Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

C-RADIOv4 (Tech Report)

Mike Ranzinger, Greg Heinrich, Collin McCarthy, Jan Kautz, Andrew Tao, Bryan Catanzaro, Pavlo Molchanov

TL;DR

C-RADIOv4 extends agglomerative foundation modeling by distilling multiple strong teachers—SigLIP2, DINOv3, and SAM3—into a unified student while enabling robust, resolution-agnostic vision capabilities. Key innovations include shift-equivariant loss and MESA-based EMA alignment to suppress non-semantic noise, DAMP weight perturbations for robustness, and a balanced, angle-based summary loss that accounts for inter-teacher dispersion. The method supports any-resolution operation, reintroduces ViTDet-mode for high-efficiency, and can replace SAM3’s vision encoder, achieving competitive performance at reduced parameter counts (SO400M) and demonstrating strong zero-shot and k-NN scaling. The work provides practical pathways for deploying versatile, open-license vision foundations in dense and open-vocabulary tasks, with broad potential impact in academia and industry.

Abstract

By leveraging multi-teacher distillation, agglomerative vision backbones provide a unified student model that retains and improves the distinct capabilities of multiple teachers. In this tech report, we describe the most recent release of the C-RADIO family of models, C-RADIOv4, which builds upon AM-RADIO/RADIOv2.5 in design, offering strong improvements on key downstream tasks at the same computational complexity. We release -SO400M (412M params), and -H (631M) model variants, both trained with an updated set of teachers: SigLIP2, DINOv3, and SAM3. In addition to improvements on core metrics and new capabilities from imitating SAM3, the C-RADIOv4 model family further improves any-resolution support, brings back the ViTDet option for drastically enhanced efficiency at high-resolution, and comes with a permissive license.

C-RADIOv4 (Tech Report)

TL;DR

C-RADIOv4 extends agglomerative foundation modeling by distilling multiple strong teachers—SigLIP2, DINOv3, and SAM3—into a unified student while enabling robust, resolution-agnostic vision capabilities. Key innovations include shift-equivariant loss and MESA-based EMA alignment to suppress non-semantic noise, DAMP weight perturbations for robustness, and a balanced, angle-based summary loss that accounts for inter-teacher dispersion. The method supports any-resolution operation, reintroduces ViTDet-mode for high-efficiency, and can replace SAM3’s vision encoder, achieving competitive performance at reduced parameter counts (SO400M) and demonstrating strong zero-shot and k-NN scaling. The work provides practical pathways for deploying versatile, open-license vision foundations in dense and open-vocabulary tasks, with broad potential impact in academia and industry.

Abstract

By leveraging multi-teacher distillation, agglomerative vision backbones provide a unified student model that retains and improves the distinct capabilities of multiple teachers. In this tech report, we describe the most recent release of the C-RADIO family of models, C-RADIOv4, which builds upon AM-RADIO/RADIOv2.5 in design, offering strong improvements on key downstream tasks at the same computational complexity. We release -SO400M (412M params), and -H (631M) model variants, both trained with an updated set of teachers: SigLIP2, DINOv3, and SAM3. In addition to improvements on core metrics and new capabilities from imitating SAM3, the C-RADIOv4 model family further improves any-resolution support, brings back the ViTDet option for drastically enhanced efficiency at high-resolution, and comes with a permissive license.
Paper Structure (14 sections, 3 equations, 9 figures, 5 tables)

This paper contains 14 sections, 3 equations, 9 figures, 5 tables.

Table of Contents

  1. Description
  2. Method Updates
  3. Updated Teachers
  4. Stochastic Resolutions
  5. Shift Equivariance
  6. Shift Equivariant Loss
  7. Shift Equivariant MESA
  8. DAMP
  9. Balanced Summary Loss
  10. Results
  11. Metrics
  12. SAM3
  13. The curious case of "person"
  14. Conclusion

Figures (9)

  • Figure 1: PCA feature visualization, comparing C-RADIOv3-H vs C-RADIOv4-H. Object boundaries are now significantly cleaner.
  • Figure 2: Visualization of DINOv3 and C-RADIOv4 (adapter) predictions. Notice the out-of-place speckles produced by DINOv3. Column 1: Input Image Column 2: DINOv3 PCA visualization. Column 3: C-RADIOv4 adapter PCA visualization. Column 4: Error heatmap between the student adapter prediction and the DINOv3 teacher.
  • Figure 3: ImageNet-1K zero-shot accuracy as a function of input resolution.
  • Figure 4: ImageNet-1K kNN accuracy as a function of input resolution.
  • Figure 5: Latency analysis on A100 for both versions of C-RADIOv4, with and without ViTDet of a specified window size. The latency difference between ViTDet mode with window size 8 and 16 is negligible.
  • ...and 4 more figures