Why DocStellar?
Astrophotography is a discipline where every image results from a dozen sensitive variables: telescope collimation, sensor flatness, tracking quality, weather, processing. A single error, a sensor tilted by 0.1°, a bad flat, a failed guide, leaves a visible mark.
Identifying those marks requires the trained eye of an experienced astrophotographer and the patience to compare the photo against hundreds of known examples. On forums, this process takes hours, sometimes days. And it does not always succeed: too many contradictory opinions, too many 'average' images to serve as reference.
DocStellar automates this first diagnosis. You submit a photo, the Doc examines it, and you leave with a list of likely defects, each with its precise area, probable causes and recommended course of action. Not to replace human expertise, but to give you a well-grounded starting point.
The Doc
The Doc is an astrophoto diagnostician, driven by a vision-language model trained on thousands of example images. It has been briefed on a taxonomy of 42 common astrophotography defects, grouped into 8 families: optics, mechanics, electronics, atmospheric, acquisition, calibration, processing, environmental.
Its method
For each image submitted, the Doc:
- Segments the image looking for suspicious areas (aberrant stars, gradients, noise patterns).
- Cross-references each observation with its knowledge base to identify the most likely defect.
- Produces a complete diagnosis: defect name, precise area, severity (from marginal to critical), confidence (in %), differential diagnosis when several causes are plausible, and recommended course of action.
The Doc always explains why it believes it sees a defect, where exactly, and how to fix it. It does not merely name the problem.
How the analysis works
The 5-step pipeline:
- Submission: JPG, PNG, FITS or XISF, up to 1 GB. For FITS/XISF, the Doc extracts a clean preview (asinh stretch on percentiles 0.5%–99.5%).
- Reduction: the image is downscaled to 2048 px maximum to stay within the vision model token budget.
- Examination: the vision-language model analyses the image, segments suspicious areas, and outputs structured JSON with one defect per observation.
- Annotation: the server draws coloured bounding boxes by defect family, with a numbered legend and zooms on the most problematic areas.
- Delivery: you receive the annotated image along with the detailed list of defects (causes, corrections, glossary pages).
The defect taxonomy is hand-coded from feedback from experienced astrophotographers, technical papers (CMOS noise modelling, optical aberrations) and experience accumulated on specialist forums.
The Doc's limitations
The Doc can be wrong. A few typical errors:
- Confusing sensor tilt with incorrect backfocus (similar visual signatures).
- Attributing to light pollution a gradient that is actually caused by an imperfect flat.
- Diagnosing amp glow on an image where it is in fact an internal reflection.
That is why each defect displays:
- an explicit confidence score (0 to 100%): a 40% diagnosis is not the same as a 90% one;
- a differential diagnosis when several causes are plausible, with the defect page explaining how to distinguish them;
- a severity rating, to weigh the urgency of acting.
Treat the Doc's recommendations as a well-grounded starting point, not a definitive verdict. When in doubt, validate with a second pair of human eyes.
Who is behind DocStellar?
DocStellar is a solo project, designed and maintained by an amateur astrophotographer. The code is written in Next.js (TypeScript) with Postgres as the database, sharp for image processing, and an AI vision-language model as the brain of the Doc.
Contributing
- A missing or poorly covered defect? Write via the contact form: the taxonomy evolves with feedback.
- An aberrant analysis? All public diagnoses are viewable in the gallery: the Doc improves when its errors are pointed out.
- A bug or feature suggestion? Contact is open.