search

alex-moon / vc

MIT License
22 stars 6 forks source link

16:9 aspect ratio #35

Closed alex-moon closed 3 years ago

alex-moon commented 3 years ago

Typical viewing ratio is 16:9 - the way to do this would be to VQGAN+CLIP at resolution 533 x 300 and then thumbnail down to 1280 x 720. This shouldn't be slower as it's the same number of pixels involved. Output will be higher res and therefore higher filesize.

alex-moon commented 3 years ago

This is failing in 3D Photo Inpainting, though it's not clear why. It's putting a large blue buffer at the top and, it would appear, cutting the bottom off.

alex-moon commented 3 years ago

This is stretching the image in VQGAN+CLIP. Again, not clear why. Worth going back to the Crowson colab to see if it does this.

alex-moon commented 3 years ago

So the stretching issue is due to the way cutouts is done. I think the NR one is just broken in fact. I've yoinked the one from pytti, but the result is images that are homogenous, which is not desirable. One thing we could do is weight the cutouts toward the centre. Worth a try. It should be as simple as replacing the call to rand with a call to normal.