FLOAT
EN FR

Block

Post

FLOAT 0.2.0 · beta of 15 November 2026

What it does
It is where the picture is finished: each workflow’s result is a layer, and every adjustment is a filter that can be changed or removed at any time.
What it shows
The final picture, the layers panel beside it, and a few sliders from 0 to 4 that start in the middle.
What can be set
Sky, Contrast and Saturation first; the default rendering is a sound starting point, and Export writes the final file, a 16-bit TIFF by default.

Post is the page where the results of the workflows are finished. It follows the layer model of image editors such as GIMP: the result of each workflow, or of any block, is a layer, with an opacity, a blend mode and a mask. Each adjustment is a filter on its layer, which can be switched off, changed or removed on its own at any time. Every workflow ends with an Output block, which brings its linear image to the screen with an automatic stretch; Post shows that stretch with a few sliders, from 0 to 4, and the filters go on top.

  1. Color
  2. Stars/background
  3. Post
  4. Report

1The Output block, and what Post receives

The Output block ends every workflow. It is automatic and cannot be cut. It takes the last image of the chain, the denoised stack if the workflow has Denoise, else the stack with its gradient removed, and writes it as a linear image in display colours with its sky taken off. If the workflow has Color, the white balance of that block is applied. It also writes the covered pixels and the statistics of the stretch, measured once on the whole image at full resolution: median, median absolute deviation and the 99.995th percentile of the luminance and of each channel.

Two repairs are made there. The periodic pattern of a colour mosaic is notched out of the image. The places where samples were clipped, the cores of the bright stars, take the colour of the light just around them, at the lowest brightness every channel allows. A combination of filters, like a single stack, is not cropped: the whole frame is shown and the user crops.

Post then renders that image. The default rendering, with every slider at 2, is a midtones transfer function from a black point at the median minus 2.8 median absolute deviations to a white point at the 99.995th percentile, the median sky brought to a display level of 0.25. With linked channels the stretch is made on the luminance, so that the colour balance of the image is kept; without, each channel is stretched on its own, which neutralises a colour cast.

2Layers

A layer is one of three kinds.

  • A workflow layer is the image of a workflow: its Output block’s image, or one of the images of its Stars/background block (the starless image, the stars alone, or the two recombined). It is rendered by the engine from its base and its sliders (section 4).
  • An image layer is the result of any block (a stack, a gradient-removed or a denoised image) sent from Inspect. It is stretched automatically and moved by the Sky, Contrast and Saturation sliders. Its pixels are never saved in the composition, only the name of the result.
  • A filter layer is one filter (section 5). It applies to the one pixel layer it is nested under, or to everything below it when it is a row of its own.

Each layer has an eye, an opacity from 0 to 100 % and a blend mode. The filters have them too. The modes are the following, a being what is below and b the layer.

Table 1.The blend modes of a layer or a filter.
ModeResult
Normalb
Lightenmax(a, b)
Screen1 − (1 − a)(1 − b)
Linear dodge (add)min(a + b, 1)
Multiplya × b
Soft light(1 − 2b) a² + 2ab

The mode is applied first, then the opacity and the mask mix the result with what is below: below + (result − below) × opacity × mask.

Stars and background

A workflow with a Stars/background block comes in as one layer with two images, each with its own filters: the starless image, stretched with the statistics of the whole image so that, recombined, it looks like the Output block’s, and the stars alone. The stars are stretched on their own, black at 0 and the white and midtone of the whole image, and go back on top in Screen, or in Linear dodge (add) or Lighten. Stretching them less is the star reduction: none is needed on the recombined image. The stars of the layer are a group: the real stars and, when the workflow has the Synthetic stars block, the synthetic ones beside them, each with its eye and its mask. Each side may also have a mask of its own, applied to its image after its filters and before the two are put together.

For a combination of filters made on starless images, the Combine layers block puts the stars of another workflow back on its image; the stars of a combination of stars are stretched with the statistics of the starless combination.

3Masks

Every layer and every filter may have a mask. A mask is made of parts, each added to what is there, subtracted from it or intersected with it. A mask shows as a red overlay on the image.

Table 2.The parts of a mask.
PartWhat it selectsSettings
H-alphaThe red excess of the image, the red minus the green (mostly Hα, with [N II] and [S II]), smoothed over 6 px, 0 at its median and 1 at its 97th percentile (measured on the whole field).None
Luminance rangeThe pixels whose luminance lies between two levels, with soft edges.From 40 %, To 100 %
Color rangeThe pixels near a hue, weighted by their saturation.Hue 0° – 359°, width 40° (10° – 120°)
Linear gradientA ramp across the image, from the top, bottom, left or right.Position 50 %, softness 40 % (1 % – 100 %)
BrushWhat is painted, with strokes that can be taken back one at a time.Size 60 px, hardness 50 %

Three settings act on the whole mask: Feather (0 to 60 px, blurs the edge), Density (0 to 100 %, the share of the layer that shows where the mask is white) and Invert. A brush added to an empty mask takes away from a full one: the layer stays shown everywhere and the first stroke hides.

The luminance and colour parts read the image they are on. A mask part is computed once for given settings and reused, so that adding a part to a mask computes that part only.

4The simplified sliders

A workflow layer is rendered from a base, a named rendering of the project or the suggestion, moved by its sliders. Each slider runs from 0 to 4 and its position 2 is the base.

Table 3.The sliders, and the setting each one moves, from position 0 to position 4, for the suggested base.
SliderGroupEndsSetting0 → 2 → 4
SkyNebula and backgrounddarker, brighterSky brightness0.15 → 0.25 → 0.37
ContrastNebula and backgroundsofter, strongerS-curve−0.35 → 0 → 0.5
Faint tonesNebula and backgrounddarker, brighterFaint tones−0.8 → 0 → 1.6
SaturationNebula and backgroundpaler, more vividSaturation0.6 → 1.2 → 1.85
ColoursNebula and backgroundfaded, trueKeep the colours0 → 0.5 → 1
H-alphaNebula and backgroundless, moreH-alpha−0.6 → 0 → 2
Noise smoothingNebula and backgroundless, moreDenoised share; faint area smoothing0.4 → 1 → 1; 0 → 0 → 1.6 px
StrengthStarssubtler, bolderStar strength0.25 → 0.6 → 1.6
Small starsStarsless, moreSmall stars1 → 0 → 0
Star sizeStarssmaller, largerStar size−3 → −1 → 1 px
Star colorStarswhiter, more colorfulStar color0.4 → 1.2 → 2.3

Faint tones raises or lowers the tones just above the sky without moving the sky: a generalised hyperbolic stretch[1] of the range above the sky, its black point and symmetry point at the sky, applied to the luminance with the colour ratios kept. Colours sets how much of the stretch keeps the colours: at 0 each channel goes through the same curve, so that the faint colours turn grey and the bright cores white; at 1 the luminance is stretched and the three channels scaled with it, which keeps the colours of the linear image.[2] The base is 0.5, so that nebulae keep their colour and the bright cores their structure. H-alpha adds the measured H-alpha emission to the red before the stretch, or takes it out; it is the red above a continuum predicted from the green (mostly Hα, with [N II] and [S II]), the red-to-green ratio of that continuum measured on the stars, as the continuum of a narrow-band Hα image is scaled on the stars.[3] Noise smoothing mixes the raw stack back into the denoised one below 2, and smooths the faint areas above, the bright ones keeping full resolution.

A slider that a layer cannot offer is greyed, and its reason is shown: the star sliders work on the stars of the Stars/background block; the stars layer holds the stars only; noise smoothing needs the Denoise block and works on the whole image; H-alpha is for colour cameras only, since in a palette combination H-alpha is its own channel and on a monochrome image there is no red over a continuum.

5Filters

A filter is a layer of its own kind. Dropped onto a pixel layer it is nested under it and acts on it alone; between two rows it acts on everything below. Each has an eye, an opacity, a blend mode and a mask. Switching it off, changing it or removing it leaves every other filter as it is.

Table 4.The filters of Post, and the method each follows.
FamilyFilterMethod
StretchGeneralised hyperbolic stretchPayne and Cranfield: a strong stretch around a chosen level, the shadows and highlights protected.[1] Factor ln(D + 1) 0 – 4, local intensity −5 – 15, symmetry point, shadow and highlight protection, black point.
Arcsinh stretchpixel = x′ · asinh(L′ · stretch) / (L′ · asinh(stretch)), the colours kept.[2] Stretch 0 – 40, black point 0 – 0.2.
Histogram transformationShadows, midtones and highlights, through the midtones transfer function (m − 1) x / ((2m − 1) x − m).
CurvesOne curve per channel (RGB/K, R, G, B, CIE L*, saturation, chroma, hue), 2 to 100 control points, Akima, natural cubic or linear interpolation.[4] Presets, the user’s own, and .acv files.
LevelsBlack point 0 – 0.3, white point 0.7 – 1, gamma 0.5 – 2.
ContrastLocal contrastA large-radius unsharp mask: strength 0 – 1.5, scale 5 – 120 px.
DehazeThe veil is measured as the background of a source extractor (meshes, 3σ clipping, mode, bicubic map)[5] and taken off with a haze model, the white kept.[6] Amount, scale 32 – 512 px.
Local histogram equalizationCLAHE on the CIE L* lightness.[7] Amount 0 – 1, contrast limit 1 – 8, kernel radius 16 – 512 px.
Highlight compressionA base and its detail by the guided filter;[8] the base is compressed above a knee, the detail kept. Strength, start 0.2 – 0.9, detail kept, scale 8 – 120 px.
ColorSaturationSaturation around the luminance, −100 to +150 %, by hue range, with a background protection that spares the noise of the sky.
Colour mixerHue, saturation and luminance of eight colour ranges (red, orange, yellow, green, aqua, blue, purple, magenta) in CIE L*c*h*. The grey background and the white star cores are left alone.
Green removalThe green held under a neutral made of red and blue; four protections, amount, preserve lightness.
Color balanceGains on red and blue (temperature) and on green (tint).
DetailSharpenUnsharp mask: strength 0 – 2, radius 0.5 – 4 px.
Multiscale sharpeningStarlet wavelet layers (B3 spline à trous) of 1, 2, 4, 8 and 16 px, each raised or lowered on its own, on the lightness or on each channel.[9]
Spot healingA spot replaced by a patch of its neighbourhood and blended in by Poisson editing.[10] The source is found automatically, 2.2 to 3.2 radii away, by the patch search of PatchMatch,[11] and can be moved.
NoiseNoise reductionA Gaussian smoothing of the luminance and of the colour.
Multiscale noise reductionSoft thresholding of the starlet layers,[9] each shrunk by k σ with σ estimated from the first layer; on the lightness, the chrominance or each channel.
Faint area smoothingSmooths the areas fainter than a level (0.05 – 0.8), the brighter ones untouched, so that the faint extensions can be raised without their grain.
CombineH-alpha to redThe engine’s H-alpha map added to the red channel.

6Versions

A project keeps several versions of the same edit, each with its own layers, filters, masks and its own Undo and Redo. A click on the version’s name opens the list and a double-click renames it. A name must be given and must be unique. New version makes a copy of the one shown, and a version can be deleted while another remains. A deleted version can be brought back with Ctrl+Z. Layers copied into a new version keep the same settings, so a version’s switch reuses what the other computed. The exported files carry the version’s name.

Nothing in Post renders in the interface thread. While a slider or the brush moves, a draft is shown (an eighth of the image on the CPU, a quarter on a graphics card); on release the quarter, then the half-resolution image of the screen; and, zoomed in, the visible region at full resolution, with a margin of 64 px so that the filters see their neighbours. A newer request replaces an older one.

7Export

Export writes the whole composition at full resolution, each layer rendered on the whole grid. The files go to the folder chosen and nowhere else; a folder inside the project’s results is refused. Each file is written under a temporary name and takes its final one in a single step: a failure leaves the earlier file as it was.

Table 5.The export formats. All are in sRGB, with the profile embedded where the format allows.
FormatFileNotes
TIFF 16-bitname_16bit.tifThe default. Zlib compression.
TIFF 32-bit floatname_32bit_float.tifZlib compression.
PNG 16-bitname.png
JPEGname.jpgQuality 93 by default (50 – 100), chroma not subsampled (4:4:4): the colour of a star is not smeared.
FITS 32-bit floatname.fitsWith the astrometric solution (WCS) and the processing history in its header, and the target, the dates and the total exposure when known.

The size is the full resolution (the drizzled grid) or a long side of 2048 or 4096 px, reduced by area averaging. A quarter turn or a vertical mirror is applied without any resampling; the solution in a FITS header follows the turn. The export can add a history file next to the image, each layer as a separate file, the masks, the night card, and the sky grid and object names drawn into the picture.

8Settings

Table 6.The settings of the Output block and of a mask.
SettingValueRangeWhat it does
StretchautoautoHow the linear image is brought to the screen: black just below the sky, the sky at a set brightness.
Linked channelsonon, offOne stretch for the three channels, which keeps the colour balance; off, each channel is stretched on its own.
Sky brightness0.250.02 – 0.6Display level of the median sky after the stretch.
Saturation1.20 – 3Colour saturation around the luminance, after the stretch.
Keep the colours0.50 – 1How much of the stretch keeps the colours.
Faint tones0−1 – 3Raises or lowers the faint tones above the sky.
Feather0 px0 – 60Blurs the edge of a mask.
Density100 %0 – 100The share of the layer that shows where the mask is white.

9Limits, and what was tried

  • Post rendering is slow on integrated AMD graphics. Tested on Windows 11 with an RTX 4060 Ti, an RTX 3070 and a laptop without an NVIDIA card. When the memory of the graphics card is full, the layer is rendered again on the CPU, and a note says so.
  • The Output stretch is automatic only. The other stretches are filters on a layer, not a setting of the Output block.
  • Layers pulled from Stacking leave the graph early. A layer taken from a block before Denoise has no noise slider, and one taken before Stars/background has no star sliders; each says why.
  • Without Color, the colours are those of the camera. Each channel’s sky is made grey and scaled to the same spread, so that the colour cast of the light pollution does not tint the preview; the gradient stays visible, since removing it is the Gradient block’s job.
  • Not in this version. Colour profiles other than sRGB, and the linear outputs.
  • The first Post’s arcsinh stretch around a symmetry point is hidden from the filter menu and kept only so that the compositions that hold it still open.

References

  1. Payne, D. & Cranfield, M., generalised hyperbolic stretch (GHSastro). ghsastro.co.uk
  2. Lupton, R. et al. 2004, Preparing Red-Green-Blue (RGB) Images from CCD Data, PASP 116, 133. arXiv:astro-ph/0312483
  3. James, P. A. et al. 2004, The H alpha Galaxy Survey I. The galaxy sample, H alpha narrow-band observations and star formation parameters for 334 galaxies, A&A 414, 23. arXiv:astro-ph/0311030
  4. Akima, H. 1970, A new method of interpolation and smooth curve fitting based on local procedures, J. ACM 17(4), 589. doi:10.1145/321607.321609
  5. Bertin, E. & Arnouts, S. 1996, SExtractor: Software for source extraction, A&AS 117, 393. doi:10.1051/aas:1996164
  6. He, K., Sun, J. & Tang, X. 2009, Single Image Haze Removal Using Dark Channel Prior, CVPR 2009 / IEEE TPAMI 33(12), 2011. doi:10.1109/TPAMI.2010.168
  7. Zuiderveld, K. 1994, Contrast Limited Adaptive Histogram Equalization, Graphics Gems IV, 474. doi:10.1016/B978-0-12-336156-1.50061-6
  8. He, K., Sun, J. & Tang, X. 2013, Guided Image Filtering, IEEE TPAMI 35(6), 1397. doi:10.1109/TPAMI.2012.213
  9. Starck, J.-L. & Murtagh, F. 2006, Astronomical Image and Data Analysis, 2nd ed., Springer. doi:10.1007/978-3-540-33025-7
  10. Pérez, P., Gangnet, M. & Blake, A. 2003, Poisson Image Editing, SIGGRAPH 2003, ACM Trans. Graph. 22(3), 313. doi:10.1145/882262.882269
  11. Barnes, C., Shechtman, E., Finkelstein, A. & Goldman, D. B. 2009, PatchMatch, ACM Trans. Graph. 28(3). doi:10.1145/1531326.1531330