Bug? `--vae-tiling` produces very obvious seams/artifacts near the edge of each "tiles"

[stduhpf]

Case and point:

.\build\bin\Release\sd.exe -m ..\models\checkpoints\sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 20 --sampling-method euler -H 1024 -W 1024 --seed 42 -p "professional photo of a girl laying down on grass, fine details, 4k resolution" --vae-tiling

I can almost count all the 64 tiles by eye. I don't think that's the expected behavior.

For comparison, here is with normal vae on cpu:

[MGTRIDER]

Case and point:

.\build\bin\Release\sd.exe -m ..\models\checkpoints\sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 20 --sampling-method euler -H 1024 -W 1024 --seed 42 -p "professional photo of a girl laying down on grass, fine details, 4k resolution" --vae-tiling

I can almost count all the 64 tiles by eye. I don't think that's the expected behavior.

For comparison, here is with normal vae on cpu:

Yes, i can confirm that these seams also happen for me when using tiled vae.

[Green-Sky]

This has been an issue for a long time now, so I am currently investigating.
Here is a test image, which replaces each tile with a single uniform shade of gray. The lines are the midway points of the overlap.

You can clearly see seams, basically everywhere beside the outermost tiles. So only the outermost with the next seem to be merged (lerped) correctly.

edit: also I'm looking into using smootherstep for the lerp, might help after the seams are fixed.

[Green-Sky]

I am relatively confident that the current interpolation function blends like this:

so I am trying to make it looks like this:

[Green-Sky]

Hmmm, this still has issues...

edit: using smootherstep it looks even nicer, but here are still seams and I am not sure why.

[Green-Sky]

old merge	new merge
no tiling	lines at mid-overlap

new merge is using lerp(a, b, smootherstep(min(x ,y)))

I think this is good enough for a pr, I will open one soon if I don't have anymore insights into whats wrong or could be better.

[stduhpf]

I think this is good enough for a pr, I will open one soon if I don't have anymore insights into whats wrong or could be better.

Please do, maybe i could help figuring it out