17.09.2026, 17:59
Hello Selur,
I was finally able to add DinoV3 in CMNET2 (see : cmnet2) and so I updated also HAVC.
I attached the new RC1 to this post.
The main changes are:
1) new parameter DeepExBackbone, that can have only 2 values: 'dinov2' (default for compatibility), 'dinov3'
2) new weights to be installed for DinoV3:
dinov3-vitb16.zip -> unzip to vshavc\colormnet\weights
DINOv3FeatureV6_LocalAtten_p369412.pth -> download to vshavc\colormnet\weights
in the GUI, the new parameter DeepExBackbone must be enabled only if is selected the "Exemplar Model" CMNET2
I hope that you have time to include this version in Hybrid.
Thanks,
Dan
P.S.
The training of the model is still a work in progress, but the released model weights are already quite good, see the images in this folder for a fair comparison
I was finally able to add DinoV3 in CMNET2 (see : cmnet2) and so I updated also HAVC.
I attached the new RC1 to this post.
The main changes are:
1) new parameter DeepExBackbone, that can have only 2 values: 'dinov2' (default for compatibility), 'dinov3'
2) new weights to be installed for DinoV3:
dinov3-vitb16.zip -> unzip to vshavc\colormnet\weights
DINOv3FeatureV6_LocalAtten_p369412.pth -> download to vshavc\colormnet\weights
weights/
│ ├── DINOv2FeatureV6_LocalAtten_s2_154000.pth # ColorMNet pre-trained weights (DINOv2 backbone)
│ ├── DINOv3FeatureV6_LocalAtten_p369412.pth # Fine-tuned weights (DINOv3 backbone, recommended)
│ └── dinov3-vitb16/ # DINOv3 ViT-B/16 backbone (HuggingFace format)in the GUI, the new parameter DeepExBackbone must be enabled only if is selected the "Exemplar Model" CMNET2
I hope that you have time to include this version in Hybrid.
Thanks,
Dan
P.S.
The training of the model is still a work in progress, but the released model weights are already quite good, see the images in this folder for a fair comparison

