Tech Splat Generator

One photo in, a 3D Gaussian splat out. A local web tool that lifts a single photograph into a point cloud, flies a camera around it, has a video model turn that fly-around into a photoreal clip, and builds a training dataset carrying the exact camera poses the orbit was authored with.

Beta, for developers and tinkerers. Tested on Windows with an NVIDIA GPU. Video generation runs on fal.ai and is paid per clip on your own account.

How it works

The camera orbit is decided before anything is generated, so the control render, the AI video and the trained splat all turn in lockstep: frame i is the same camera in every one of them. That is the whole point of the design — when the path is authored rather than solved after the fact, pose error is removed by construction instead of being estimated.

The app walks you through five steps in a left-hand rail, with a 3D viewport on the right that previews everything live.

Step	What happens
1 Source	Drop a photo. MoGe-2 measures the lens automatically.
2 Cloud	The photo is lifted into a point cloud with the depth model you pick: MoGe-2 (metric, default), Depth Anything V2 Small / Large, or Depth Anything 3 Metric / Mono. Clean it up or isolate the subject; it re-lifts as you change settings.
3 Shot	Author the camera orbit (length, sweep, radius, aim height, lens). The server renders a control video of the point cloud from exactly those cameras.
4 Generate	The control video and your photo go to a fal.ai video model (LTX 2.3 render-to-real, Wan 2.2 VACE depth-control, Wan 3.0 Prime, MiniMax H3). An approval window shows exactly what will be sent before anything is paid for.
5 Review	Watch the clip, cut a matte of the subject (BiRefNet, SAM 2.1, RMBG-1.4, optional MatAnyone), retime if needed, then Build Dataset: the frames, their mattes, the authored poses and an init point cloud, written to projects\<name>\dataset\.

That dataset is where this tool stops. Open the folder in Brush and train there, with whatever step count and settings you want. Brush writes its .ply exports into the project's splat_out\ folder, and the viewport's checkpoint switcher loads any that are there.

Every file the tool reads, writes or uploads is logged with its full path in the console at the bottom.

Requirements

• Windows 10 or 11. Other platforms are untested (Brush paths and some defaults assume Windows).
• NVIDIA GPU, 12 GB of VRAM recommended, with a recent driver.
• Python 3.12
• Git — two packages install straight from GitHub.
• A fal.ai account and API key → https://fal.ai/dashboard/keys
• Brush, the Gaussian splat trainer → https://github.com/ArthurBrussee/brush/releases
• Several GB of free disk space for Python packages and model downloads.

Install

Full step-by-step instructions, including the optional extras and how to verify the install, are in INSTALL.md. The short version, in PowerShell:

git clone https://github.com/Techcopter/splatlab.git
cd splatlab

py -3.12 -m venv .venv
.venv\Scripts\activate

# PyTorch with CUDA first, then everything else
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu129
pip install -r requirements.txt

copy .env.example .env      # then put your fal.ai key in it

Put brush_app.exe at brush\brush_app.exe, then:

python run.py

Open http://127.0.0.1:8771. The Environment chip in the header turns red if the fal key or Brush is missing.

The first lift is slow: MoGe-2 downloads and loads once. After that, depth is cached per photo and re-lifts take well under a second.

The server has no auto-reload. Restart it after changing any .py file, and hard-refresh the browser (Ctrl+F5) after changing anything in web/.

Using it

• Drop a photo in Source. The lens is measured and the points are lifted without any button presses; Cloud opens by itself.
• Anything that runs shows a working card over the viewport with its progress, then a tick, a warning or a cross when it ends. Buttons show busy / done / failed on themselves.
• Playhead under the viewport: ← → step a frame (Shift for ten), Space plays, Home / End jump.
• Generate re-renders an out-of-date control video for you, then opens the approval window. Nothing is uploaded until you press Approve & send there.
• Build Dataset at the end of Review is the last thing the tool does; it stays greyed out until the matte is cut. Training is Brush's job.
• Projects live in projects\<name>\ (not committed). Use the ⋯ menu to rename, copy, export or import a project.

Configuration

config.json is optional. Copy config.example.json to create one. Every key has a default:

Key	Default	Meaning
brush	./brush/brush_app.exe	The Brush executable
projects	./projects	Where projects are stored
python	the running interpreter	Shown in the Environment panel
host	127.0.0.1	Keep it local: the server has no authentication
port	8771	Change it if the port is taken
log_ring	5000	Console lines kept in memory

Models and licences

No model weights are stored in this repository. Each is downloaded from its original source the first time a feature needs it, and each has its own licence. Read them before any commercial use.

Model	Used for	Source
MoGe-2	Depth, point cloud, lens	microsoft/MoGe, Ruicheng/moge-2-vitl-normal
Depth Anything V2 (depth option)	Relative depth	depth-anything/Depth-Anything-V2-Small-hf (Apache-2.0), depth-anything/Depth-Anything-V2-Large-hf. Large is non-commercial (CC-BY-NC-4.0)
Depth Anything 3 (optional depth option)	Metric or relative depth	ByteDance-Seed/Depth-Anything-3, depth-anything/DA3METRIC-LARGE, depth-anything/DA3MONO-LARGE (Apache-2.0)
BiRefNet / BiRefNet-HR	Subject matte	ZhengPeng7/BiRefNet, ZhengPeng7/BiRefNet_HR
SAM 2.1 + Grounding DINO	Tracked and text-prompted matte	facebook/sam2.1-hiera-small, IDEA-Research/grounding-dino-tiny
RMBG-1.4	Fast matte	briaai/RMBG-1.4
MatAnyone (optional)	Soft-edged video matte	PeiqingYang/MatAnyone. Non-commercial (NTU S-Lab 1.0)
fal.ai video models	Photoreal clip from the control video	Remote, billed by fal.ai under its terms
Brush	Splat training	Separate download

Troubleshooting

• "FAL_KEY is not set": check .env is next to run.py, then restart the server.
• "Brush not found": check the path shown in the Environment panel.
• CUDA out of memory: close other GPU apps, or use ⋯ → Free VRAM.
• Port already in use: set another port in config.json.
• Something failed: the console at the bottom has the full error and every file path involved. The error filter shows only failures.

More cases, with fixes, are in INSTALL.md.

Project layout

run.py              starts the server
server/             FastAPI routes, job runner, config
steps.py            geometry, rendering, depth, matting, datasets
falclient.py        fal.ai engines and prompts
web/                the browser app (plain ES modules, no build step)
web/vendor/         three.js, PlayCanvas, GSAP
projects/           your work (created on first run, not committed)
brush/              the Brush trainer you download (not committed)

Licence

The code in this repository is released under the MIT licence. That covers this project's own code only, not the models it downloads or the third-party libraries in web/vendor/, which keep their own licences.