Dear all,

thanks again for the library. So far, I haven't found anything equivalent to exploit the conditional flow matching paradigm.

To learn the library, I started from the example on conditional generation. I quickly stumbled upon the fact, that the enclosed UnetModel assumes very strongly that a condition (if provided) is a class label. In my case, the condition is a single floating point number or a float tensor.

I sat down and adapted the UnetModel class accordingly. Everything seems to work there. I am currently moving towards testing an entire training loop. I am positing this here to hear, if

a) people are interested in this?
b) this is implemented already and I just didn't see it?
c) it would be interesting to hear from @atong01 or @kilianFatras if you'd be happy to receive a PR when I am done?

Some PRs look quite stale and I am unclear if you intend to maintain this library in the near future.

Best,
P