Yah new sutff, a vision transformer CLipV

[torzdf]

They also seem to have a ''State-of-the-art face alignment and face parsing model''

That's already in hand, but it's going to take a while. Most likely there will be an iterative FAN update first whilst I build a training set.

[Ryzen1988]

very cool to hear that, should you ever be in need of high quality/high res faces sets for development or stuff i be more that happy to contribute.
Other than that I'm looking forward to these improvements.

[Ryzen1988]

Having trained a couple of different clip setups up and to around 200000 its, it has become clear that the clipv_vit-l-14 models really stand head and shoulder above the other ones, it's crazy almost every setup using one of the 2 vit-l models gives crazy fast and crazy good results.
Only just a short spurt of training with encoder frozen and when you unfreeze it the magic starts happening fast.
Will post some results in the weekend but this has been a real hero (and last model i picked to try out )

It is crazy heavy on Vram, a setup with the normal vit-L encoder with mixed precision i can just fit on my 4090 with output res of 384 and batch size of 6.
The vit-L-336 big brother with output of 512px just fills up the 48gb vram with batch size 4 as if it was nothing to speak off
But very interesting results, am doing a split / half shared FC model and it really looks crazy good.
keep ya guys updated but if you can fit it, these are really worth checking out

[MaxHunter]

So, would you recommend the 336 for a 512 output?

Though I've thought about it, I ultimately haven't tried these yet because I felt they wouldn't be very good for 512 output.

[Ryzen1988]

the clipv L encoder is certainly very suitable for 448-512px output.
The farl B i would personally use for 256-384px.
The clipv L - 334px is probably good for 512 and above but the encoder itself is already very massive so vram is a big issue, so whatever comes behind it as model needs to be skinny.