Neural Network Encoder for .gif is extreme bad for certain situation

[byzod]

For gifs that have high saturation, many colors or small area of bright pixels, Neural Network can hardly make it

I just found a practical example when I was about to share a game recording to friend

Original video: https://www.reddit.com/r/noita/comments/k735yn/is_that_a_new_dragon_ball_attack_from_anonymous/

Original recording project: bug1.zip

Output with Neural Network (sampling 1, use global color table):

Output with Neural Network (sampling 1, don't use global color table):

With global color table, the result overall color is dim (in situation like this, it's basically white&black), but no visible banding

Without global color table, the color looks better, but there're heavy banding problem

Thus, Neural Network is unable to handle gif like this, there's no way to encode it correctly currently

[NickeManarin]

The issue here is that the color quantizer is only using the parts of the image that changes between frames to calculate the colors.

I need to execute the quantizer before cropping/removing unchanged frames.
This will reduce the color quality for gifs with many colors, but it will improve the color consistency of the bar,