SPHEREx Blink

Type a star name or coordinates and flip through every image NASA's SPHEREx telescope has taken of that spot. Anything that moved or changed brightness between visits jumps out.

Live: spherexblink.vercel.app  ·  Pitch deck: PDF, PowerPoint

Built in 48 hours at NASA Space Apps 2026 for the Planet X and SPHEREx challenge by team Skill-Issue (Delhi): Arnav Aggarwal and Soumil Saini.

_{WISE 0855-0714, the coldest brown dwarf known, in three SPHEREx visits between May 2025 and May 2026 (detector D6, 4.7 to 5 µm). Watch the ring.}

Why we built it

Clyde Tombaugh found Pluto in 1930 by flicking between two photographs of the same stars until one dot jumped. From 2017, volunteers on Backyard Worlds did the same with NASA's WISE images and found more than 3,000 brown dwarfs, many of them with WiseView, a flipbook tool one of the volunteers wrote.

WISE stopped observing in 2024, when its NEOWISE mission ended. SPHEREx now photographs the whole sky every six months in 102 colours, so it is the survey to blink next. Its images come as 66 MB files behind an archive query language, though, and there was no flipbook for them. We built one.

What it does

• Search by name (resolved through CDS Sesame) or by RA/Dec in decimal or sexagesimal. Fast movers like Barnard's Star are centred on where they are in 2026, not where catalogues list them for 2000.
• Four views of the same patch:
  • Epochs: one stacked image per visit. SPHEREx comes back to each part of the sky about every six months, so this is the view for slow movers and brightness changes.
  • Difference: each visit minus a reference visit. A moving object shows up as a dark/light pair.
  • Frames: every single exposure, in time order or sorted by wavelength.
  • Static removed: single exposures minus the median sky.
• Any of the six detectors (0.75 to 5 µm) and a 3, 6, 12 or 20 arcmin field.
• A spectrum from every exposure. SPHEREx's filters change wavelength across the detector, so each exposure sees the target at a slightly different wavelength. Measuring the target in all of them gives a 0.75 to 5 µm spectrum per visit. When visits disagree at the same wavelength, the source changed.
• Measured motion. The target is located in each visit stack and a straight line is fitted through the positions. That gives a proper motion you can compare with the catalogue value, which is drawn as a track on the image.
• Gaia DR3 overlay. Stars from ESA's Gaia catalogue are circled on the image, so you can tell known stars from everything else.
• NEOSSat cross-match. Lists observations of the same field by the Canadian Space Agency's NEOSSat telescope, from the CADC archive.
• Blink like it's 1930. A/B blink between two frames, pan and zoom, Space to pause, arrow keys to step, and export the loop as a WebM clip.
• Flags. Pixels SPHEREx marked as saturated, ghosts/artifacts or transients can be tinted, bad pixels are dropped, and transients are left out of stacks.
• Share and embed. The URL holds the target, detector, view and field, so a link opens the same thing. /embed?... is a stripped-down player for an iframe.

Try these

Target	Opens with	What to look for
WISE 0855-0714	D6, 3′	Coldest known brown dwarf. Invisible at short wavelengths, clear at 4 to 5 µm, moves about 8″ a year west.
Barnard's Star	D4, 6′	Over 10″ a year north. The core saturates, so its position comes from the blanked-out centre.
Epsilon Indi B	D4, 6′	Brown dwarf pair. Its brightness swings with wavelength inside D4 (methane band at 3.3 µm).
RW Aurigae	D3, 6′	Young star whose disk keeps dimming it. Compare visits in the spectrum.
R Draconis	D3, 6′	Mira variable near the north ecliptic pole, so it has many more visits than anything else.
Comet 3I/ATLAS	D4, 6′	Interstellar comet. The preset is one fixed point on its path, so step through single exposures in the Frames view.
(1) Ceres	D3, 6′	The largest body in the asteroid belt, at its position during SPHEREx's first survey.
Betelgeuse	D3, 6′	Red supergiant in Orion and a pulsating variable. Bright enough to saturate.

Results on real data

Proper motions the app measured from public SPHEREx images (DR1 and QR3), next to the catalogue values:

Target	Measured	Catalogue
WISE 0855-0714 (D6)	RA: -7.9 ± 1.0″/yr	-8.1″/yr
Barnard's Star (D4)	Dec: +8.9 ± 1.2″/yr	+10.4″/yr
Epsilon Indi B (D4)	RA, Dec: +3.5, -2.2 ± 0.9″/yr	+4.0, -2.5″/yr

Three visits over one year are good to about 1″/yr, so the measurement is only useful for fast movers. Barnard's Star saturates, so it has to be located from its blanked-out core.

The Epsilon Indi B spectrum (247 clean points from 4 visits) shows the deep methane dip near 3.3 µm that marks a cold brown dwarf rather than a star.

How it works

browser ──/api/search──► IRSA TAP (spherex.plane + spherex.artifact)
        ──/api/frames──► HTTP range requests into each Level 2 image on IRSA
                          └► reproject onto a north-up grid, co-add per visit
        ──/api/spectrum► aperture photometry in every exposure, all detectors
        ──/api/motion──► gaussian centroids in each visit stack

Finding images. One ADQL query against IRSA's spherex.plane and spherex.artifact tables returns every Level 2 spectral image whose footprint contains the position. When the same exposure is in more than one release, the DR1 version wins over the weekly QR3 one. Exposures are grouped into visits wherever there is a gap of more than 30 days, and continuous stretches longer than 60 days (near the ecliptic poles) get split into separate visits.

Reading only what we need. Each Level 2 file is about 66 MB (2040×2040 image, flags, variance, zodiacal model, PSF cube, wavelength table). Instead of downloading files or using the cutout service, the backend sends HTTP multi-range requests for the header blocks, the image rows around the target, the compressed flag rows and the wavelength table at the end of the file. That is three or four round trips per exposure, and 368 KB for a 3′ field. The result matches IRSA's own cutouts pixel for pixel.

_{All 258 SPHEREx exposures of WISE 0855-0714 across the six detectors, sorted by wavelength from 0.75 µm (blue) to 5 µm (amber). Downloading the files whole would take about 17 GB. For a 3′ field the app reads about 95 MB.}

Making the movie. Each cutout is background-subtracted, then resampled with its own WCS onto a common north-up, east-left grid centred on the target. Visit stacks are inverse-variance weighted means with robust 3σ MAD clipping, so cosmic rays and satellite trails drop out. Every frame keeps its wavelength, which is read from the file's spectral WCS at the target position.

Display. Frames go to the browser as raw float32 arrays (a small binary format, see backend/app/wire.py) so stretch, contrast and colour map change instantly without another request.

Everything the backend fetches is cached on disk. demo_cache/ holds the search results, stacks, spectra, Gaia stars and NEOSSat matches for the presets, so they open instantly even if IRSA is slow.

Run it locally

Backend (Python 3.11+):

python -m venv .venv
.venv/Scripts/pip install -r backend/requirements.txt    # .venv/bin/pip on mac/linux
.venv/Scripts/python -m uvicorn app.main:app --app-dir backend --port 8000

Frontend, in another terminal:

cd frontend
npm install
npm run dev

Then open http://localhost:5173. Vite proxies /api to the backend.

Tests: .venv/Scripts/python -m pytest backend/tests

Deploy

Vercel runs the live site. vercel.json builds frontend/ as static files and runs the whole FastAPI app as one Python function through api/index.py, with requirements.txt at the repo root. Only /tmp is writable there, so the cache goes to /tmp/blink-cache. Vercel also refuses responses over 4.5 MB, so BLINK_MAX_BYTES=4000000 thins the single-exposure views to fit.

vercel deploy --prod

Docker works anywhere else (the image is set up for Hugging Face Spaces). The backend serves the site on port 7860:

docker build -t spherex-blink .
docker run -p 7860:7860 spherex-blink

Settings

Variable	Default
SPHEREX_RELEASES	dr1,qr3	Release preference order. New releases are picked up even if not listed.
BLINK_CACHE	./cache	Where search results, cutouts and stacks are cached.
BLINK_WORKERS	16	Parallel requests to IRSA.
BLINK_MAX_BYTES	0 (no cap)	Largest single-exposure response, for hosts with a size limit.

Keeping up with new data

SPHEREx publishes new quick-release images every week. Search results are cached for 24 hours, so new visits show up within a day without anything being redeployed. GET /api/search?ra=..&dec=..&fresh=true forces a fresh query, and python scripts/prebake.py rebuilds the demo cache.

API

Endpoint	Returns
GET /api/resolve?q=	Position, proper motion and parallax for a name or coordinate string
GET /api/targets	The preset targets
GET /api/search?ra&dec	Every exposure covering the position, grouped into visits
GET /api/frames?ra&dec&det&view&fov&ref	Image stack in the SPXF binary format
GET /api/spectrum?ra&dec&pmra&pmdec	Forced photometry in every exposure (JSON)
GET /api/motion?ra&dec&det&fov&pmra&pmdec	Per-visit centroids and fitted proper motion
GET /api/gaia?ra&dec&fov	Gaia DR3 stars in the field with G-band magnitudes
GET /api/neossat?ra&dec&fov	CADC CAOM-TAP observations for CSA NEOSSat

Limitations

• Spectrum fluxes are summed in a fixed 12″ aperture with no aperture correction. SPHEREx pixels are 6.15″, so a neighbouring star inside that radius adds its light (WISE 0855's blue end is mostly a background star).
• Very bright stars saturate. Their points are flagged and hidden in the spectrum by default.
• The catalogue track ignores parallax, which is under 0.6″ for every preset.
• Motion measured from three visits over one year is good to about 1″/yr. Only trust it for fast movers.
• On Vercel the first request after a quiet spell starts a fresh Python function, so it can take several seconds longer.

Data and credits

This uses data from SPHEREx (Spectro-Photometer for the History of the Universe, Epoch of Reionization and Ices Explorer), a NASA mission led by Caltech and managed by JPL, served by the NASA/IPAC Infrared Science Archive (IRSA). Names are resolved with CDS Sesame (SIMBAD, NED, VizieR). Gaia DR3 stars come from the ESA Gaia archive, and NEOSSat observations from the Canadian Astronomy Data Centre. The idea comes straight from WiseView and the Backyard Worlds volunteers.

Built with Astropy, NumPy, SciPy, PyVO, FastAPI, React and Vite. More detail on the data in docs/data.md.