Optional inputs / generative tasks

[brunosan]

This is low priority for v0, but I think it is worth assessing level of effort.

Would it make sense to make all inputs optional?

I understand we currently provide as an input everything, always: the image (with masked out patches), location, time, instrument, ...

I want to understand if we can drop some of those inputs, specifically the image one. These are some applications where this can unlock a lot of value:

• If we don't provide an image it would be extremely valuable to make a finetuned application to generate the expected image. In essence creating a "semantically compressed" model of Earth. To the degree that this works, it could provide e.g. cloudfree RGB images, synthetic SAR, extremely fast approximation of a basemap in device...
• If we don't provide a location we can still try to infer semantics. This is useful to leverage images without locations information. We can also finetune to predict locations, which can help locate, or fix locations of images.

I can see implementing this making the MAE 100% of the image some of the times, and/or masking other inputs (and adding a flag). I expect the loses much greater in those cases, but it can hopefully force the model to learn absolute anchors in the model, instead on relative to the available patches on the image.

Does this make sense?