SVD weights release

[dsvilarkovic]

Hello,
Given that only Cosmos setup was released and that the whole strength of spatial maxpooling idea comes from SVD not having temporal encoding (relevant issue: #38), does that mean we can easily setup Gen3C to work with vanilla SVD weights, with only maxpooling configured? Or finetuning is indeed important and there is an SVD weights file that I missed?

Kind regards,
Dusan

[dsvilarkovic]

Another relevant issue: #22

[7yzx]

Any progress? I want to change cosmos1 to more powerful model(like wan2.2)

[dsvilarkovic]

Nope, didn't change a lot. But I provided in a forked repo a Multi-View support for Cosmos: https://github.com/dsvilarkovic/GEN3C/blob/main/README_MULTIVIEW.md, and from some of my small tests it works good enough for 3D reconstructions (4D will not work due to spatiotemporal compression). Look into git diff.

If you spend some time with Wan2.2, I think you can pull off similar spatial max-pooling and make it work.
Please share if you succeed.