DINO3?

[vitacon]

Thanks for this project - CMNET2 is really a huge step forward from CMNET. =)

I've been tinkering with its code with Gemini because I use separate color files in my project and Gemini pointed out that while CMNET is build around DINO2, Meta has already released DINO3 which should be more accurate.

I suppose DINOv2FeatureV6_LocalAtten_s2_154000.pth is derived from DINO2 models but it was created by yyang81 for ColorMNet.

How difficult it would be to make CMNET3 based on Dino3? I suppose it would require a lot of model training..?

[dan64]

The upgrade of the model to DINOv3 (with probably SAM3 for better segmentation) it's something I'd like to do, but I don't have the hardware resources to undertake this training. I'll give it a try this winter; right now it's too hot to keep the GPU busy for days or maybe weeks...

[vitacon]

"I'll give it a try this winter; right now it's too hot to keep the GPU busy for days or maybe weeks..."

That sounds like a very sad joke. =} Running this on a home computer for weeks would be rather annoying. By the way, what VGA do you currently use?

The original ColorMNet team is not interested in updating their project? Or some other university team that could access a decent hardware?

[Elefant-Freeciv]

If you have training code, I could try to train the weights. I have 2 x 7900 XTX and a number of other cards. I dunno how long it would take though, so no guarantees.

Edit: Looks like it would be about 2 days to train. Based on training the original colormnet, with poor gpu utilization, due I think to insufficient CPU workers.