C-RADIOv4: Multi-Teacher Vision Transformer

• C-RADIOv4 is an agglomerative vision transformer backbone that unifies multiple teacher models to boost visual task performance.
• It employs a plain ViT architecture enhanced with per-teacher MLP adapters, shift-equivariant losses, and advanced regularization techniques for efficient high-resolution inference.
• The model supports continuous resolution training from 128 to 1152 px, achieving competitive segmentation and downstream metrics while maintaining constant computational complexity.

C-RADIOv4 is an agglomerative vision transformer backbone designed as a unified student model trained via multi-teacher distillation. Building upon the AM-RADIO and RADIOv2.5 frameworks, C-RADIOv4 aims to consolidate and enhance the distinct capacities of multiple state-of-the-art teacher models, offering strong performance improvements on vision tasks while maintaining constant computational complexity. The model introduces innovations in training with diverse teachers, loss formulation, any-resolution support, and high-resolution inference efficiency, and is released under a permissive license (Ranzinger et al., 24 Jan 2026).

1. Model Architecture and Design

C-RADIOv4 utilizes the “plain” Vision Transformer (ViT) backbone foundational to both AM-RADIO and RADIOv2.5, specifically a standard ViT with a single summarization token and per-teacher MLP adapters. Significant architectural changes from prior versions include:

• Teacher Set Update: The teacher set is revised to SigLIP2, DINOv3, and SAM3, with DFN CLIP removed.
• Full Multi-Resolution Training: The model samples all resolutions from discrete low- and high-resolution integer sets, eliminating the need for separate “low” vs “high” modes and abolishing the “mode-switching” pathologies observed in AM-RADIO.
• Shift-Equivariant Distillation Losses: To mitigate memorization of teacher-specific fixed-pattern noise, shift-equivariant variants are adopted for dense distillation.
• Balanced Summary-Token Loss: The summary-token loss is balanced so that no single teacher’s summary feature dominates the student representation.
• Regularization Enhancements: DAMP weight noise and MESA regularization are used in addition to standard AdamW for robustness and better generalization.

A specialized ViTDet mode is available, where most self-attention layers operate within non-overlapping W×W token windows, with global attention in sparse layers—substantially accelerating high-resolution inference. The vitdet_window_size flag configures this, and empirical results demonstrate 2–3× speedups at 1152 px without increasing FLOP count.

2. Multi-Teacher Distillation Framework

C-RADIOv4 employs a multi-teacher distillation setup integrating the following teacher models:

• SigLIP2-g-384: A 412M-parameter multilingual vision-language encoder providing strong text–image alignment and global summary/dense features.
• DINOv3-7B: A self-supervised dense representation model with robust semantic segmentation performance, generating high-quality dense features.
• SAM3 (Perception Encoder): A promptable segmentation teacher (ViT-L+ backbone) producing maskable features through various prompts.

Loss is computed as:

$$L_{\text{total}} = \sum_{t\in T} \left(\lambda_t^\text{feat} L_\text{spatial}^t + \lambda_t^\text{sum} L_\text{angle}^t \right) + \lambda_\text{mesa} L_\text{mesa} + \lambda_\text{damp} L_\text{damp}$$

where:

• $L_\text{spatial}$: Spatial (dense) MSE loss with shift-equivariance.
• $L_\text{mesa}$: Shift-equivariant regularization using EMA student outputs.
• $L_\text{angle}$: Balanced summary-token (“angular”) loss, normalized by each teacher's directional spread.
• $L_\text{damp}$: DAMP weight noise regularization.

Training Schedule and Key Techniques:

• Stochastic resolution sampling from six low‐resolution and four high‐resolution sets.
• FeatSharp upsampling for scaling SigLIP2 inputs, and mosaic augmentation for SAM3 at large sizes.
• 300K steps with AdamW and cosine learning rate decay (following RADIOv2.5).
• Shift-equivariant cropping and DAMP/MESA regularization throughout.

3. Capabilities and Improvements

Any-Resolution Support

Continuous sampling from low- and high-resolution bins enables stable performance across 128–1152 px, eliminating abrupt performance transitions seen in earlier models. Comparative segmentation metrics (ADE20k):

Model	512 px	1024 px	1536 px
DINOv3-7B	55.9	57.3†	57.8†
C-RADIOv4-H	55.20	57.02	57.72

C-RADIOv4-H matches the largest teacher’s quality within 0.1 percentage points at a substantial reduction in parameter counts.

New Capabilities from SAM3 Imitation

C-RADIOv4 directly supports replacing SAM3's Perception Encoder, providing promptable segmentation masks with identical decoders and memory stacks to SAM3. Notably, it empirically resolves a known SAM3 issue where "person" queries fail. Instance segmentation results on the SA-Co/Gold cgF1 benchmark:

Model	Avg cgF1
Human	72.8
SAM3	54.1
C-RADIOv4-H-VDT12	44.2
C-RADIOv4-H-G	44.7

4. Performance Evaluation

Downstream Metrics Overview

Key results for zero-shot, kNN, segmentation, and geometric evaluation are as follows:

Model	Params	Z-Shot	kNN	ADE20k	VOC	NAVI	SPair
RADIOv2.5-H	631 M	82.51	85.81	51.58	85.97	60.89	56.24
C-RADIOv4-H	631 M	83.09	86.59	55.20	87.24	63.44	60.57

On full Probe3D:

Model	Depth	Surf Norm	NAVI	SPair
C-RADIOv4-H	85.55	61.70	63.44	60.57

C-RADIOv4-H attains strong gains relative to predecessors, equaling or exceeding larger teacher models.

Computational Efficiency

Model Variant	Params	High-Res Latency (A100, ms/im, W=8/VDT)
SO400M	412 M	~10
SAM3 ViT-L+	632 M	~25

The ViTDet windowed-attention configuration achieves a 2–3× actual reduction in latency at 1152² px, with parameter and FLOP counts matching standard ViTs of equivalent size.

5. Licensing and Deployment

C-RADIOv4 is distributed under a commercially permissive (Apache 2.0–style) license, imposing no usage restrictions beyond standard patent and attribution requirements. Integration in both academic and commercial settings is allowed, and all associated resources—including code, pretrained models, and deployment examples—are accessible via the referenced GitHub and Hugging Face repositories (Ranzinger et al., 24 Jan 2026).

6. Model Variants and Public Resources

Two main variants are available:

Name	Parameter Count
SO400M	412 million
H	631 million

Both variants support multi-resolution training, ViTDet high-resolution inference, and full SAM3 prompt compatibility. Detailed instructions for backbone replacement in SAM3, together with public demos and checkpoints, are provided online.

---

C-RADIOv4 represents a mature integration of multi-teacher vision distillation, high-throughput ViT architectures, and robust regularization and training techniques, benchmarked across a comprehensive suite of visual understanding metrics (Ranzinger et al., 24 Jan 2026).