Files
Frank Schwenk fd024d31e9
CI / lint-and-test (pull_request) Successful in 2m3s
docs: update README and add ARCHITECTURE
Dev setup (venv, ruff, pytest), config table entries, architecture
overview, and cleanup completion report.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-18 17:12:26 +02:00

4.3 KiB

live.f12.rocks

Event photo pep pipeline: guests drop a photo into a Syncthing folder, a worker runs it through rembg (background removal) + gmic (random filter/blend compositing, ported from make_random.py), and a small web UI shows the results with a "remix" option to try different filters on the same photo. Finished jobs sync back via the same share.

No web auth. This is an event tool, not a DAM.

Services (compose.yml)

Service What Port
worker Watches incoming/, runs the compose pipeline none published
web FastAPI + Jinja UI, browse jobs, remix via Traefik only

Both share one host directory (DATA_HOST_DIR) mounted at /data. Syncthing runs on the host — not in this compose.

Data layout (Syncthing share)

Prod path: /home/frank/sync.schwenk.online/data/livef12

incoming/                 # phone drop — worker only ever reads/copies from here
variants/
  {stem}_original.jpg     # preprocessed private copy
  {stem}_rembg.png
  {stem}_v1.png           # final composed images
intermediates/
  {stem}_v1_bg_filtered.png
  {stem}_v1_post_0.png    # …
meta/
  {stem}.json             # status + manifest for the web UI
processed.json            # worker bookkeeping: which incoming files were handled

Phone tip: .stignore can drop meta/ and processed.json if you only want images on the device.

assets/ (filter lists + trimmed filters.json) lives in the repo and is bind-mounted read-only into worker/web at /data/assets.

Running it

cp .env.example .env
# local: set DATA_HOST_DIR=./data
# prod: leave unset (defaults to the Syncthing share path)
docker compose up -d --build

Drop a photo into incoming/ (via Syncthing or locally). After a few seconds (poll interval + processing time) it shows up on the web UI; flat files land in variants/ and intermediates/ for the phone.

Development

Python 3.11 (same as Dockerfile). On a host with newer Python, use Docker or a 3.11 venv — pinned onnxruntime does not publish wheels for 3.14+.

python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt

ruff check app/ tests/       # lint
ruff format app/ tests/      # format
pytest -q                    # unit tests (no gmic/rembg)

CI (.gitea/workflows/ci.yml): ruff + pytest on push/PR.

Architecture notes: ARCHITECTURE.md. Cleanup log: CLEANUP_PLAN.md.

Configuration (.env, see .env.example)

Var Default Notes
DATA_HOST_DIR /home/frank/sync.schwenk.online/data/livef12 Syncthing share root; use ./data locally
OUTPUT_COUNT 3 Variants generated per uploaded photo
BLEND_OPACITY 30% Default blend opacity for auto-generated variants
FILTER_TIMEOUT 120 Seconds before a single gmic call is killed
MAX_FILTER_ATTEMPTS 8 Random-filter retry budget per bg/fg pick
MAX_EDGE_PX 2000 Longest edge after ImageMagick preprocess
PREPROCESS_TIMEOUT 120 Seconds for preprocess step
NICE_LEVEL 18 nice -n level for gmic/rembg subprocesses

Deploy notes / caveats

  • Firewall: Traefik must already route live.f12.rocks — this repo only adds the router labels; it assumes the external traefik docker network exists. No SFTP port needed.
  • rembg model download: first background-removal call downloads the u2net ONNX model (~176 MB) from GitHub. This needs outbound internet on first run and can take a minute or two depending on the link; the model is cached in the rembg_cache named volume afterwards, so restarts don't re-download it.
  • Incoming is append-only: the worker only ever copies out of incoming/ and never deletes or moves anything there — plan disk space accordingly, or clean up incoming/ manually between events.
  • Sequential processing: the worker handles one photo at a time (cpus: "1.0", no concurrency) — fine for an event pace, but a burst of uploads will just queue up and get processed in order.
  • Deploy itself (docker compose up -d on the server) is on you — this repo doesn't run it automatically.