Hi Michał! Sorry for the extremely late reply, I somehow completely overlooked your comment until now. That’s actually a great idea. Both parametric t-sne as well as UMAP would be good options, and I’ve actually also had good results in the past with training an additional “compression VAE” just for compressing existing embeddings to 2D (the nice thing here is that it’s also fully reversible, you can then go from 2D to the higher-D embedding, and from that back to actual data). I should try that at some point! :)