Try flow matching ! #22
Description
https://www.youtube.com/watch?v=5ZSwYogAxYg
https://www.youtube.com/watch?v=7NNxK3CqaDk
Seems extremely promising and simple enough to make this project functional !
I think we should go towards a pre processing step that does not take random blocks all the time but a fixed block.
One idea to make inpaiting from this flow matching would be to subdivide t by the number of blocks (t / (16 * 16 * 16)) and then instead of smapling from noise to have something new, we sample from (noise - blocks already placed) + blocks placed and set t to t - (number of block placed)
This shit might work