ComfyUI/custom_nodes:
git clone https://github.com/mariobilly/msch-a2v.git
Open a terminal in the cloned folder and install requirements using the same Python environment as ComfyUI:
python -m pip install -r requirements.txt
Windows portable, from ComfyUI_windows_portable:
.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\msch-a2v\requirements.txt
Restart ComfyUI and refresh the browser. Load a JSON workflow from examples/ and select the supplied demo input or your own media. Keep only one installed copy of each package to avoid duplicate node registrations.
Requirements and behavior
Requires the separate comfyui-minimax-h3-audio-T8 pack and compatible MiniMax H3 model, text encoder, video VAE and audio VAE. Model files are not bundled. The sequencer creates timing and prompts; the H3 models generate the video. Generation depends on your GPU, model setup and source audio.
Documentation and examples
docs/NODES.md documents every input, default, range, choice and output. examples/README.md explains which inputs and other nodes each workflow needs and how the included results were produced.
ComfyUI Manager
The release includes Comfy Registry metadata and a GitHub publishing action. Version 0.1.0 was uploaded successfully to Comfy Registry under publisher mariobilly on 2026-09-06. Registry reported Pending at verification; Manager installation availability is not yet confirmed. A separate ComfyUI Manager node-list registration has been submitted; listing is pending maintainer acceptance. Git installation works independently. See PUBLISHING.md.
Validation
Imports and input schemas were checked against the local ComfyUI environment with Python 3.12.10, PyTorch 2.10 and CUDA available. Example render coverage is documented per workflow; this is not a claim of compatibility testing on every platform or of every MiniMax H3 configuration. The linked INT4 H3 demo completed full generation on the stated local setup; pixel-upscale refinement remains untested.
License
Project code: MIT. Third-party assets and optional model weights keep their own licenses.