On using Eevee for compositing images (Patreon)

Ebu brought this point up when using Eevee for compositing images.

This is a good point and something you should be aware of if you're using Eevee as a compositor for some reason. You're going to want to make sure your render plane is as close to the camera's size as possible (I always start with scaling up a plane by 1920x1080 and scaling it down by 0.01 and then more to size it to the camera precisely by zooming in as close to one of the corners as possible).

Here's the ground truth version of frame 3538:

This is the render from my Eevee compositor setup:

This is the two images on top of each other, with the blend mode set to difference (in Photoshop cs6). It looks like there's no difference, however...

 This is the image after auto contrast. There is a difference, an introduction of a noise pattern and some color differences (the black splotches where colors approaching white are lost and that red area around the head of their appendage means some color was lost). Of course there would be a difference, it's an image mapped onto a plane in 3d space in a render engine. It won't be a perfect match!

 And like... yeah, it's such a small difference but I think I'll use the compositor instead. For more extreme stuff like the VHS style effect I did awhile where information loss and image degradation is the whole point, Eevee compositing is perfect. But I want results that are as exact as possible when mixing the images...

So, thanks to Ebu for getting me to investigate this further!

Edit: RavenWorks, a good friend of mine, suggested comparing the results after video compression. So I went ahead and did that!

I rendered both images as videos with the usual settings I use for mp4. H.264 with perceptually lossless compression, both 24 frames long. Same settings for both videos. Then, in the Blender VSE, I put one on top of the other with the difference blend mode. Once again, the difference is not noticeable without further processing.

Here's a much higher contrast version.

And here's the difference between the ground truth image and a frame of the video of the ground truth (so ground truth PNG vs. ground truth PNG after video compression) with auto contrast in photoshop:

And here's the difference between the render from the Eevee compositing setup and the render of that image after video compression. Looks pretty much the same:

 And also this is the difference between the two above images:

 So... really not a huge difference it seems? Most of the difference seems to be from the noise pattern the Eevee compositing introduced propagating into the video compression but it's like... not that extreme or anything. I'd say using Eevee as a compositor is probably fine.