ComfyUI + FastH3 Omni: clean public package
A downloadable ComfyUI source app and one FastH3 Omni NVFP4 QAD reference-to-video workflow, extracted from the existing Pixaroma setup. This is the most quantized H3 variant selected in that setup: an NVFP4 denoiser and an NVFP4/AWQ 32B text encoder, with FP16 video and FP32 audio VAEs. It is not a claim to be the smallest H3 model available anywhere.
Requires NVIDIA Blackwell, such as an RTX 5090 or RTX PRO 6000 Blackwell, for this NVFP4 path. Use a Linux x86-64 environment with Python 3.12, a driver supporting CUDA 12.8, and enough RAM, VRAM and disk for these files and runtime dependencies. Smaller non-Blackwell GPUs are not a drop-in target for these exact weights.
Files
comfyui-h3-source.tar.gz: clean ComfyUI app source with the required Pixaroma, KJNodes, rgthree and Spectrum nodes, the workflow, and installation/start scripts. Original upstream licenses are included.FastH3 Omni - Reference to Video (Pixaroma).json: the standalone workflow for an existing compatible ComfyUI installation.models.json: exact public source revisions, byte counts, SHA-256 checksums and destination paths for all four weights.download_models.py: downloads only those four files, verifies them, and installs the workflow. It uses anonymous downloads (token=False); no Hugging Face token is required for these sources.runtime-sources.json: exact GitHub source commits and archive hashes.validation.json: static validation and privacy-scan results for this package.
The app archive is a source distribution. It does not contain a preinstalled CUDA/Python environment or the 35.27 GB model weights. Dependencies are installed on the recipient's machine. It is not a drop-in replacement for a previously qualified private runtime archive.
Install the clean app
Download comfyui-h3-source.tar.gz from this repository and run on Linux x86-64:
tar -xzf comfyui-h3-source.tar.gz
cd comfyui-h3
bash install.sh
.venv/bin/python download_models.py --comfyui ComfyUI --download
bash start.sh
Open http://127.0.0.1:8188. For a remote machine, forward port 8188 over SSH. The start script binds to loopback by default; this archive contains no account login or saved credentials.
The installer uses PyTorch 2.11.0 with CUDA 12.8 and the matching published SageAttention wheel. It installs the requirements of the pinned source packages. A fresh installation still needs an import, node-registration and actual GPU generation check before being treated as qualified.
Use an existing ComfyUI installation
Use the core and custom-node versions in runtime-sources.json. With the requirements installed in that ComfyUI environment, place the standalone workflow, models.json and download_models.py together, then run:
python -m pip install -r requirements-download.txt
python download_models.py --comfyui /path/to/ComfyUI
python download_models.py --comfyui /path/to/ComfyUI --download
python download_models.py --comfyui /path/to/ComfyUI --verify
The command without a mode prints the download plan and makes no changes. An existing different model or workflow is not overwritten. Download concurrency is two and Xet high-performance transfers are enabled.
Exact models: 35.27 GB total
| Model file | Size | Source |
|---|---|---|
fasth3_omniref_qad_nvfp4_comfy.safetensors |
13.77 GB | public download |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
15.69 GB | public download |
minimax_h3_video_vae_fp16.safetensors |
5.21 GB | public download |
minimax_h3_audio_vae_fp32.safetensors |
0.61 GB | public download |
The public sources were checked without a token. Complete weight files were not downloaded again for this publication; the hashes are the publishers' pinned LFS SHA-256 identities.
Workflow controls
Load the workflow and upload your own references into its two image loaders, or disconnect an unused reference input. Reference filenames and media have intentionally been cleared. Edit the generic prompt, preserving <Picture 1> and <Picture 2> tags for the references you use. The graph keeps the eight-step Euler/simple setup, CFG 1, video/audio shifts 12/3, and 24-fps output controls. The default duration is five seconds.
The graph uses the matching PDD-aware ComfyUI core and Spectrum/KJNodes sources. Do not replace the selected files with similarly named BF16, GGUF or other runtime-format weights. Disabled interpolation/upscaling branches were removed from this export; they require no extra model downloads.
Quantization and approximation patches can affect output quality. This publication was statically validated, not newly GPU-rendered. No new rental-to-output timing or quality claim is made, and no GPU was rented to prepare it.
Privacy and scope
The app was assembled exclusively from pinned public source archives. No used installation, environment files, Grok/xAI keys, saved API settings, cookies, access tokens, user database, logs, training data, reference media or generated outputs were copied. The workflow was separately sanitized. Credential-pattern and archive-content checks are recorded in validation.json.
This publication does not change the live AutoScale installation or add H3 to the current SleevrApi launch allowlist. Source and model terms remain as described in NOTICE.md.
- Downloads last month
- -