Native Qt desktop app for building MiniMax H3 reference packages with ComfyUI
Python
1
3 commits
updated Oct 6, 2026
A native PySide6/Qt desktop app for creating one MiniMax H3 reference package per character, asset, or location.
Requires Linux, Python 3.11+, and FFmpeg on PATH. Install the desktop app in its own environment:
git clone https://github.com/cyberworm1/refmodBuilder.git
cd refmodBuilder
python3.11 -m venv .venv
.venv/bin/pip install -e .
./launch.sh
The app has its own Python environment. It uses the existing local ComfyUI server for encoding, and works offline for reference editing and project browsing.
Project saves are automatic after edits, with an explicit Save Project button. Removing a reference removes it from the package; the source copy is retained in the project folder. Each build preserves a workflow snapshot. The project library is under ~/.local/share/refmodBuilder/projects/. Settings are under ~/.local/share/refmodBuilder/settings.json.
Exports default to ~/ComfyUI/models/refmods/<type>/. Settings can change the local ComfyUI URL, ComfyUI folder, export folder, and VAE filenames. The URL and folder must refer to the same local installation.
The paths below use ~/ComfyUI as a generic example. Set your actual installation and export folders in Settings.
Use a local ComfyUI installation with native MiniMax H3 support, the H3 video/audio VAEs, and MiniMaxH3Mod installed under:
~/ComfyUI/custom_nodes/ComfyUI-MiniMaxH3Mod
Tested MiniMaxH3Mod revision: f9462081e28794389b5a6c5067eb327412ad8ee7.
Install the extension's requirements in ComfyUI's own environment and provide OpenCV or imageio for its video loader. New custom nodes require a ComfyUI restart; arrange this when existing jobs have finished. The app checks node availability before submitting work. ComfyUI defaults to http://127.0.0.1:8188; configure the URL and local folders in Settings.
The app never restarts ComfyUI, interrupts jobs, clears its queue, or unloads models. A build is appended to the normal queue. Stop monitoring stops only this app's monitoring; ComfyUI continues its job. Reopen the project and choose Resume build to collect the result. A saved prompt ID prevents automatic duplicate submissions. If a network failure leaves submission outcome unknown, inspect ComfyUI history before retrying; this version deliberately blocks automatic resubmission.
Each source reference is prepared in a unique ComfyUI/input/refmodBuilder/<build-id>/ folder. Images are orientation-corrected and cropped. Trimmed videos are encoded as silent H.264 at 24 fps; selected audio is prepared separately as 32 kHz stereo WAV. Originals remain intact.
Each member is encoded separately through the upstream RefMod nodes. Intermediate members are saved by ComfyUI under its registered RefMod root, in refmodBuilder_work/<build-id>/. After the app sees its own completed history record, it copies the member tensor bytes into one v5 safetensors bundle. BF16 data is preserved without loading a second model or converting tensors. Independent audio clips remain independent members.
The combined token limit is checked before publishing the final bundle. Overflow reports an error instead of silently dropping references. Full mode and compressed mode use upstream H3 behavior; category and description are metadata, not training instructions. Compressed mode uses a 16×16 grid with 100 refinement steps. Generation quality and voice identity are not guaranteed.
Unique export names and atomic publication prevent overwriting previous packages. Intermediate files are retained for inspection/recovery; automatic cleanup is not included in this version.
python3.11 -m venv .venv
.venv/bin/pip install -e . pytest numpy safetensors
QT_QPA_PLATFORM=offscreen .venv/bin/python -m pytest -q
Tests cover source preservation, save/reopen, ordering/removal, mixed image/audio bundle interoperability with safetensors, BF16 byte preservation, truncated input rejection, token-limit failures, export overwrite prevention, image preparation and audio trimming via FFmpeg, independent member workflow construction, queue append behavior, completion collection, missing-node handling, and ambiguous-submission protection.
Local media, RefMod packages, generation workflows, runtime records, and review screenshots are excluded from this repository. Package-creation workflows are constructed by refmod_builder/backend.py; no external workflow template is required.
Upstream: https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod
See the maintainer handoff for architecture, branch and release status, operational constraints, validation, and future work. Coding agents should also read AGENTS.md.
Native Qt desktop app for building MiniMax H3 reference packages with ComfyUI
Python
1
3 commits
updated Oct 6, 2026
A native PySide6/Qt desktop app for creating one MiniMax H3 reference package per character, asset, or location.
Requires Linux, Python 3.11+, and FFmpeg on PATH. Install the desktop app in its own environment:
git clone https://github.com/cyberworm1/refmodBuilder.git
cd refmodBuilder
python3.11 -m venv .venv
.venv/bin/pip install -e .
./launch.sh
The app has its own Python environment. It uses the existing local ComfyUI server for encoding, and works offline for reference editing and project browsing.
Project saves are automatic after edits, with an explicit Save Project button. Removing a reference removes it from the package; the source copy is retained in the project folder. Each build preserves a workflow snapshot. The project library is under ~/.local/share/refmodBuilder/projects/. Settings are under ~/.local/share/refmodBuilder/settings.json.
Exports default to ~/ComfyUI/models/refmods/<type>/. Settings can change the local ComfyUI URL, ComfyUI folder, export folder, and VAE filenames. The URL and folder must refer to the same local installation.
The paths below use ~/ComfyUI as a generic example. Set your actual installation and export folders in Settings.
Use a local ComfyUI installation with native MiniMax H3 support, the H3 video/audio VAEs, and MiniMaxH3Mod installed under:
~/ComfyUI/custom_nodes/ComfyUI-MiniMaxH3Mod
Tested MiniMaxH3Mod revision: f9462081e28794389b5a6c5067eb327412ad8ee7.
Install the extension's requirements in ComfyUI's own environment and provide OpenCV or imageio for its video loader. New custom nodes require a ComfyUI restart; arrange this when existing jobs have finished. The app checks node availability before submitting work. ComfyUI defaults to http://127.0.0.1:8188; configure the URL and local folders in Settings.
The app never restarts ComfyUI, interrupts jobs, clears its queue, or unloads models. A build is appended to the normal queue. Stop monitoring stops only this app's monitoring; ComfyUI continues its job. Reopen the project and choose Resume build to collect the result. A saved prompt ID prevents automatic duplicate submissions. If a network failure leaves submission outcome unknown, inspect ComfyUI history before retrying; this version deliberately blocks automatic resubmission.
Each source reference is prepared in a unique ComfyUI/input/refmodBuilder/<build-id>/ folder. Images are orientation-corrected and cropped. Trimmed videos are encoded as silent H.264 at 24 fps; selected audio is prepared separately as 32 kHz stereo WAV. Originals remain intact.
Each member is encoded separately through the upstream RefMod nodes. Intermediate members are saved by ComfyUI under its registered RefMod root, in refmodBuilder_work/<build-id>/. After the app sees its own completed history record, it copies the member tensor bytes into one v5 safetensors bundle. BF16 data is preserved without loading a second model or converting tensors. Independent audio clips remain independent members.
The combined token limit is checked before publishing the final bundle. Overflow reports an error instead of silently dropping references. Full mode and compressed mode use upstream H3 behavior; category and description are metadata, not training instructions. Compressed mode uses a 16×16 grid with 100 refinement steps. Generation quality and voice identity are not guaranteed.
Unique export names and atomic publication prevent overwriting previous packages. Intermediate files are retained for inspection/recovery; automatic cleanup is not included in this version.
python3.11 -m venv .venv
.venv/bin/pip install -e . pytest numpy safetensors
QT_QPA_PLATFORM=offscreen .venv/bin/python -m pytest -q
Tests cover source preservation, save/reopen, ordering/removal, mixed image/audio bundle interoperability with safetensors, BF16 byte preservation, truncated input rejection, token-limit failures, export overwrite prevention, image preparation and audio trimming via FFmpeg, independent member workflow construction, queue append behavior, completion collection, missing-node handling, and ambiguous-submission protection.
Local media, RefMod packages, generation workflows, runtime records, and review screenshots are excluded from this repository. Package-creation workflows are constructed by refmod_builder/backend.py; no external workflow template is required.
Upstream: https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod
See the maintainer handoff for architecture, branch and release status, operational constraints, validation, and future work. Coding agents should also read AGENTS.md.