Qwen image edit mask vs Krea inpaint vs GrabCut

For a qwen image edit mask job, Qwen darkened the background best on one header: only 4.2 percent of the masked pixels stayed light, against 18.5 percent for Krea inpaint and 0.3 percent for GrabCut. Qwen painted the sky, but it left a light fringe on the edge. This is one run, so treat it as a reading, not a ranking.

The setup was the same for all three. The header is the one used in the SEO post, and the mask is the GrabCut background mask.

source:   header 06, light
mask:     GrabCut background mask, 56.3 % of the frame
qwen:     seed 7, 25 steps, output 1344x768
check:    outside the mask grown by 40 px, difference 0

How long each one took

The Qwen job spent most of its life waiting. GPU time was about 90 s (load plus sampling), against about 95 s for the Krea inpaint run. GrabCut is a pixel operation, and the source gives it no timing, so it is left out of the table.

The queue wait depended on other lanes using the card, so it says little about the model. A separate Krea batch log lists header 06 dark at 49.05 s of job time and 50.7 s of wall time, but that is a different run. Do not set it against the Qwen total.

What the pixels say

Outside the mask, nothing moved: the Chebyshev difference was 0 against the light source, the same as Krea. Inside the mask, the three outputs separate.

GrabCut is darkest by the numbers, but darkest is not the goal. The mean colour shows the difference: navy for Qwen, grey-blue for Krea, near black for GrabCut.

Where each one wins

Qwen paints a dark navy gradient, faint stars, dim night clouds and a warm beam that fades into the sky. GrabCut only darkens pixels, so the beam keeps a hard edge. Krea sits in between: lighter, and less clean.

Qwen’s halo shows around the magnifier, the browser window and the floor line. The source suggests tightening the mask edge, by eroding it 2-4 px or lowering the blend setting, but that fix was not retested here. So whether it removes the fringe is unknown.

What this does not show

This is one header, one seed, one run. Nothing here says Qwen would beat GrabCut on another image, and no human verdict was collected. The pixel counts are measured, but “cleaner” is a reading of numbers. For other model comparisons on the same box, see the logo build with Opus.

Verdict

Use Qwen when the sky is the point, and GrabCut when the edge is. Run Qwen, check the boundary, and expect to fix the fringe by hand until the erode step is tested. Krea inpaint left the most light behind in this run.

Check the run yourself: read the header 06 proof