Dependencies & Installation

Everything required to run the Splat Factory pipeline — converting DJI Mini 4 Pro footage to 3D Gaussian Splats. Ordered by installation sequence. Run items marked Manual before anything else.

Pipeline Overview

inbox/DJI_0001.MP4
        │
        ▼  worker.py  (inbox watcher → atomic job claim)
        Creates queue/project_timestamp.job → moves to working/
        │
        ▼  STAGE 1: ffmpeg
        Decode MP4 (NVDEC if available, software fallback)
        Apply LUT (DJI D-Log → REC709)
        Extract at 2fps, scale 1920px, mild sharpen
        → local_temp/project/frames/ (~240 frames)
        │
        ▼  STAGE 2: COLMAP
        Feature extraction (SIFT GPU), Sequential/Exhaustive matching
        Sparse reconstruction → quality check (warns if <75% registered)
        Undistort → images/ + sparse/
        → local_temp/project/colmap/
        │
        ▼  STAGE 3: 3DGS Training
        GTX 1080 safe flags auto-applied (data_device cpu, 20k iters)
        gsplat (pip) or GraphDECO repo — whichever found
        → local_temp/project/splat/*.ply
        │
        ▼  Copy to NAS
        splat/ → projects/project_name/splat/
        done/project.job  +  .done_X checkpoint flags

01 System Requirements Manual

NVIDIA GPU

GTX 1080 min · 8GB VRAM

Required

CUDA-accelerated COLMAP feature extraction and 3DGS training. GTX 1080 safe mode auto-applied (data_device=cpu, 20k iterations).

CUDA Toolkit

11.8 (cu118)

Required

Must match your GPU driver. GTX 1080 → CUDA 11.8. Install before PyTorch.

Downloaddeveloper.nvidia.com/cuda-toolkit

Python 3.10+

3.10 or 3.11 recommended

Required

Runs all pipeline stages. Use 3.10/3.11 for best gsplat compatibility.

Windowspython.org/downloads

wingetwinget install Python.Python.3.11

02 External Tools Manual — add to PATH

ffmpeg

6.x · NVDEC build

Required

Video decoding (NVDEC hardware acceleration), LUT application, frame extraction at 2fps, scaling, sharpening filter.

Windowsgyan.dev/ffmpeg/builds (full build)

wingetwinget install Gyan.FFmpeg

Linuxsudo apt install ffmpeg

COLMAP

3.9+ · CUDA build

Required

Structure-from-Motion reconstruction. Extracts camera poses from frames. Use the CUDA build for GPU-accelerated feature extraction. Exhaustive matcher for ≤500 frames, sequential for longer sequences.

Downloadgithub.com/colmap/colmap/releases

Docscolmap.github.io

git

2.x

Optional

Required only if using GraphDECO gaussian-splatting repo instead of gsplat pip package. Not required when using gsplat from pip.

Windowsgit-scm.com/download/win

Linuxsudo apt install git

03 PyTorch Manual — install first

torch + torchvision

2.x · cu118 build

Required

CUDA-enabled PyTorch. Must match your CUDA version. GTX 1080 → cu118. Install manually before gsplat — do not let pip resolve it automatically.

Selectorpytorch.org/get-started

cu118 whldownload.pytorch.org/whl/cu118

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118 COPY

04 Python Packages Auto-installed by run.py

gsplat

pip · preferred trainer

Required

Memory-efficient 3DGS trainer. Preferred over GraphDECO repo in 2026. Installs as a pip package with CUDA extensions compiled on install.

PyPIpypi.org/project/gsplat

Repogithub.com/nerfstudio-project/gsplat

pip install gsplatCOPY

numpy

1.24+

Auto

Numerical arrays. Used by gsplat, COLMAP output parsing, and frame validation.

pip install numpyCOPY

Pillow

10.x

Auto

Image I/O for frame validation and sample generation.

pip install PillowCOPY

tqdm

4.x

Auto

Progress bars for long-running COLMAP and training stages.

pip install tqdmCOPY

watchdog

3.x

Auto

Cross-platform filesystem watcher for the inbox/ folder. Triggers job creation on new .MP4 files.

pip install watchdogCOPY

nvidia-ml-py

pynvml

Optional

VRAM monitoring during training. Falls back to nvidia-smi subprocess if not installed.

pip install pynvmlCOPY

05 NAS Folder Structure Manual — create before first run

Folder Layout

Create this structure on the ASUSTOR NAS share before starting any workers. Workers use os.replace() for atomic job claims — all dirs must pre-exist.

splat-factory/
  inbox/         # drop .MP4 files here
  queue/         # auto-created .job files
  working/       # jobs in progress
  done/          # completed jobs
  projects/      # output .ply + frame samples
  luts/          # DJI D-Log LUT files (.cube)
  tools/         # ffmpeg, colmap binaries
    windows/
    linux/

DJI D-Log LUT

.cube format

Optional

Place DJI D-Log to REC709 LUT file in luts/. If absent, pipeline skips colour grading (still works). DJI Official LUTs available from DJI support or third-party packages.

DJI LUTsdji.com (official LUT download)