.env.example

# Environment variable reference
#
# How to use: copy this file to `.env` (same directory) and edit the values.
#             The backend loads `.env` automatically (python-dotenv) and it is
#             git-ignored, so putting your password here is safe.
#
# Priority: process environment variable > .env file > defaults in code
#
# Administrators use three fixed accounts: 001 through 003.
# THREEDGS_ADMIN_PASSWORD is no longer used for login.

# ---- Optional ----
# Data directory (photos + database + training output). Defaults to ./data.
# Point it at a large disk if you have one:
# THREEDGS_DATA_DIR=E:\campus-splat-data
#
# Or keep only the photos on another disk (photos are the bulky part).
# Absolute path, or relative to the repo root. Defaults to <data>/uploads.
# With an external photo folder the database stores absolute paths, so moving
# that disk later means updating this variable too.
# THREEDGS_UPLOAD_DIR=F:\campus-photos

# Login session lifetime in hours (default 720 = 30 days)
THREEDGS_SESSION_TTL_HOURS=720

# Per-photo size limit (MB) and how many files a single upload may contain
THREEDGS_MAX_UPLOAD_MB=60
THREEDGS_MAX_FILES_PER_REQUEST=20

# ---- Quality check of uploads ----
# An upload returns as soon as the file is on disk; the heuristic check runs in a
# background worker, so the volunteer can start the next batch immediately and
# watches each photo flip from "质检中" to its verdict (and checks the photo
# again after a crash/restart). Set THREEDGS_QUALITY_INLINE=1 to check inside the
# request instead — the reply then carries the verdict right away (handy for a
# tiny single-user setup; that is also what the test suite uses).
# THREEDGS_QUALITY_INLINE=0
# How many photos may be checked at once (1-4). One worker keeps the club-room
# machine responsive while volunteers are uploading.
# THREEDGS_QUALITY_WORKERS=1

# Login rate limiting: an IP is locked out after too many failures in the window
THREEDGS_LOGIN_MAX_FAILURES=8
THREEDGS_LOGIN_LOCKOUT_SECONDS=300

# ---- Training (see docs/training-pipeline.md) ----
# mock (default; walks the whole graded pipeline without doing real work) | real
THREEDGS_TRAINING_MODE=mock
# Path to the colmap executable (real mode)
THREEDGS_COLMAP_BIN=colmap
# COLMAP GPU switch. Set to 0 when COLMAP was built without CUDA (CPU-only):
# SfM on CPU is slower but perfectly usable — and it is the only stage that can
# fall back to CPU. 3DGS training itself still needs an NVIDIA GPU.
# THREEDGS_COLMAP_USE_GPU=0
# COLMAP vocabulary tree. Indoor scopes (thousands of unordered phone photos)
# need it: without it the matcher falls back to sequential matching.
# Download: https://demuc.de/colmap/ (vocab_tree_flickr100K_words*.bin)
# THREEDGS_VOCAB_TREE=D:\colmap\vocab_tree_flickr100K_words256K.bin
# Command template for the original 3DGS trainer (real mode).
# Placeholders: {source} {model} {images} {iterations} {resolution} {data_factor}
# {source} and {model} are mandatory (that is how the produced ply is found);
# any other {...} is passed through untouched.
# THREEDGS_GS_COMMAND=python train.py -s {source} -m {model} --iterations {iterations} --resolution {resolution}
# Same, for the gsplat toolchain (about 1/5 of the VRAM, docs/training-toolchain.md §二).
# Install it with: powershell -ExecutionPolicy Bypass -File scripts\setup_gsplat.ps1
#   --disable_viewer  unattended runs must not wait for a browser viewer
#   --save_ply        gsplat defaults to save_ply=False: without it no ply is written
#   --ply_steps N     write the point cloud at step N (step counting starts at 0)
#   --data_factor N   image downsampling, controlled by the training page
#   --eval_steps 1e6  skip evaluation (a single room has no held-out split)
# THREEDGS_GSPLAT_COMMAND=D:\gsplat-env\Scripts\python.exe D:\gsplat\examples\simple_trainer.py default --data_dir {source} --result_dir {model} --max_steps {iterations} --save_steps {iterations} --eval_steps 1000000 --ply_steps {iterations} --save_ply --data_factor {data_factor} --disable_viewer
# Or override the whole thing with your own pipeline script:
# THREEDGS_TRAINING_COMMAND=python scripts\my_pipeline.py
# How many runs may execute at once (a single 3090 -> keep this at 1)
THREEDGS_TRAINING_MAX_CONCURRENT=1
# GPU memory used for the preflight warning (GB) and the gaussian-per-GB rule.
# 24 GB / ~500k gaussians per GB is the club-room 3090.
THREEDGS_VRAM_GB=24
THREEDGS_GAUSSIANS_PER_GB=500000