emerging
C-RADIOv4 (Tech Report)
· NVIDIA · 2026-01 · arXiv:2601.17237
Why it matters
The current agglomerative frontier, and a useful signal about consensus: its teachers are now DINOv3, SigLIP 2 and SAM 3, which is the field converging on the same three encoders that the generative-latent work also reaches for.
What this paper does
read: full textBefore this
AM-RADIO and RADIOv2.5 established agglomerative distillation with recipe fixes for resolution mode switching, teacher imbalance and excess token count, but the student still absorbed fixed-pattern teacher artifacts such as SigLIP2 border holes, SAM ViTDet window artifacts and DINOv3 noise patches, and summary-token matching stayed imbalanced across teachers with different angular dispersions.
The problem
The report distills a newer, stronger teacher set while removing the remaining position- and teacher-specific artifacts, and improves any-resolution support and high-resolution inference efficiency.
The idea
Keep the RADIO agglomerative recipe but upgrade the teachers to the current best-in-class encoders and add shift-equivariant losses plus perturbation-based regularization, so the student stops memorizing each teacher's fixed-pattern noise and generalizes better across resolutions.
How it works
The supervision is multi-teacher agglomerative distillation. Frozen SigLIP2-g-384, DINOv3-7B and SAM3 teachers distill into two student sizes, C-RADIOv4-SO400M at 412M parameters and C-RADIOv4-H at 631M parameters. Training uses stochastic low- and high-resolution partitions spanning 128px to 1152px, a shift-equivariant loss that randomly and independently shifts the crop seen by the student and by each teacher to strip out position-specific noise, a shift-equivariant MESA EMA regularizer, DAMP multiplicative weight perturbation for robustness, and a balanced summary loss that normalizes angular distance by each teacher's own cone radius. FeatSharp upsampling and mosaic augmentation adapt SigLIP2 and SAM3's native resolutions to the shared schedule, and a ViTDet windowed-attention mode is offered for efficient high-resolution inference. The frozen resulting student is evaluated purely as a discriminative encoder for segmentation, correspondence, classification, and as a drop-in replacement for SAM3's own vision tower in instance segmentation. The report contains no mention of diffusion models or of C-RADIO features feeding a generative image model anywhere in the text, so like its predecessor this is representation learning only as far as the paper's own framing goes.
Evidence
C-RADIOv4-H reaches 83.09 percent ImageNet-1k zero-shot accuracy against RADIOv2.5-H's 82.51 percent. ADE20k mIoU is 55.20 for C-RADIOv4-H against RADIOv2.5-H's 51.58 and against the DINOv3-7B teacher's own 55.9. NAVI correspondence rises to 63.44 from RADIOv2.5-H's 60.89, and SPair to 60.57 from 56.24. As a substitute for SAM3's own vision tower in instance segmentation, C-RADIOv4-H reaches 44.7 percent cgF1 against SAM3's own 54.1 percent.
Limitations
The paper concedes an uneven ability to replace SAM3's vision encoder, with larger performance gaps on out-of-domain data such as sports equipment and Wikipedia-common imagery compared to natural images, and states that closing this gap is an open research direction.
Why it matters
The teacher set converging on DINOv3, SigLIP2 and SAM3 is itself a signal of field-wide consensus on which encoders carry the most useful structure, the same three families, self-distillation, language-aligned contrastive pretraining and promptable segmentation, that generative-latent work such as RAE also draws its frozen encoders from.
Abstract, in the authors' own words
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.
Research line
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a emerging entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.