comfyui-vrgamedevgirl: a scene-by-scene video builder inside ComfyUI

A node pack that outgrew its description

The repository description calls it custom ComfyUI nodes for film grain, color matching and video enhancement, which is accurate but undersells what is there. The README describes a growing collection of custom ComfyUI nodes for AI video creation, music videos, storyboarding, image generation, video enhancement, face repair, editing, LoRA training and workflow automation.

The directory listing backs that up. There are directories for the builder, storyboard, core, general and post_process code, plus flow_automation, prompt_creator, llm, runner, browser, minimax, optional_nodes, web and assets folders, a Workflows directory holding the shipped workflow JSON files, an mcp_server directory, an agent_api directory, scripts, tests, docs, LUTS and srt_files. There are also AGENT_GUIDE.md, CLAUDE.md, CHAT_AGENT_PROMPT.md, MUSIC_VIDEO_AGENT_PROMPT.md and Api_Endpoints.md, which tells you the project has been built with, and partly for, AI coding assistants.

The pack is written in JavaScript and Python, and the top-level listing includes a __init__.py, which is how ComfyUI discovers a custom node folder. The project metadata in pyproject.toml names it comfyui-vrgamedevgirl at version 9.1.1, requires Python 3.10 or newer, and carries the ComfyUI publisher fields with a publisher id of vrgamedevgirl and the display name VRGameDevGirl Custom Nodes.

The licence is the one detail to check first. The repository shows a custom licence that GitHub cannot classify, with the terms in the LICENSE file, and pyproject.toml points the licence field at that same file rather than naming a standard licence. For a personal creative tool that is a low-stakes question; for a studio shipping client work, read it.

The AI Video Builder and the engines behind it

The flagship tool is the AI Video Builder, described as a scene-by-scene production workspace for LTX 2.3, MiniMax H3 and future video engines that brings planning, prompting, media generation, timing, review and final assembly together inside ComfyUI.

That list of stages is the feature set. You can build a project from songs, audio, SRT files, lyrics or manually timed scenes, so a project can start from a track with lyrics and a subtitle file rather than from an empty timeline. The Wizard, Storyboard Builder and Reference Builder are the three guided entry points, and the plan-and-generate step supports character and location references so that a face or a place stays consistent between scenes.

On the generation side, the Builder covers image-to-video, text-to-video, ID-LoRA and First/Last Frame scenes, and it lets you chain or keep independent First/Last Frame sequences for continuity. The lyric review step maps singers and marks instrumental or B-roll sections, which is what you need when the video has to match a vocal arrangement rather than a metronome.

The finishing tools are part of the same workspace: face repair, clip enhancement, post-processing and result comparison, a visual timeline for previewing timing, beat marker calibration and stitching the final video, and saving, branching, exporting and importing portable Builder projects so a session can be resumed or shared.

Text generation is wired in through built-in LLM, local LM Studio, API and Browser AI options where supported. The distinction matters for a production user, since the same workflow can run against a hosted model, a local model on your own machine, or a browser-based one.

Installing through the Manager or by hand

The recommended route is ComfyUI Manager. You open Manager, go to Install Custom Nodes, and search for vrgamedev or install from the repository URL:

https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Then you restart ComfyUI and hard refresh the browser page. That last step appears three times in the README and it is not boilerplate, because a browser that has cached the old JavaScript UI files will keep showing a stale interface after a node pack updates.

The manual route is a clone into the custom_nodes directory:

git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

After that you install the Python requirements using the same Python environment that runs ComfyUI. For the Windows portable build, the README gives these commands to run from the ComfyUI_windows_portable directory:

python_embeded\python.exe -m pip install --upgrade pip setuptools wheel
python_embeded\python.exe -m pip install Cython scikit-build-core
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\comfyui-vrgamedevgirl\requirements.txt

After installation you add the node named VRGDG AI Video Builder UI, create a project, add audio or SRT timing or scenes, plan with the Wizard or the Storyboard Builder, then generate and approve scene images, render the scene videos and stitch the result.

The quick start also advises starting with a short project or a few scenes before committing to a long render, and saving often.

The dependency wall and the Python version trap

This is where a casual installation goes wrong, and the README is unusually direct about it.

The first two pip commands prepare build tooling needed by voxcpm and llama-cpp-python, especially on Windows with Python 3.13. The README then adds that llama-cpp-python may also need a working CMake/Ninja and a C++ compiler when a compatible prebuilt wheel is unavailable, and that Python 3.12 is the safer choice for older Windows portable environments.

So the failure mode is predictable. On a portable Windows install with Python 3.13 and no compiler, the dependency install fails, and the error appears at install time rather than when you open the Builder. The fix is either an older interpreter or a full build toolchain, and the README's advice is to pick the older interpreter.

The dependency set itself is large. The pyproject list includes kornia, librosa, imageio, torchcodec, google-generativeai, av, stable-ts, demucs, transformers, accelerate, huggingface_hub, Cython, scikit-build-core, voxcpm and llama-cpp-python, while requirements.txt lists a slightly different set that includes opencv-python and drops kornia, imageio, av and voxcpm. That difference between the two manifests is itself a small warning, since the two files do not describe exactly the same environment.

The classifiers say Operating System :: OS Independent and Environment :: GPU :: NVIDIA CUDA. The independence claim is about the OS, not the hardware: the video models this pack drives are GPU workloads, and the resource monitor section talks about NVIDIA VRAM readings through the driver's nvidia-smi.

The README also warns that some advanced tools need additional models or external components, and points to the Builder guide for supported workflows, required custom nodes, model locations and optional setup. That is the honest scope of what the pack does out of the box.

The nodes that stand on their own

Not everything requires the Builder, and for many workflows the standalone nodes are the reason to install the pack.

The three lightweight effects are grouped as Fast Film Grain, Color Match and Sharpening nodes, described as adding cinematic grain, matching a reference palette, and restoring edge detail efficiently across image batches. There is a broader set of LUT, colour, grain, sharpening, resize, combine and general video-processing nodes alongside them.

The heavier tools are more interesting. The Face Fix node set detects and tracks a face, enhances guided anchors, processes the crop through LTX, and composites the repaired face back into the source video. The Video Enhance node set creates guided enhancement anchors, processes a full video through LTX, and restores the exact original resolution and frame count, which is the part that matters: enhancement passes that silently change your frame count or rescale your footage break the timeline downstream.

For training, there is a VRGDG LoRA Dataset Creator UI for building captioned datasets for styles, characters and experimental edit pairs, plus training workflows for MiniMax H3, LTX 2.3 and Z-Image covering standard video, audio, audio/video and Speed LoRAs, and a Musubi-Tuner Installer for supported training environments.

For audio there is a VoxCPM2 Voice Clone / TTS node for speech from text, voice design, continuing spoken audio, and cloning a voice from a reference clip, alongside audio loading, splitting, timing, transcription and silent-audio helpers. There are also LLM prompt tools for structured image and video prompting, image comparison tools, and preview and grid plot nodes for comparing checkpoints, prompts, strengths and generated video folders.

The resource monitor and the optional canvas strip

The AI Video Builder carries a small monitoring feature that is worth understanding because of what it does when it cannot read a sensor.

The Builder shows host RAM and NVIDIA VRAM usage in its top toolbar, refreshing roughly every two seconds while the Builder is open, and the compact display hides itself automatically when the toolbar is too narrow. GPU readings require the driver's nvidia-smi, and the README is explicit that RAM remains available when GPU sensors are unsupported. That fallback is the interesting part: on a machine without NVIDIA tooling, the RAM readout keeps working instead of the whole panel disappearing.

The optional canvas strip is enabled under Settings, VRGDG, Resource Monitor. It adds GPU load, temperature, RAM and VRAM bars, optional fan, clock and power details, and an idle-only Clear memory action. The idle-only restriction is a safety measure: the action only becomes available when the GPU is not busy, so you cannot free memory out from under a running render.

Readings come from the machine running ComfyUI, which is worth remembering for a shared or remote setup: if ComfyUI runs on a server, the panel reports the server's hardware, not the browser's.

The rest of the pack is oriented at automation and agent access. There is an mcp_server directory with a start_mcp_server.bat script, an agent_api directory, Api_Endpoints.md, and a set of agent prompt documents. For a team that wants a model to drive the workflow, that is the part to read, and it is also the part with the largest security surface, since it exposes an interface that a tool can call.

Licence, branches and the state of the repository

Three maintenance facts are visible and they should shape your expectations.

First, the default branch is Beta2.0, not main. For a project of this size that is a strong signal about how the code is expected to be consumed, and it means the branch you get by default is labelled as a beta.

Second, there are no GitHub releases. The version number in pyproject.toml is 9.1.1 and the repository has an update_notes.json, so change tracking happens through the repository rather than through published artefacts. There is no tag to pin and no changelog release entry to diff against.

Third, the repository is not archived and the last push was on 2026-10-16, so work is landing. The upstream ComfyUI ecosystem moves quickly, and a node pack that ships workflows referencing LTX 2.3, MiniMax H3 and Z-Image models is tied to model releases and to the custom nodes it expects to find alongside it.

The licence is custom and unclassified by GitHub, with the terms in the LICENSE file. Since the pack bundles workflows, prompts and node code, and leans on many external models and custom nodes, the practical licensing question for a commercial user is not only the pack's own terms but whether the models and the assets you feed it carry their own restrictions. This is not legal advice.

Taken together, the pack is a large, actively changing, GPU-dependent creative toolchain with a real learning curve in its documentation, whose documentation lives in a Workflows subfolder rather than in the root README. It rewards someone who reads the guide before installing, and it punishes someone who clones it and restarts ComfyUI expecting a drop-in node.

Editorial conclusion

Adopt this pack if you are producing music videos or storyboard sequences inside ComfyUI and want planning, timing, review and assembly in one place instead of a folder of loose workflows. Do not adopt it for a single clean upscale or a grain overlay, where the standalone nodes are more than enough, and do not install it into a portable ComfyUI on Windows without reading the Python version guidance, because voxcpm and llama-cpp-python can require a compiler toolchain. Before you commit, check the requirements.txt against your environment, confirm you have an NVIDIA GPU with the CUDA stack for the video models, and read the Video Builder guide in Workflows/LTX-2_Workflows/Video_Builder/readme.md, since that file, not the repository root, is where the setup details live.

Frequently asked questions

Official sources

1. Issues
2. README
3. vrgamegirl19/comfyui-vrgamedevgirl on GitHub