So I'm using alpha cutoff transparency for foliage, and mipmaps were always a nuisance. Simple averaging can make the model disappear at a distance if the alpha channel isn't dense enough.
Experimenting with max alpha (1st row) instead of average alpha (2nd row) for mipmap generation, seems to preserve the overall shape better (3rd vs 4th rows).
@lisyarus *takes notes*
@lisyarus I've had very nice results with hashed alpha (although that is a bit noisy)
@scherzog You mean smth like dithered alpha?
@lisyarus Not quite, it's pseudorandom, unlike dither: https://research.nvidia.com/labs/rtr/publication/wyman2017hashed/
@lisyarus But still consistent over time unlike pure stochastic alpha testing
@lisyarus It seems like this could be done better by trying to preserve highest possible frequencies. Maybe try flipping between min/max alpha on each column?
Maybe downsample in frequency domain?
Or brute-force/genetic algorithm optimizing for number of blades after thresholding.