AstroBlink & AIsaac

Fast visual culling for astrophotography sessions on macOS — now with AI-powered session analysis.

Blink, mark, stack — triage your astro XISF, FITS subs in seconds. GPU-stretched viewer with LightspeedStacker, QuickLook, SmartCull quality scoring, and the fastest keyboard workflow on macOS. NEW: AIsaac, your built-in AI astrophotography assistant.

AstroBlink lets you blink through hundreds of FITS and XISF sub-exposures in seconds, mark the bad ones, and move them out of the way — without ever permanently deleting a single file. Inspired by PixInsight's Blink, built from the ground up for Apple Silicon.

Super lightweight (~5 MB app size), yet high performance — Metal GPU compute for real-time auto-stretch, star detection, and eccentricity analysis. No Electron, no bloat. Pure native Swift + Metal, engineered for Apple Silicon with zero-copy GPU buffers and concurrent decode pipelines. Loads 500+ subs in seconds on M-series Macs.

Nice side effect: Finally you have a native XISF and FITS Quicklook for macOS. (press Spacebar shows a quick preview of the image). Was missing that for long! :-)

---

What's New in v5.28.0 (Build 87)

• Auto-Rotate Pipeline Overhaul — Full rework after an empirical survey of 6,210 local frames exposed multiple failure modes on multi-setup, multi-session targets. The rotator-delta fallback (which covered ~80% of non-WCS frames) turned out to be unreliable because cross-session recalibrations invalidate naïve rotator math — it was silently overriding every other signal. Removed.
• New Pixel Fingerprint Signal — A 32×32 spatial signature computed from the raw decoded buffer (< 1 ms per frame, 1024 bytes, session-scoped). Works as the last-resort orientation detector when headers are silent (20.8% of frames — mostly RASA), and as an independent veto against bad WCS plate-solve transforms (catches 90°-off false solutions).
• Unified Reference Between Header + WCS Pipelines — Previously the two paths picked different reference frames of the same target; WCS-aligned frames and PIERSIDE-flipped frames rendered in different orientations. Both pipelines now use the same WCS-median reference.
• Per-Setup Orientation Grouping — Each physical rig (telescope + focal length + camera) has its own internally-consistent orientation group. Cross-setup frames render in setup-local orientation — a limitation we hit when there's no plate-solve data on one side.

Previously in v5.27.0 (Build 86)

• Transform-vs-Header Conflict Fix — When a frame had clear header evidence of a meridian flip but the star-matching aligner landed on a spurious near-identity match (rotation-invariant fields), the transform overrode the headers and the frame rendered un-flipped. Now the transform's rotation is cross-checked against header signals; transforms that disagree get discarded in favour of a clean 180° flip.
• Centroid-Mirror Flip Detection — 4th signal in the auto-rotate OR chain: star-distribution centroid mapped to its 180° mirror position; fires when headers are silent but the star field is asymmetric enough to disambiguate.

Previously in v5.26.3 (Build 85)

• Overlay mini-map pink tint fix — Bypassed CoreImage, rendered via direct Metal pipeline, CGImage built with noneSkipFirst so alpha byte is ignored. Mini-map colors now match main viewer for both OSC (post-debayer RGB) and mono frames.
• Filter + exposure label — Overlay shows Ha — 300s / L — 0.5s in mixed-size attributed string (whole seconds / sub-second formatting).

Previously in v5.26.1 (Build 83)

• Floating Image Overlay — Top-left overlay pinned to the viewer showing the big filter letter, capture time + observing-night date, a 5%-size mini-map of the full image, and your current 1-3 star rating. Stays in viewport space while you zoom and pan. Default on; toggle with ⌘⇧O or View → Show Image Overlay. Auto-on in Blind Curation.
• Live Viewport Indicator — A white dashed rectangle on the mini-map tracks the currently-visible sub-region in real time while you zoom and pan. Hides when the whole image fits the viewer.
• Mini-map thumbnail via CoreImage — Fixes a SIGSEGV in AGX when the display texture uses a GPU-private tile-swizzled layout.

Previously in v5.26.0 (Build 82)

• Stage 1.5 Severe-FWHM False-Positive Fix (Algorithm v23) — Systematic 10-finding code review of the quality pipeline followed by empirical validation against 4540 user-rated frames from the Frame History DB. The Stage 1.5 session-sanity check's single-flag "severe FWHM" demote path was removed after the data showed it uniquely fired on frames where FWHM was the only issue at ~34% precision (65 false positives per 33 true catches). Net +32 frames correctly classified on the curated set; genuinely bad frames continue to fail multiple metrics and are caught by the 2-flag rule.
• Stage 4 Rescue Preserves Session-Sanity Reasons — When the FWHM-sanity pass lifts a z-score-trash frame back to borderline, the original session-sanity and historical-baseline reasons now carry over. Rescued borderline frames show "REVIEW — " in the recommendation label and tooltips, giving Autopilot and manual culling clearer context.
• Uncertain-Tier Reasoning Coherence — Frames flipped to the uncertain tier after a rescue now display "Small group — low confidence" instead of stale rescue narrative.
• No More Silent Frame Drops — A frame that couldn't produce any z-score (e.g. the only measured frame in its group) previously vanished silently from the result. Now surfaces as an uncertain breakdown with explicit "No comparable frames in group" reasoning.
• Full Analysis Document — wiki/quality-pipeline-review-2026-04-18.md captures the methodology, confusion matrix (566 FP / 169 FN at entry), per-finding empirical numbers, and decision rationale — including three findings explicitly rejected as empirically net-negative. Re-review-ready for future contributors.

Previously in v5.25.2 (Build 80)

• Mixed-Sensor Scoring Fix (Algorithm v22) — Sessions with multiple cameras (e.g. ASI6200MM full-frame + ASI2600MC APS-C on the same telescope) now correctly separate quality scoring by sensor resolution.
• Mixed-Dimension Warning — Detects sessions with different resolutions and shows an informative alert per sensor group.
• Auto-Rotate Dimension Guard — Meridian flip rotation skipped for frames with different sensor dimensions than the reference.
• 1-Star Garbage UX — 1-star rating displays as outline star (☆); pressing 1 auto-marks for pre-delete.

---

Rosette Nebula in full color with AIsaac's quality breakdown — per-filter analysis, recommendations, and actionable next steps.

---

Side-by-side comparison with star eccentricity overlay. AIsaac explains what to look for and guides your culling decisions.

---

AIsaac — Your AI Astrophotography Assistant

Named after Isaac Newton — astronomer, physicist, and the father of reflecting telescopes. Newton didn't just observe the stars, he built the tools to understand them. AIsaac carries that spirit: an AI that doesn't just look at your data, but understands it, explains it, and helps you make better decisions.

AIsaac knows your equipment, remembers your imaging history, understands light pollution at your location, and speaks your language. He's a knowledgeable friend who happens to know everything about astrophotography — with a sense of humor.

We'd love to hear from you! AIsaac is brand new and we're actively improving him based on real-world feedback. If you try it, let us know what works, what doesn't, and what you'd like to see next:

• GitHub Issues
• AstroBin

---

Highlights

• AIsaac AI Assistant — built-in astrophotography AI powered by Claude. Quality summaries, smart culling suggestions, filter advice, imaging plans, voice input, equipment memory. Free Sonnet tier included; bring your own API key for Opus.
• Target Catalog Browser — browse 533+ deep-sky objects with real-time visibility charts, weather overlay, FOV simulation for your equipment, and filter gap analysis. Plan your next imaging session without leaving the app.
• SmartCull Quality Engine — 5-stage pipeline (garbage detection → session sanity → z-score ranking → rescue rules → FWHM sanity) auto-classifies ~99% of frames. Target-aware scoring — 533+ deep-sky target database adjusts metric weights by object type (galaxies prioritize resolution, nebulae prioritize SNR, IFN keeps every photon). FL-aware MAD floor prevents penalizing insignificant differences. Severity-dependent trailing detection with orientation consensus (industry first). Atmospheric attenuation detection (cloud/dew/fog). Cross-group session sanity. Self-calibrating per-setup baseline.
• LightspeedStacker — GPU warp+accumulate stacking, 10-20x faster than CPU. Hash-based triangle matching, Lanczos-3 interpolation, min/max pixel rejection. Color Combine for mono filter palettes (SHO, HOO, LRGB).
• GPU Post-Processing — real-time stretch, sharpening, contrast, dark level, color saturation, bilateral denoise, Richardson-Lucy deconvolution, gradient removal, Wiener deconvolution, structure enhancement.
• Metal GPU Rendering — PixInsight-compatible STF auto-stretch, zero-copy Apple Silicon buffers, <32ms navigation on cache hit.
• Compare with Best (C key) — side-by-side synchronized zoom/pan with star eccentricity overlay and PA direction arrows.
• Culling Autopilot — one-click auto-marking (Conservative/Balanced/Aggressive) with integration impact preview. Convergence guard warns when quality spread is too tight for further culling. Session Spread stats show per-metric distribution.
• VLM Check — AI Visual Anomaly Detection (ALPHA) — experimental feature that generates tiled mosaic wallpapers from your session and sends them to Claude Vision for visual inspection. Includes deviation maps, computational center-anomaly detection, and click-to-mark tiles. Current LLM vision models have not yet demonstrated sufficient accuracy for reliable detection of instrumental artifacts (ice, frost, optical defects) in astronomical sub-exposures. Available for experimentation — results should not be relied upon for culling decisions. Cancellable. Supports highlighted selection (uses all highlighted frames regardless of mark status) or all unmarked frames.
• Blink Playback & Video Export — play/stop button with adjustable delay (0.1–2s) cycles through visible images endlessly. Export as animated GIF (size-constrained) or HEVC .mov with scale, loops, and crop-to-zoom. Multi-select support.
• SSWEIGHT & PSFSignalWeight Export — writes PixInsight-compatible SSWEIGHT and PSFSWGHT keywords into FITS/XISF headers. GPU PSF fitting (Metal Gauss-Newton) for accurate flux. Batch keyword deletion via Batch Rename.
• Blind Curation Mode (⌘⇧B) — Rate frames 1-3★ without seeing algorithm scores. Your ratings train the scoring model — uploaded anonymously to improve quality thresholds across setups. Star shapes are the primary visual signal; the algorithm handles SNR, gradients, and decentering that you can't judge zoomed in. 1★ carries highest weight (may detect sensor artifacts like ice/frost invisible to metrics).
• User Confidence Rating — Press 1/2/3 to rate frames with 1-3 yellow stars (same key toggles off). Persisted in Frame History DB. Filter with rating:1, rating:2, rating:3.
• Quality Feedback (A key) — Tell the algorithm whether you agree with its quality tier. Press A to cycle Agree → Disagree → Partly → Clear. FB column right next to Q with thumbs-up / thumbs-down / sideways-pointing-hand icons (green / red / orange). Context menu submenu for mouse users. Persisted in Frame History DB and — because the DB is keyed on a content-based fileHash and iCloud-synced — the same feedback automatically reappears when you open the session on a different Mac. Uploaded anonymously (per filter/exposure group) to the community baseline so thresholds can be tuned from real agreement rates.
• PixInsight Bridge — PI script launches AstroBlink from PixInsight, imports CSV with quality scores, and prepares WBPP file lists. Available via PI Update Repository.
• GPU PSF Fitting — Metal compute kernels: circular (3 params) and elliptical (5 params: A, σx, σy, θ, B) Gaussian PSF fitting per star. Elliptical fit derives eccentricity and PA analytically. CPU FWHM fallback when GPU fit quality is poor.
• Native FITS/XISF QuickLook — Finder thumbnails and spacebar previews with debayer support for color cameras.
• Bortle Sky Quality (VIIRS 2024) — real satellite-measured light pollution for every frame. Fractional Bortle (e.g. B4.8) from NOAA VIIRS 2024 annual composite via Supabase lookup. Computed from SITELAT/SITELONG FITS headers. Offline fallback via embedded Falchi 2015 atlas.
• Frame History Database — persistent SQLite tracking all per-frame quality metrics across sessions. Cross-session scoring, global frame IDs, iCloud backup, archive NAS scanning.
• Temperature vs HFR Metrics — scatter plot correlating ambient temperature with focus quality (HFR). Per-filter colored dots, 1°C-binned rolling average trend line. Night picker for single-night detail or all-nights longterm view.
• Rain Forecast — hourly precipitation probability bars in the Target Catalog weather section, alongside cloud cover.
• Setup Management — rename, merge, delete, and fix focal length for equipment setups. Clean up duplicate entries from plate-solve variations.
• In-App Messaging — server-driven announcements, feedback collection, email signup without app updates. Rich actions, user targeting, feature unlocking.
• iCloud Settings Sync — all preferences and calibration data sync across Macs. Frame History DB with automatic sync and conflict detection.
• Zoom Controls — Cmd+0 fit to view, Cmd+1 actual pixels (100%), Cmd+2 double (200%), Cmd+/- 25% step zoom. Option+drag to pan. True pixel zoom % overlay.

For the full version-by-version changelog, see CHANGELOG.md.

---

Performance

Up to 5x faster session loading on local SSD. Up to 8x faster on network volumes (NAS/10GbE).

AstroBlinkV2 was rewritten for full hardware utilization on Apple Silicon. On a Mac Studio M3 Ultra with 300 FITS files (~100 MB each), total session load time dropped from minutes to under 45 seconds.

What changed	Before	After	Gain
FITS decode	1 file at a time	Up to 6 concurrent	4x throughput
Memory per decode	116 MB copy	Zero-copy GPU buffer	-116 MB/image
Downsampling (50 MP)	30–150 ms (CPU)	< 1 ms (GPU)	100x faster
Header reading (300 files)	~9 s	~1.5 s	6x faster
STF statistics	~50 ms	~17 ms	3x faster
Prefetch pattern	Batch 4, wait all	Sliding window	50% less stall
NAS file transfer	Single stream	4 parallel streams	3–4x faster

End-to-end (300 files, local SSD):

Phase	v0.9.7	v2.0.0+
Header reading	~9 s	~1.5 s
First image display	~250 ms	~170 ms
Full session prefetch	Minutes	~30–45 s
Navigation (warm cache)	< 32 ms	< 32 ms

---

Why AstroBlinkV2?

After a night of imaging you might have 200-600 sub-exposures. Some have clouds, tracking errors, satellite trails, or planes. You need to find and remove them before stacking. AstroBlinkV2 makes this fast:

1. Open your session folder (Cmd+O) — images load instantly with metadata parsed from filenames and headers
2. Blink through frames — arrow keys with key repeat let you scan frames like a flip-book
3. Mark the bad ones — hit Space on anything that looks wrong (clouds, trails, blur)
4. Hide and skip — press H to hide marked frames from the list, K to skip them during navigation
5. Pre-delete — Cmd+Backspace moves all marked files to a PRE-DELETE subfolder — nothing is ever permanently deleted
6. Undo if needed — full undo stack lets you restore any pre-delete operation (Cmd+Z)
7. Review your session — Session Overview shows per-filter integration times and generates a shareable Fact Sheet
8. Stack — select your best subs and hit LightspeedStacker for an instant stacked preview

---

Complete Feature List

Image Viewing & Rendering

• Metal GPU rendering — 50-megapixel images display in milliseconds on Apple Silicon
• Auto STF stretch — PixInsight-compatible Screen Transfer Function makes raw linear data visible
• Lock STF (S key) — freeze exact c0/mb stretch params from current image for brightness comparison
• Apply All — bake current stretch + post-processing into all cached previews for instant navigation
• Adjustable stretch strength — slider from 0% (linear) to 100% (maximum stretch)
• GPU post-processing — real-time sharpening (unsharp mask), contrast (S-curve), and dark level sliders
• Doubled slider ranges — Stretch 0–100%, Sharpening -4/+4, Contrast -2/+2, Dark Level 0–1.0
• Zoom & pan — click-drag zoom (Photoshop-style), trackpad pinch, +/- keys, scroll to pan
• Double-click to reset zoom to fit-to-view
• Persistent settings — all sliders, toggles, and column layout remembered across sessions

LightspeedStacker — GPU Stacking

• LightspeedStacker — GPU warp kernel, hash-based triangle matching, parallel star detection, full-res centroid refinement, two-pass least-squares alignment. ~15s for 16 frames.
• Select 3+ images and stack with one click — no plate solving required
• Triangle pattern matching for scale-invariant star alignment
• Affine transform alignment (rotation + translation + scale)
• Min/max pixel rejection — automatically removes satellite trails and hot pixels
• Lanczos-3 interpolation — optional sharper kernel for large dithers
• GPU bin2x pre-processing for ~4x faster stacking
• Live blue star crosses showing detected stars during processing
• Full result window with Photoshop-style zoom and all adjustment sliders
• GPU Metal compute kernel for instant slider response in result window
• Save as PNG with smart filename (object_date_filters_camera.png)
• Same-target validation — warns if you accidentally select images of different objects
• Color Combine — mono filter stacks to RGB (SHO, HOO, HSO, LRGB, HaRGB, Custom presets)

OSC Debayer

• Automatic Bayer pattern detection (RGGB, GRBG, GBRG, BGGR) from FITS/XISF headers
• Toggle on/off (D key) — debayer OFF (default) for fastest caching, ON for color preview
• GPU-accelerated bilinear interpolation Metal compute kernel
• Debayer indicator only visible when session contains OSC images

Night Mode

• Red-on-black UI — preserves dark-adapted vision at the telescope
• Press N — toggle night mode on/off, affects all UI elements including file list, status bar, and overlays

Search & Filter

• Spotlight-style search — real-time filtering in the toolbar, reduces file list as you type
• Plain text search — searches across all columns (filename, object, filter, camera, etc.)
• Column syntax — filter:Ha, file:Veil, type:LIGHT, fwhm:>4, stars:<500, exp:300
• Quality filter presets — q:trash, q:excellent, q:borderline, trail:>0.5
• Column aliases — short forms like fil, obj, cam work as column prefixes
• Numeric operators — >, <, >=, <=, = for FWHM, HFR, stars, exp, gain, etc.
• Mark/Unmark filtered — batch checkmark all search results, then move or delete

Blink Workflow & File Operations

• Space — mark/unmark images for pre-deletion (single or multi-select)
• K — skip over already-marked images during navigation
• H — cycle view: all files → hide marked → show only marked → all
• Cmd+Backspace — move all marked files to a PRE-DELETE subfolder (never permanent deletion)
• Cmd+M — move checkmarked files to any folder (with "Create New Folder" dialog)
• Full undo stack — Cmd+Z undoes both PRE-DELETE and Cmd+M moves, unlimited depth
• Multi-select — Shift/Cmd+click in the file list, then Space to mark all selected at once
• Arrow keys stop at boundaries (no wrap-around)
• Page Up/Home and Page Down/End for jump to first/last image

Computed Star Metrics (HFR & FWHM)

• GPU-accelerated star detection — during session loading, every frame is analyzed for stars using a Metal compute kernel on the GPU (~3-5ms per image)
• HFR & FWHM measurement — Half-Flux Radius and Full Width at Half Maximum are computed via Gaussian fitting for the brightest unsaturated stars in each frame
• Automatic quality scoring — computed metrics feed into the SmartCull quality engine for multi-stage frame ranking
• Works without NINA — even if your capture software doesn't provide HFR/FWHM, AstroBlinkV2 computes them from the actual image data
• Per-group source consistency — when mixing images with and without capture-software HFR, quality scoring uses a single consistent measurement method per group to ensure fair comparison

SmartCull Quality Engine

• 5-stage quality pipeline — garbage detection, session-wide sanity check, z-score ranking, pattern-based rescue rules, and FWHM sanity check. Handles ~99% of quality decisions automatically. Uncertain tier (blue ?) for small groups with ambiguous scores
• Orientation consensus trailing detection — industry first: detects tracking errors by analyzing whether star elongation directions agree (tracking error) or are random (optical aberration)
• Severity-dependent trailing penalties — mild narrowband trailing stays reduced (~0.3×), severe trailing escalates toward full luminance penalty (1.0×). Absolute ceiling catches severe trailing regardless of filter when consensus confirms tracking error
• Atmospheric attenuation detection — star count drop + SNR drop with normal FWHM = cloud, dew, or fog. Catches signal loss that z-scores alone can't detect
• Focal-length-adaptive thresholds — automatically adjusts eccentricity tolerance based on FOCALLEN from headers (short FL = more tolerance)
• Stage 3 rescue rules — frames with good FWHM + acceptable noise rescued from trash; star count dips recognized as transient events
• Quality reasoning ("Why?") — hover any quality icon for a human-readable explanation of the scoring decision
• Culling Autopilot — one-click auto-marking with Conservative/Balanced/Aggressive modes, each showing impact before applying
• Per-setup self-calibration — learns your equipment's quality baseline over time; after 30+ frames, calibrated frames are locked as KEEP
• SSWEIGHT export — writes PixInsight-compatible weight keyword (0-100) into FITS/XISF headers for WBPP weighted integration

Target Catalog Browser

• 533+ deep-sky objects — galaxies, nebulae, star clusters, planetary nebulae, supernova remnants, and more. Searchable and filterable by type, constellation, magnitude, and size.
• Real-time visibility charts — altitude/azimuth curves for your location showing when each target is observable, with twilight and moon indicators.
• Weather overlay — live cloud cover and transparency forecast integrated into the visibility timeline, so you can plan around the weather.
• FOV simulation — frame your target with your actual equipment setup (telescope + camera). See exactly how the object fits your sensor before imaging.
• Filter gap analysis — shows which filters you've already captured for each target and which still need integration time, based on your Frame History database.
• Data sources — target coordinates, magnitudes, and angular sizes from OpenNGC (CC-BY-SA-4.0) enriched with hand-curated filter recommendations and astrophotography metadata. Weather from 7Timer (seeing/transparency) and Open-Meteo (hourly cloud/wind/temperature). Sky survey thumbnails from NASA/STScI DSS (public domain).

Metadata & Session Overview

• NINA filename parsing — automatically extracts target, filter, exposure, gain, temperature, HFR, star count, and more
• FITS/XISF header reading — pulls metadata directly from file headers (filter, exposure, camera, telescope, mount, coordinates, pier side, etc.)
• Header Inspector (I key) — floating window with all FITS/XISF keywords, search filtering, highlighted important keywords, scroll position preserved, multi-row selection with Cmd+C copy and "Copy All" button
• Session Overview — per-object/filter/exposure breakdown with total integration time
• Quality Overview — per-filter noise, background, and SNR statistics with color-coded bars
• Interactive quality help — click ? for beginner-friendly explanation with real-world examples and rules of thumb
• Fact Sheet generator — one click copies a ready-to-paste summary with hashtags for Astrobin, Instagram, or forums
• AutoRotate — WCS plate-solve based visual alignment. Every frame of a target is pixel-locked to a single reference using the FITS/XISF WCS data (CD matrix + CRPIX + CRVAL) via direct matrix algebra — takes microseconds per frame, is mathematically exact, and works for any filter, exposure, night, rotator angle, or pointing offset. Smart median-CRVAL reference selection keeps outliers from becoming the baseline. Falls back to a synthetic rotator-based transform for the rare frames without plate-solve data. Works with ASIAir, NINA, and any other capture software that writes WCS keywords. Blink mode shows stars pixel-locked across frames — tracking drift, focus changes, and trailing pop out as the only things moving.
• Astronomical observing night — sessions spanning midnight are treated as one night. Quality scoring and grouping use the evening date, not the calendar date

File List & Sorting

• 20+ sortable columns — click any column header to sort, drag to reorder
• Columns include: #, Filter, Q (quality), SNR, FWHM, HFR, Night (observing night), Time, Object, Filename, Type, Camera, Exposure, Temps, Gain, Size, Stars, Subfolder, and more
• Quality tooltips — hover over Q column for score explanations or why a score is missing
• Right-click context menu — copy filename, file path, or full path
• Smart folder scanning — opens root images only when present, scans subfolders when root is empty, automatically excludes calibration frames (DARK/FLAT/BIAS)
• Individual file selection — select specific files instead of entire folders
• File size column with human-readable formatting (MB/GB)

Format Support

Format	Compression	Library
XISF	Uncompressed, LZ4, LZ4HC, zlib, zstd, ByteShuffle	libxisf
FITS	Uncompressed, fpack (Rice, GZIP)	cfitsio

Network Volumes

• Images from NAS/SMB shares are automatically cached locally for fast browsing
• Stop/continue caching at any time with inline controls
• 4 parallel network streams for maximum throughput
• Cache is cleaned up automatically on quit

QuickLook Extensions

• Thumbnail provider — FITS/XISF thumbnails in Finder
• Preview provider — full-size FITS/XISF preview in QuickLook (press Space in Finder)
• OSC debayer support — color camera previews show debayered color

---

Screenshots

macOS — AstroBlink & AIsaac

Color Image with AI Quality Analysis:

Compare View with Star Eccentricity Overlay:

Quality Summary with AIsaac Chat:

Header Inspector with Quality Metrics:

Classic Views

Session Overview with Header Inspector:

Night Mode — red-on-black for dark-adapted vision:

iOS — AstroFileViewer

M42 Orion (OSC color, debayered)	NGC 6960 Veil (mono)	FITS/XISF Headers

iPad — M42 Orion Nebula with Stretch & Debayer controls:

---

Keyboard Shortcuts

Key	Action
← →	Previous / next image
Page Up/Home	Jump to first image
Page Down/End	Jump to last image
+ -	Zoom in / out
Space	Toggle pre-delete mark (single or multi-select)
Cmd+Backspace	Move marked files to PRE-DELETE folder
Cmd+M	Move marked files to a chosen folder
Cmd+Z	Undo last move operation
S	Toggle Lock STF (freeze stretch params)
K	Toggle skip-marked during navigation
H	Cycle view: all → hide marked → only marked → all
I	Toggle FITS/XISF header inspector
D	Toggle OSC debayer (when Bayer images detected)
N	Toggle night mode (red-on-black)
1 2 3	Set confidence rating (1-3 stars, same key clears)
C	Compare with best frame
U	Unmark all
Cmd+ Cmd-	Increase / decrease font size
Cmd+0	Reset font size
Cmd+O	Open folder or select files
ESC	Stop blink playback
Double-click	Reset zoom to fit-to-view

---

AstroFileViewer — iOS Companion App

AstroFileViewer is a companion iOS/iPadOS app for viewing FITS and XISF astrophotography files on your iPhone or iPad.

Features

• Open FITS and XISF files directly from the Files app, iCloud Drive, or any document provider
• Auto STF stretch — same PixInsight-compatible algorithm as the macOS app
• Adjustable stretch strength — slider from 0% (fully linear) to 100%
• Dark point slider — raises the black point to clip faint background noise (same as macOS AstroBlink)
• GPU bilateral denoise — edge-preserving noise reduction that smooths background while keeping stars sharp
• Auto gradient correction — adjustable removal of linear light pollution gradients (0–300%)
• File history & swipe navigation — cached files with swipe left/right to navigate between recently viewed images
• Auto-rotate — landscape images automatically rotated to fill the screen in portrait mode
• OSC debayer — automatic Bayer pattern detection with GPU-accelerated bilinear interpolation
• Sharpening — real-time unsharp mask with adjustable strength
• FITS/XISF header viewer — browse all metadata keywords with priority sorting
• Save to Photos — export stretched images as JPEG to your Photo Library
• Persistent settings — slider positions and gradient toggle saved across launches
• Universal app — optimized for both iPhone and iPad
• Automatic bin2 display — large sensor images (e.g. ZWO ASI6200MM at 9576×6388) are automatically downscaled for smooth display

iOS Screenshots

iPhone — M42	iPhone — Veil	iPad — M42

The iOS app source code is included in this repository under AstroFileViewer-iOS/.

---

Requirements

macOS (AstroBlinkV2)

• macOS 14 Sonoma or later
• Apple Silicon required (M1/M2/M3/M4) — engineered for unified memory architecture
• Metal-capable GPU

iOS (AstroFileViewer)

• iOS 16.4 or later
• iPhone or iPad with Metal support

---

Installation

Download Release (macOS)

1. Download the latest release from the Releases page
2. Unzip and drag AstroBlinkV2.app to your Applications folder
3. Double-click to launch — the app is signed and notarized by Apple

AstroFileViewer (iOS)

Download AstroFileViewer from the Apple App Store (link coming soon).

Build from Source

1. Clone the repository:

   git clone https://github.com/joergs-git/AstroBlinkV2.git
   cd AstroBlinkV2

2. macOS app: Open AstroTriage.xcodeproj in Xcode 15+ and build (Cmd+R)

3. iOS app: Open AstroFileViewer-iOS/AstroFileViewer.xcodeproj in Xcode 15+ and build for your device

The project includes vendored C/C++ libraries (libxisf, cfitsio) as a local Swift Package — no external dependencies to install.

---

How It Works

AstroBlinkV2 decodes FITS and XISF files using cfitsio and libxisf through a C bridge, renders them on the GPU via Metal compute shaders with a PixInsight-compatible STF auto-stretch, and displays them in an MTKView. Navigation is keyboard-first with full key repeat support. Metadata is extracted from both filenames (NINA token patterns) and file headers, merged with header values taking priority.

The workflow is non-destructive by design: marking a file only sets a flag in memory, and the "pre-delete" action physically moves files to a dedicated subfolder — never to Trash, never permanently deleted. A full undo stack allows you to reverse any pre-delete operation.

LightspeedStacker uses triangle pattern matching on the brightest stars in each frame, computes affine transforms for alignment, and mean-combines aligned frames — all without external plate solving. The warp+accumulate step runs on the GPU via a Metal compute kernel with hash-based triangle lookup for near-instant matching.

Floating windows (Session Overview, Header Inspector, Quick Stack Result) stay above the main AstroBlinkV2 window while working but go behind other apps when you switch away.

---

Supported NINA Filename Tokens

AstroBlinkV2 parses the standard NINA filename pattern:

2026-03-06_IC1848_23-54-58_RASA_ZWO ASI6200MM Pro_LIGHT_H_300.00s_#0016__bin1x1_gain100_O50_T-10.00c__FWHM_4.15_FOCT_4.46.xisf

Extracted tokens: date, target, time, telescope, camera, frame type, filter, exposure, frame number, binning, gain, offset, sensor temp, FWHM, focuser temp, HFR, star count.

---

FAQ

Why is LightspeedStacker so fast — and is it accurate?

LightspeedStacker achieves 10-20x speedup over CPU stacking while maintaining near-identical alignment quality:

1. GPU warp+accumulate — The biggest bottleneck in stacking is warping each frame to match the reference. LightspeedStacker offloads this to a Metal compute kernel that runs thousands of GPU threads in parallel.

2. Parallel star detection + full-res centroid refinement — All frames are analyzed simultaneously via Swift TaskGroup. After coarse detection on subsampled data, each star position is refined using a 9×9 weighted centroid on the full binned-resolution image — the same approach used by professional astrometry tools (SExtractor, DAOPHOT).

3. Hash-based triangle matching — Quantizes triangle shape ratios into hash buckets for O(1) lookups — same 455 triangles from 50 stars, but matched ~100x faster than brute-force O(N²).

4. Two-pass least-squares refinement — The initial 3-point affine from triangle matching is refined using ALL inlier star correspondences (typically 20-40 pairs) via least-squares normal equations. A second pass with a tighter threshold (4px) converges to sub-pixel alignment accuracy.

5. GPU restretch — The result window sliders use a Metal compute kernel (restretch_float) for instant response (<16ms) instead of CPU processing.

Why don't the computed HFR/FWHM values match what NINA reports?

Different software measures HFR/FWHM differently. NINA measures during capture (often on a subframe), while AstroBlinkV2 measures post-capture on the full saved frame with its own star detection algorithm. The absolute numbers will differ, but the relative ranking (which frames are better/worse) should be very consistent. AstroBlinkV2's quality scoring always uses the same measurement method across all frames in a group to ensure fair comparison.

Is the quality scoring scientifically accurate?

The quality scoring is designed for triage — quickly identifying the best and worst frames for stacking decisions. It uses robust statistical methods (median HFR, FWHM, noise estimation, z-scores within groups) but is not intended to replace scientific photometry tools like PixInsight or AstroImageJ. Think of it as a fast, automated version of the visual blinking you'd do in PixInsight's Blink tool, enhanced with quantitative metrics.

Is this scientific stacking?

No. LightspeedStacker is designed for visual preview only — a quick "what does my data look like stacked?" without leaving the triage app. It is not a replacement for dedicated stacking software like PixInsight, Siril, or APP. Specifically:

• No astrometric calibration — alignment uses triangle pattern matching, not plate solving with a star catalog
• No calibration frames — no dark, flat, or bias subtraction
• Min/max rejection only — removes single-frame outliers (satellite trails, hot pixels) but no sigma clipping or advanced rejection

LightspeedStacker does support optional Lanczos-3 interpolation and min/max pixel rejection. The stacked result is meant to give you a quick visual impression of your session's potential — not a final image.

---

Author

joergsflow

• Astrobin Gallery
• Instagram
• GitHub
• joergsflow@gmail.com
• AstroBlink Homepage

---

License

This project is licensed under the GPLv3 License — see LICENSE for details.

libxisf is licensed under GPLv3. cfitsio is licensed under the NASA Open Source Agreement.

---

Clear skies and happy imaging!