As a developer, the most common and most frustrating false positives in image moderation are with very small images. The threshold for “very small” is up for debate but I think 8×8 is a good starting point.
It’s very hard to get something innapropriate in a single very small image. The main counter argument seems to be that we can composite small images into an inappropriate image. However, mind that anyone can do that with 1×1 Frames quite easily. It’s probably even easier.
Small images are everywhere in games. We need them for icons, solid colors, small UI textures, etc. Having to enlarge them to get them past the filter is a waste of GPU RAM and overall leads to more unnecessary work from resizing the images with bilinear filtering (which we should also be able to turn off btw). This is rather negligible on PC and console but means less battery life and potentially a noticeable performance hit on mobile.
In short, the current policy of automatically rejecting very small images does very little for moderation and is a huge annoyance for developers. Attracting developers from other platforms to Roblox seems to be a goal that Roblox has. An outside developer would likely find the current policy on small images laughable given their expectations of what a game dev platform is.
Inappropriate pixel art can still be made with something like 8x8 grids. Even a 3x3 grid can have inappropriate symbols (ex: Color pixels (1,1), (2,2), (3,2), and (1,3)). I would prefer them to not be rejected outright for their size, rather than automatically approved.
Take any 3 letter insult (for example “ass”), take the ASCII values for the 3 letters, use those values for the RGB value of your pixel and you got yourself a very insulting pixel.
Infact, you could upload entire books this way, now we just need a way to actually read the RGB-values of individual pixels…