The easy way (2.331.0): run the installer for your system. It sets up everything - its own Node.js and Python, ComfyUI, every custom node pack, the Hugging Face models - tuned for your graphics card, and opens the studio when it is done. Plan for an hour or two; almost all of it is downloads (ComfyUI with PyTorch is several GB, the MiniMax H3 models about 70 GB).
The manual route (sections 1-9) is still here for anyone who wants it.
One file (2.616.0): install_all.bat (Windows), install_all.sh (Linux) or install_all.command (macOS) is
enough on its own - put it in the folder you want bitGen Studio in and run it. It installs Git when it is missing
(Windows: winget, else the newest Git for Windows; Linux: apt / dnf / yum / pacman / zypper / apk with sudo; macOS:
Homebrew, else Apple's Command Line Tools) and Node.js 18+ (Windows: winget Node.js LTS, else the LTS .msi;
Linux / macOS: the first start puts Node.js LTS inside the app folder, runtime/node, no sudo) - on Windows also
Python 3.10+ (start.bat needs it). Then it clones this repository into Bitgen-Studio next to it (or uses the
folder it is in, when that is a bitGen Studio folder) and starts the studio: start.bat on Windows,
installer/bootstrap.sh the first time on Linux / macOS (Python 3.12, ffmpeg, the packages, the app window, a launcher,
the setup window) and misc/bitgen-studio.sh after that. BITGEN_NO_START=1 installs without opening it;
BITGEN_REPO clones another URL.
- The installer
- 1. Prerequisites
- 2. Get bitGen Studio
- 3. First launch (start.bat)
- 4. The MiniMax H3 models
- 5. ComfyUI node packs
- 6. The local LLM (screenplays, chat, ✨ Enhance)
- 7. Optional engines
- 8. Updating
- 9. Troubleshooting
| System | File | How to run it |
|---|---|---|
| Windows 10 / 11 | bitGen-Studio-Setup-<version>.exe |
double-click. No administrator rights - it installs for your user into %LOCALAPPDATA%\Programs\bitGenStudio (change the folder on the second page; a big, fast drive is best) |
| Linux (x64 / arm64) | bitgen-studio-setup-<version>-linux.sh |
sh bitgen-studio-setup-<version>-linux.sh (optional --dir /big/drive/bitGenStudio) |
| macOS | bitgen-studio-setup-<version>-macos.command |
double-click, or sh bitgen-studio-setup-<version>-macos.command. If macOS blocks it: right-click -> Open |
Build them from a full release zip with sh scripts/build-installers.sh dist/bitgen-studio-v<version>-full.zip (the Windows one needs makensis,
from apt install nsis or choco install nsis).
What happens
- The app and its runtime. The app is copied, then
installer/bootstrap.ps1(Windows) orinstaller/bootstrap.sh(Linux / macOS) puts everything it runs on INSIDE the app folder - nothing system-wide: Node.js LTS (runtime/node), Python 3.12 through uv (runtime/python), git (MinGit, Windows, only if you have none), ffmpeg (only if you have none), the app window (Electron) and the Python packages. On Windows the Visual C++ runtime PyTorch needs is installed if it is missing (Windows asks once). A Start-menu + desktop shortcut (Linux: the app menu; macOS:~/Applications/bitGen Studio.app) opens the studio straight in its window. - The setup wizard opens on the first start:
- Your hardware - the GPU check (nvidia-smi / the Mac's GPU), VRAM, RAM and what the studio will do with them.
- ComfyUI - Install a new ComfyUI (the folder is yours to choose; it
gets its own Python environment and the right PyTorch for your card:
CUDA 13.0 / 12.8 wheels for RTX 50, 12.8 / 12.6 for RTX 20-40, Metal on a
Mac, CPU otherwise) or Use my existing ComfyUI with a Browse... button
- a portable build, a git install or a folder with a venv. The wizard shows which of the studio's 14 node packs it already has.
- Custom nodes - every pack the studio's workflows use is installed if it is missing (git clone, or GitHub's zip when there is no git), with its requirements, then ComfyUI's own import check says whether each one loads.
- OmniVoice (voices), optionally faster-whisper (lyric timing), and the assistant's engine (llama.cpp).
- Viggle (native) (optional, unticked by default): Viggle-Animate in
bitGen's own engine. The box shows the download size: its Python
environment (~5 GB) plus the model parts your ComfyUI folder does not
already have (up to ~80 GB). Ticked, it installs when the studio first
opens (
components.txt:viggle native=0); unticked, it is not installed (viggle native=2). The swap dialog's Viggle (H3) engine runs on ComfyUI and works without it. - Settings for your hardware - render size and per-run length for your VRAM, the text encoder placement below 14 GB, where the LLM runs (a second card if you have one), speech / transcription on the GPU or CPU, ComfyUI starting with the app.
- Models from Hugging Face. A Hugging Face token box comes first, then
the list - every model the studio's workflows load, with live sizes, those
right for your GPU ticked. Without a token only the open (not gated)
models can be downloaded; with one, gated models whose licence your
account has accepted can be too (FLUX.2 Klein 9B and Lightricks' LTX files are
gated - accept their licences on Hugging Face). Downloads resume
if they are interrupted - run setup again and they continue.
Download with (2.341.0): the Hugging Face CLI (
hf download- faster on big files, resumes by itself; Install hf CLI puts it in the studio's Python) or the built-in downloader. The default uses the CLI when it is installed. - Done. With nothing to report, the studio opens by itself 8 seconds later
- ComfyUI starts with it and you can generate right away. Anything that
needs your attention is listed instead (and in
setup.log).
- ComfyUI starts with it and you can generate right away. Anything that
needs your attention is listed instead (and in
Headless / terminal: node scripts/setup-cli.js --comfy new --models recommended [--hf-token hf_...] [--downloader hf] [--install-hf-cli] does the
same as the wizard (--help for the options; --detect prints the hardware
check). Just some model files into an installed studio:
node scripts/setup-cli.js --only-models --models ltx23_omninft,taeltx23 --downloader hf (--list-models prints the ids).
Models later: Settings → Models → ⤓ Model catalog (2.341.0) is the same
model page inside the studio - token, Hugging Face CLI or built-in, tick,
Download (it keeps going if you close the dialog). The whole setup again:
start the app with H3_SETUP=1, or delete .setup-done in the app folder.
Uninstalling (Windows): Settings -> Apps -> bitGen Studio. It asks whether to keep your projects, ComfyUI, models and settings.
| Install | Why | How |
|---|---|---|
| NVIDIA driver (current) | CUDA for ComfyUI / llama.cpp | nvidia.com/drivers — RTX 50-series cards need a CUDA 13 driver |
| Node.js 18+ | runs the bitGen Studio server | nodejs.org (LTS) |
| Python 3.10+ | runs the workers | python.org — tick Add python.exe to PATH |
| Git | installs ComfyUI and its node packs | git-scm.com or winget install Git.Git |
| ffmpeg | cutting, joining, trimming, previews, the timeline compile | winget install Gyan.FFmpeg, then open a new terminal so it is on PATH |
Check them in a new PowerShell window:
node --version; python --version; git --version; ffmpeg -versionPick a folder on a drive with at least 100 GB free — ComfyUI, its Python environment and the models all live under the app folder by default.
cd D:\
git clone https://github.com/Tim-byte-dev/Bitgen-Studio.git bitgen-studio
cd bitgen-studioNo Git? Use Code → Download ZIP on the GitHub page and unzip it — but you still need Git installed for the ComfyUI step.
Double-click start.bat. Start installs bitGen Studio when it is not installed yet and runs it when it is - it is
also how you open the studio every day. The first time, it:
- checks Node.js and Python (Python 3.10+: if it is not found, start says so - install it from python.org with
Add python.exe to PATH ticked and run
start.batagain), - installs the Node dependencies (
npm install) and the Python ones (pip install -r misc\requirements.txt), - downloads the app window (Electron, ~150 MB, once) — set the
environment variable
H3_NO_SHELL=1if you would rather use your browser, - starts the server and opens the app.
On the first run the setup window opens. It shows what your computer can run and
asks which version to install - Editor (no AI), Editor + small AI (6-12 GB
card), Editor + mid-size AI (12-20 GB) or Everything (20 GB+), with one
recommended for your graphics card. Then ComfyUI: pick Use a ComfyUI I already
have and select its folder, or let it install a new one - choose a folder (default
comfy\ inside the app) and it will
git cloneComfyUI and create its own Python venv with CUDA PyTorch,- write
comfy\ComfyUI\run.bat(what the app uses to launch it), - install ComfyUI-Manager and the packs the app relies on (KJNodes, Spectrum for MiniMax H3, MMH3 Ultimate Upscale).
Keep the black console window open — closing it stops the server.
Already have ComfyUI? Pick Use a ComfyUI I already have in the setup window - or decline it, then in Settings → ComfyUI set ComfyUI folder to your install (the folder with
main.py,run_nvidia_gpu.bator arun.bat) and the ComfyUI URL (defaulthttp://127.0.0.1:8188). Declined by mistake? Deletecomfy\.declinedand restart.Browser mode: run
powershell -ExecutionPolicy Bypass -File misc\get-comfy.ps1yourself, then open http://127.0.0.1:8311.
bitGen Studio renders video through ComfyUI, so the model files go into
ComfyUI's models folder (comfy\ComfyUI\models\ for the built-in
install). They come from the Comfy-Org/MiniMax-H3
repository, whose folders already match ComfyUI's:
| File | Folder | Size | Used for |
|---|---|---|---|
minimax_h3_ref2va_pruned_int8_convrot.safetensors ★ |
diffusion_models |
~21 GB | Ref / Cont scenes, character swap |
minimax_h3_fl2va_pruned_int8_convrot.safetensors ★ |
diffusion_models |
~21 GB | First / Trans scenes, F/L mode |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors ★ |
text_encoders |
~17 GB | the text encoder the shipped workflows load |
minimax_h3_video_vae_fp16.safetensors |
vae |
1.4 GB | video VAE |
minimax_h3_audio_vae_fp32.safetensors |
vae |
0.4 GB | audio VAE |
The quickest way is the Hugging Face CLI (installed by start.bat):
hf download Comfy-Org/MiniMax-H3 `
diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors `
diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors `
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors `
vae/minimax_h3_video_vae_fp16.safetensors `
vae/minimax_h3_audio_vae_fp32.safetensors `
--local-dir comfy\ComfyUI\modelsAlternatively open any of the shipped workflows (comfy_workflows\*.ui.json)
in ComfyUI — its Missing models dialog lists each file with a download
link, including the optional 4-step turbo LoRA some workflows use.
The
fp8_scaleddiffusion files are the fallback ifint8_convrotwill not load on your setup. The GGUF files in Settings → Models /misc\download-models.ps1are for the older sd-cli engine only; ComfyUI mode does not need them.
Current ComfyUI includes the MiniMax H3 nodes (update ComfyUI if a
MiniMaxH3… node is reported missing). Some features need extra packs — install them from ComfyUI-Manager → Install Missing Custom Nodes
(open the matching workflow from comfy_workflows\ first), or with
git clone into ComfyUI\custom_nodes\:
| Feature | Node pack |
|---|---|
| Native-audio scene board with a song | ComfyUI-H3-NativeAudioLock (MiniMax-H3-NativeAudio-MusicVideo-Workflow) |
| Character swap | ComfyUI-VideoHelperSuite, rgthree-comfy |
| ControlNet preprocessors | comfyui_controlnet_aux |
| Face fix | ComfyUI-Impact-Pack + ComfyUI-Impact-Subpack (and models\ultralytics\bbox\face_yolov8m.pt) |
| RefMod character identity | ComfyUI-MiniMaxH3Mod |
| Sage attention / Spectrum speed-ups | installed by the setup window (KJNodes, Spectrum) |
| Upscaling a take | installed by the setup window (MMH3 Ultimate Upscale) |
When a render needs a node ComfyUI does not have, the job fails with the node's name and the pack to install — nothing fails silently.
The screenplay writer, the prompt chat and ✨ Enhance run a local model through llama.cpp:
- Settings → LLM → ⬇ Install engine (or run
powershell -ExecutionPolicy Bypass -File misc\get-llama.ps1). It downloads the newest Windows CUDA build ofllama-serverintollama-cpp\. - Put a chat model (
.gguf, any instruct model your VRAM allows) intollm-models\. - Pick it in the LLM dropdown under the prompt box.
For Describe image (the vision model), run misc\get-qwen3vl.ps1 — it adds
Qwen3-VL and its projector next to the same llama.cpp.
With two GPUs, Settings → LLM → Device puts the LLM on the other card (ComfyUI keeps its own); on one card, pick an LLM small enough to sit beside the video model or unload it before long renders.
| Engine | Install | Used by |
|---|---|---|
| OmniVoice (voice cloning / TTS) | Settings → Audio → Install OmniVoice | voice clones, TTS, Read replies |
| faster-whisper (speech-to-text) | Settings → Audio → Install faster-whisper | transcription, 🎤 Speak |
| demucs (stems) | the Stems docker's ⚙ Install | stem separation |
| YuE2 (songs) | Settings ▸ Add-ons ▸ YuE2 music (or misc\get-yue2.ps1) |
Music docker → YuE2 |
| MiniMax Music3 (songs) | ComfyUI workflow audio_minimax_music3.json |
Music docker → MiniMax |
| sd-cli (legacy engine) | misc\get-sd-cli.ps1 |
only when Settings → ComfyUI → Engine is sd-cli |
Each engine gets its own venv / folder under the app; nothing is installed globally.
cd D:\bitgen-studio
git pullthen run start.bat again. Your work is never in the repository:
projects\, styles\, graphs\, custom.settings (your settings), the
models and the engines are all ignored by Git, so a pull never touches them.
| Symptom | Fix |
|---|---|
| ComfyUI is not running and could not be started | Settings → ComfyUI: check the folder and URL; the Console docker shows ComfyUI's own output. |
| A job fails naming a node type | install that node pack (section 5), restart ComfyUI. |
OutOfMemoryError / a render dies near the end |
lower the resolution or the scene length; leave Free VRAM before every generation on. The app remembers the largest clip your card finished and splits longer ones. |
| Very slow first render | ComfyUI loads ~40 GB of weights the first time; later renders reuse them. |
You need pytorch with cu130 in the ComfyUI console |
RTX 50-series: update the driver and reinstall ComfyUI's torch (see minimax-fixed.txt). |
| No sound on a take | the take's badge says whether the file has an audio stream; a native-audio scene always has one. |
| The app window never appears | run with H3_NO_SHELL=1 and use the browser at http://127.0.0.1:8311. |
| A black or frozen app window, or the UI dragged off the screen | close bitGen Studio and run misc\reset-layout.bat in the app folder: it puts back the default layout and theme and clears the desktop app's caches (projects, settings and models are kept; what it resets goes to backups\layout-<time>\). In a browser, the same reset is the line the bat prints for the F12 console. |
Still stuck? Open an issue with the Console docker's output and the failing job's log from the Queue docker.