WebP shrink-on-load rounding error produces "extract_area: bad extract area"

[asilvas]

Appears there is some sort of rounding error going on. As I validated the inputs and found that if I change the top position from 1 to 0, it works.

Code to reproduce (sharp 0.15.0):

var sharp = require('sharp');

var image = sharp('./test/files/bad-crop.jpg');

image
  .toFormat('webp') // if i don't do this step, it won't fail
  .toBuffer(function(err, outputBuffer, info) {
    console.log('info:', info); // reports 1080x607
    sharp(outputBuffer)
     .ignoreAspectRatio()
     .resize(233, 131)
     .extract({ top: 1, left: 0, width: 233, height: 130 }) // notice, height 130 + top of 1 = height of 131 - VALID
     .toFormat('jpeg')
     .toFile('result.jpg')
     ;

  });

Results in:

Unhandled rejection Error: extract_area: bad extract area

    at Error (native)

It only happens with this image: https://img1.wsimg.com/isteam/stock/5518eebf9a5d704c916de1acd4fe355e2c14463b39e4852697e1ec62859b5f5a?download

[lovell]

Hi Aaron, v0.15.0 had a small rounding error fixed as part of #452. Are you able to test with v0.15.1?

[asilvas]

Using v0.15.1 I get:

Unhandled rejection Error: enum 'VipsOperationBoolean' has no member 'null', should be one of: and, or, eor, lshift, rshift
enum 'VipsOperationBoolean' has no member 'null', should be one of: and, or, eor, lshift, rshift
extract_area: bad extract area

    at Error (native)

This is not a warning or "noise" as indicated in #509 -- it is failing.

[lovell]

Thanks for confirming this fails with v0.15.1 also (PR #511 on master should have fixed the contents of the error message). I'll make time to investigate further over the next couple of days.

[lovell]

I've created https://github.com/jcupitt/libvips/pull/489 for this.

[lovell]

Commit a5bd68e recalculates scaling factors after any use of shrink-on-load, as discussed at https://github.com/jcupitt/libvips/pull/489

[asilvas]

Seems this bug is impacting more images than I thought. Any idea when you think this patch would be made available? Thanks @lovell !

[lovell]

@asilvas This will be in v0.16.0, which will ship with libvips v8.3.3 (v8.3.2 fixes a buffer overflow but introduces a performance regression - see https://github.com/jcupitt/libvips/blob/master/ChangeLog#L36-L42).

[lovell]

v0.16.0 is now available via npm.