Astro test frames

Raw frames kept for developing and validating the astro render mode. Measurements below come from the sensor data itself, not from a rendered preview.

Both current frames are a7R IIIa, 6024×4024, white level 16383, stated black level 512 per channel, RGGB with the pattern given as R,G],[G,B.

_DSC2589.ARW — wide field, vignetting

Download (24 MB)

The primary vignetting test frame. What the sensor data shows:

• Falloff of 0.68 centre to corner, about 0.55 stops. Smooth, monotonic and concentric, with the brightest cell landing within one cell of the geometric centre, so a radial fit centred on the sensor is a good starting model and the centre-offset parameters will have little to do.
• Nearly achromatic: corner/centre ratios are R 0.739, G 0.726 and 0.712, B 0.735. The colour cast of the vignette on this lens is only around 3%, so per-plane fitting will mostly confirm rather than correct. A faster or adapted lens should show far more.
• A planar sky gradient rides on top of the radial falloff: left 240 DN against right 215, and a 34 DN spread between the four corners, against a sky background of roughly 290 DN above black. This single frame therefore exercises both terms at once, and is good evidence that the radial and planar fits must be solved together rather than one after the other.
• Star cores clip on about 0.001% of photosites, which is enough to exercise per-channel highlight reconstruction.
• Green split gives a clean Poisson slope: variance rises from about 1010 at 125 DN to about 1320 at 500 DN, with a read floor near 30 DN. The G1 − G2 noise model works on real files.

Outstanding caveat: without a matching flat there is no way to separate how much of the 0.55 stops is lens falloff and how much is sky. That is what makes a genuine flat the most valuable frame still to come.

_DSC2596.ARW — lunar, the case the fit must refuse

Download (24 MB)

• Moon centred in an empty sky. Peak 2680 DN against a white level of 16383, so about 2.6 stops of headroom and nothing clipped anywhere in the frame. Being underexposed is the right error for a test frame: no data is lost, and exposure lift is exercised.
• Far-corner sky sits at a median of 512.0 DN raw against a stated black level of 512, mean 511.19. The camera's declared black level is accurate to within a DN, so black level subtraction needs no correction term for this body.
• An annulus-median fit reads this frame as roughly 29 stops of falloff, because beyond half radius the sky is black. This is the clean argument for an automatic guard rather than relying on the toggle: real lens vignetting is smooth, monotonic and essentially never exceeds two and a half to three stops, so a measured profile steeper than that, or one shaped as a plateau followed by a cliff, is a subject rather than optics. The correction should decline to apply itself and say why.

Still wanted

• A genuine flat, or an evenly lit wall or blue sky at the same lens and aperture, to separate lens falloff from sky gradient.
• A landscape astro frame with a treeline or horizon, the known failure case for annulus statistics.
• The same four cases from the a6000, for an APS-C body with larger, older photosites and different black level and green split behaviour.
• Ideally a frame from a fast or adapted lens with falloff worse than the 0.55 stops above.