Supporting advanced metrics for image generation

Nobody has claimed this yet.

## 🚀 Feature

Proposal to include advanced metrics for image generation that helps to evaluate related models more completely. By completeness, I mean the following factors (as laid out in [here](https://arxiv.org/pdf/2103.01209.pdf)):

* Transparency and compositionality
* Disentanglement
* Image diversity

### Motivation

Image generation models often only use FID, IS, and KID for evaluating quality. While this has been helpful but it has shortcomings:

* FID, KID, IS tend to work better for class-conditional generation (especially where there's an alignment between classes during training the generative model and the classes on which the underlying Inception model was pre-trained)
* Therefore, these metrics might not present a complete picture of the evaluation, especially when the generation is unconditional or use free-from natural language queries (Stable Diffusion, for example).

### Pitch

Therefore, it might be helpful to approach the evaluation from many different aspects as laid out above. Here are some snapshots for each of those aspects:

#### Image diversity

#### Disentanglement

I am currently unsure how to best support these as far as API design is concerned.

### Alternatives

### Additional context

No files, tests, or entry points are named. Start with the proposal and linked paper; before implementation, clarify the API and concrete metric definitions for transparency and compositionality, disentanglement, and image diversity, then establish tests for the agreed behavior.

Written by the indexing model from the issue text.