ComfyUI-Spectrum-MiniMax-H3
Spectrum-based acceleration for ComfyUI’s native MiniMax H3 audio-video model. Forecasts post-transformer features with Chebyshev ridge regression to skip selected transformer evaluations, with adaptive scheduling, sampler-aware safeguards, CPU/VRAM history storage, and native fallbacks.
What is it?
What it is
A custom ComfyUI node that implements spectral feature forecasting for the MiniMax H3 audio-video model.
Why it exists
To reduce expensive H3 transformer evaluations during sampling, acting as an approximate accelerator to increase throughput.
Who should use it
Users seeking faster MiniMax H3 sampling who can tolerate potential trajectory deviations or localized quality degradation in fast-moving details (e.g., eyes, fingers) and are using supported samplers (Euler, RES multistep, RES multistep CFG++) with compatible ComfyUI versions.
Who should avoid it
Users requiring maximum fidelity to the native MiniMax H3 trajectory, doing quality-critical work, using unsupported samplers or ancestral methods, or working with multi-GPU parallel sampling where forecast validation is lacking.
How it works
A quick walkthrough in plain English
How ComfyUI-Spectrum-MiniMax-H3 works
Step 1 of 3
You interact with it
Open ComfyUI-Spectrum-MiniMax-H3, send a request, or connect it to your stack.
Features
Advantages
- Reduces expensive H3 transformer evaluations during sampling
- Significant potential for end-to-end wall-clock speedup
- No third-party Python dependencies required
- Minimal memory overhead when using system_ram storage
- Predictive scheduling maintains sampler-safe minimums
Disadvantages
- Not a lossless or bit-identical execution path
- Potential for trajectory deviations in fast or brief actions
- Localized quality degradation in rapidly moving small details (e.g., eyes, fingers)
- Increased VRAM usage when using vram history storage
- Requires specific ComfyUI versions and model types
Installation
native
Clone the repository into `ComfyUI/custom_nodes`: ```bash cd ComfyUI/custom_nodes git clone https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3.git ``` Restart ComfyUI.
FAQ
What is the primary purpose of the Spectrum Apply MiniMax H3 node?
The node reduces expensive H3 transformer evaluations during sampling by fitting a Chebyshev ridge model to actual post-transformer hidden features and forecasting those features on selected future solver steps, while still executing native output heads, video/audio reconstruction, sigma mapping, and return structure on every step.
How does enabling Spectrum affect output fidelity compared to native MiniMax H3 sampling?
Spectrum is an approximate accelerator that changes the denoising trajectory, potentially causing trajectory deviations (e.g., different motion, pose, or gaze) and localized quality degradation in fast-moving or briefly visible details like eyes, fingers, or fingernails. Outputs may differ even with identical prompts, seeds, models, samplers, and workflows.
Which samplers are currently supported for forecasting with Spectrum Apply MiniMax H3?
Forecasting is allowlisted for Euler (`sample_euler`), RES multistep (`sample_res_multistep`), and RES multistep CFG++ (`sample_res_multistep_cfg_pp`). Ancestral samplers and multi-GPU parallel sampling remain native due to incompatibility with the deterministic feature trajectory assumption.
What is the difference between `history_storage=system_ram` and `history_storage=vram` in terms of performance and memory usage?
With `system_ram`, history is stored in CPU memory, reducing VRAM usage but adding CPU-GPU transfer overhead. With `vram`, history stays on the GPU, avoiding transfers but requiring sufficient free VRAM (e.g., >6.1 GiB for native 1344x768, 124-frame examples). VRAM history may offer variable timing benefits but is not a guaranteed speedup and risks OOM if headroom is insufficient.
What are the conservative default parameters for Spectrum Apply MiniMax H3, and why are they considered conservative?
The conservative preset uses: `blend_weight=0.50`, `degree=4`, `ridge_lambda=0.10`, `window_size=2.0`, `flex_window=0.75`, `warmup_steps=5`, `tail_actual_steps=1`, `max_history=8`, `history_storage=system_ram`. These settings prioritize stability and fidelity over speed, using moderate blending, lower polynomial degree, and smaller adaptive windows to reduce forecasting errors.
What happens if Spectrum Apply MiniMax H3 encounters an incompatible model or sampler during execution?
The node falls back to the native MiniMax H3 path when forecasting is unsupported or unsafe (e.g., sampler incompatibility, missing branch labels, topology changes, or prediction shape failures). In such cases, the entire `predict_noise` attempt is discarded and rerun as an actual step, preserving scheduler state and releasing history on teardown.
Featured in Videos
YouTube tutorials and walkthroughs for ComfyUI-Spectrum-MiniMax-H3
Alternatives
Similar projects ranked by category, topics, and text overlap.
