Export with resizing, extreme risk of wron colors

[NotMyFault]

I made a tiny test document to show one big principle issue with Affinity's apprach for resizing and resampling (during export).

In case you use the move tool and shift the pixel layer in 0.1 steps in x and/or y direction, the colors will shift dramatically due to resampling artifacts.

The source document is simply downscaled by factors 2, and contains a high frequency pattern of red/green/blue/white pixels. By downsizing, all colors average out to grey. But if you misalign the layer position, resampling will create absurd color artifacts.

For more craziness, slightly reduce the layer size and the again change starting positions.

Of course this document is specifically carfted to demonstrate the effect, never the less we get frequent complaints from users in real-word situaions like https://forum.affinity.serif.com/index.php?/topic/200893-losing-fine-detail-of-polygons-when-zoomed-out-and-in-scaled-export/#comment-1187203

A better way would be to provide a result which preservs the global color balance (grey) and creates only local color seams.

1. merge visible,
2. then resize the result and resample from the temporary "merge visible" layer.

Resample with downsizing.afphoto

Edited March 15, 2024 by NotMyFault

[Old Bruce]

That is a very interesting document. I tried to use the flood fill tool on it and couldn't get the colour of the pixel layer to change.

[NotMyFault]

1 hour ago, Old Bruce said:

That is a very interesting document. I tried to use the flood fill tool on it and couldn't get the colour of the pixel layer to change.

This should be absolutely no problem. But you need to set tolerance to 100%, and deactivate continuouse. Otherwise you may hit only 1px or 1 row/colum.

[NotMyFault]

After further testing it is a bit more complex.

When exporting, the resample method has a big impact.

bilinear (default):

if the layer is resized between 50% and 100% at least in one axis or both, the issue occurs as described.

In case both layers are resized below 50%, the issue vanishes.

In case using bicubic or lanczos, issue vanishes.

Using layer rotation and/or skew is hit and miss, depending on angles.

It seems there is a specific performance optimization for (default) bilinear resampling, but restricted to scaling of *2 resp /2.

This is consistent with older findings when scaling pixel layers and getting unexplainable antialiasing effects at layer edges, depending on scaling factors and aspect ratio of the layer.

[Ldina]

@NotMyFault I’m not  near my computer at present, but I was wondering if this color shifting happens when you move in 1 pixel increments with Force Pixel Alignment active (and move by whole pixels disabled). Thanks.

[NotMyFault]

3 minutes ago, Ldina said:

@NotMyFault I’m not  near my computer at present, but I was wondering if this color shifting happens when you move in 1 pixel increments with Force Pixel Alignment active (and move by whole pixels disabled). Thanks.

No, the document is crafted specifically so this happens only with fractional positions

[Ldina]

4 minutes ago, NotMyFault said:

No, the document is crafted specifically so this happens only with fractional positions

Thanks! 

[Lee D]

@NotMyFault So I can log this with the developers with as much information as possible, including your further testing. Can you confirm the steps taken to create the source document as well.

[NotMyFault]

thanks for following up.

To see the source, just use transform panel and increase layer size by factor 2 to see the layer in original form (100 zoom size / DPI matching document DPI).

it is a simple pattern of pixels in either R, G, B (created by a procedural texture filter, but you could do the same by bitmap fill, pattern layer). 
no additional tricks like alpha, blend range, blend gamma etc. just a plain fully opaque pixel layer.