The distinction is that textures are dense and have rigid topology (a grid) while a vertex mesh can be sparse, and has arbitrary topology. The distinction is extra clear in a voxel/SDF approach, where you see harsh grid boundaries because detail is undersampled, but you can't crank up the resolution without oversampling elsewhere. The rigid topology of textures can really hurt you.
You can even automatically generate the optimal UV allocation by balancing resulution-to-area distortion against polygon edge distortion and optimizing with a constrained least-squares solver.
Could you expand on the problem in sdf approaches? I’ve always thought or assumed that if there’s high frequency features in a signed distance function rep you could use gradient / derivative info to do adaptive sampling ?
I'm referring to a discrete SDF, where you have a 2D/3D grid storing distance samples, e.g. font textures, or mesh distance field textures [0]. You can use a tree of textures to get adaptive sampling, but a single texture gives you a fixed resolution grid.
