Iterate-and-compare astrophotography stacking. Tweak a slider, see the result, undo, branch, publish.
astrolab is a local-first workbench that wraps Siril, GraXpert, and StarNet++ behind a typed pipeline graph and a content-addressed cache. Each "project" is a live edit of a stack: changing a knob in the UI submits a fresh job, but every upstream node it didn't touch (calibrate, register, stack) is a free cache hit, so re-renders feel cheap. Every revision shows up in a history strip; promote keepers to a public gallery.
It is built around the Dwarf 3 capture layout for now, but nothing in the pipeline is Dwarf-specific. Want your scope supported? See the Adding a scope section in CONTRIBUTING.md; it starts with running a small collector script, not uploading raw files.
- Wraps Siril 1.4 for calibration, registration, stacking, and pixel ops.
- Drives GraXpert for background extraction and ML denoise.
- Drives StarNet++ v2 for star removal, replacement, and recombination.
- Caches every node's output by
(inputs + params + node version)so editing a downstream knob never re-runs an upstream node. - Tracks a per-project edit history (revert, branch, compare, publish).
- Surfaces a Tonight planner with altitude/visibility for the targets you've already captured, so you know what's worth re-imaging.
- Gallery view surfaces the renders you opt in to publish; everything else stays in your private history.
No local install needed. The image bundles Siril 1.4 and GraXpert 3.0.2. New to Docker? See INSTALLING.md for a step-by-step setup guide.
CPU-only (most users):
docker run -d \
--name astrolab \
-p 8000:8000 \
-v ~/Pictures/Siril:/captures:ro \
-v astrolab-data:/data \
ghcr.io/bscholer/astrolab:latestOpen http://localhost:8000. The container auto-scans /captures on first run, so your sessions should appear within a few seconds.
NVIDIA GPU (for GraXpert and StarNet++ AI nodes):
You need the NVIDIA Container Toolkit installed on the host first.
docker run -d \
--name astrolab \
--gpus all \
-p 8000:8000 \
-v ~/Pictures/Siril:/captures:ro \
-v astrolab-data:/data \
-e ASTROLAB_ENABLE_STARNET=1 \
ghcr.io/bscholer/astrolab:cudaASTROLAB_ENABLE_STARNET=1 tells the entrypoint to download StarNet++ v2 on first run and store it in the persistent /data volume. Omit it if you don't need star removal.
Mount points:
| Path | Purpose |
|---|---|
/captures |
Your raw capture tree (read-only is fine; astrolab never writes here) |
/data |
Persistent state: cache, projects, settings, optional StarNet++ binary |
git clone https://github.com/bscholer/astrolab.git
cd astrolab
# python deps + venv (uv reads pyproject.toml)
uv sync
# UI deps
cd ui && npm install && cd ..
# run the API on :8000
.venv/bin/uvicorn server.api:app --reload --host 0.0.0.0 --port 8000
# in another terminal: vite dev server on :5173 (proxies /api to :8000)
cd ui && npm run devOpen http://localhost:5173, point the captures path at your Siril folder in Settings, hit Refresh, and projects show up.
For a single-port setup mirroring production (http://localhost:8000), cd ui && npm run build once instead.
The Docker image is the supported runtime. It's amd64-only and runs anywhere Docker does (Linux, macOS via Docker Desktop, Windows via WSL2).
- CPU only (the
:latestand:cputags) works on any amd64 box. GraXpert and StarNet++ fall back to CPU; stacking is fine, AI nodes are noticeably slower. - NVIDIA GPU (the
:cudatag, plus--gpus alland the NVIDIA Container Toolkit) is recommended if you want GraXpert and StarNet++ to fly. Tested on Ampere and Ada. - From source (Linux only, see Quick Start above) works for dev. macOS source installs run everything except the Siril paths; the test suite mocks the Siril subprocess so the dev loop stays full-featured, but final validation belongs on a Linux box.
The image is amd64; on M-series Macs it runs through Rosetta. Siril, the calibrate / register / stack / save_image pipeline all work fine under Rosetta. GraXpert and StarNet++ do not — their PyInstaller-bundled Python stack hits a Rosetta amd64 emulation bug and segfaults at startup. Astrolab's pipeline will therefore crash on those nodes when run on Apple Silicon via Docker Desktop. The Siril-only path through stack works end to end and produces a usable PNG. For full pipelines with GraXpert / StarNet, run on a Linux or amd64-native host.
Phase 0. Not stable, not packaged, schemas can change between commits. Useful enough to process my own captures every weekend; rough enough that I'd hate to ship it as v1.0.0 today. The V1 milestone tracks the punch list.
The actual image processing is done by external tools that astrolab orchestrates:
- Siril (GPLv3) - calibration, registration, stacking, plate-solving.
- GraXpert (GPLv3) - background gradient extraction and ML denoise.
- StarNet++ v2 (proprietary freeware) - star-nebula separation.
The calibrate_register_stack template follows the pipeline shape from Naztronomy's smart-telescope script. No code is copied; astrolab reimplements the same Siril command sequence as graph nodes so each step is cacheable. Stage order is his.
The multi-scope ingest classifier (NINA, ZWO ASIAIR, Seestar) and the cross-scope filter alias table were derived clean-room from public header dumps in starbash (Kevin Hester, GPLv3). No starbash code is in astrolab; the mappings are factual observations about how each capture program writes its FITS headers.
Full notes in LICENSES/THIRD_PARTY.md.
- Design doc, the original README. Pipeline contracts, cache shape, node authoring guide.
- Contributing. Dev setup, test/lint commands, "add a node" walkthrough.
- Issue tracker. V1 board.
MIT, with a CC BY-SA 4.0 carve-out for the vendored OpenNGC catalog under catalogs/openngc/ (Mattia Verga). See LICENSE for both.
Tonight thumbnails come from the Digitized Sky Survey (STScI/NASA), coloured and tiled by CDS and hosted by ESA ESASky. The server downloads them on demand and does not bundle them. See the DSS copyright terms.
The Docker image bundles Siril (GPLv3) and GraXpert (GPLv3) as separate processes; astrolab's MIT license applies only to astrolab's own source code. See LICENSES/THIRD_PARTY.md for details.





