To improve texture mapping quality when textures are shrinked we should see if mipmapping improves the quality enough to justify the upto 33% more memory usage, also:
Can the mipmap textures be generated on level-load?
Linear/Bilinear/Trilinear filter?
On-the-fly dithering?
Is blending between mipmap levels necessary?
Should there be an abstract Sampler interface (with compile-time implementation)?
To improve texture mapping quality when textures are shrinked we should see if mipmapping improves the quality enough to justify the upto 33% more memory usage, also: