DINOv3: Self-supervised learning for vision at unprecedented scale

TL;DR

Meta releases DINOv3, a 7B-parameter ViT vision foundation model trained with self-supervised learning on 1.7B unlabeled images — 7× the model size and 12× the data of DINOv2. The headline claim is that a single frozen SSL backbone now matches or beats weakly-supervised models (SigLIP 2, Perception Encoder) on classification while drastically widening the gap on dense prediction (detection, segmentation, depth). Ships with a distilled family (ViT-B/L + ConvNeXt T/S/B/L), a satellite-imagery backbone trained on MAXAR, downstream evaluation heads, and a distillation pipeline, all under a commercial license. Already cited as the encoder-of-record across the wiki’s video/world-model line.

Key claims

DINOv3 is trained label-free on 1.7B images at 7B parameters — 12× the data and 7× the model size of DINOv2 [§Scalable and efficient visual modeling without fine-tuning].
A single frozen DINOv3 backbone outperforms specialized solutions on dense prediction tasks (object detection, semantic segmentation) for the first time among SSL models [§Intro].
DINOv3 matches or exceeds SigLIP 2 and Perception Encoder on many image classification benchmarks while widening the gap on dense prediction [§Unlocking high-impact applications with self-supervised learning].
Performance gains transfer to non-web modalities: a DINOv3 backbone trained on MAXAR satellite imagery reduces tree canopy height measurement error in a Kenya region from 4.1 m (DINOv2) to 1.2 m [§Intro].
A distilled family (ViT-B, ViT-L, ConvNeXt T/S/B/L) is released to span the inference-compute envelope; distilled DINOv3 variants outperform comparable CLIP-based models across the broad evaluation suite [§Scalable and efficient visual modeling without fine-tuning].
Backbone is evaluated across 15 visual tasks and >60 benchmarks; the distillation pipeline itself is open-sourced [§Scalable and efficient visual modeling without fine-tuning].
Because the backbone is frozen, a single forward pass can serve multiple downstream heads simultaneously — relevant for edge deployment [§Scalable and efficient visual modeling without fine-tuning].

Method

The post does not disclose the training algorithm in detail (the technical report is a separate artifact). What’s stated: DINOv3 builds on the DINO self-distillation algorithm with “novel refinements”, consumes only a fraction of prior compute, and produces “rich, dense features” — per-pixel float vectors that parse objects into parts and generalize across instances/categories. Lightweight adapters (a linear model on top of a few annotations) recover strong dense predictions; with a more sophisticated decoder on the frozen backbone, DINOv3 achieves SOTA on detection, semantic segmentation, and relative depth without backbone fine-tuning. Smaller variants are distilled from the ViT-7B teacher; ConvNeXt variants are produced by the same distillation procedure to target different compute envelopes.

Results

Quantitative numbers in the blog are sparse — the headline empirical claim is the SSL-vs-weakly-supervised reversal on dense prediction tasks with a frozen backbone, but no benchmark tables are reproduced. The one concrete number disclosed is the WRI deforestation use case: tree canopy height error in Kenya drops from 4.1 m (DINOv2) to 1.2 m (DINOv3 trained on satellite + aerial imagery), a ~3.4× reduction. The rest is announcement framing: “matches or exceeds SigLIP 2 and Perception Encoder on many image classification benchmarks”, “drastically widens the gap” on dense prediction, “state-of-the-art performance” on object detection / segmentation / relative depth under the frozen-backbone constraint. Full numbers live in the (unfiled here) technical report.

Why it’s interesting

DINOv3 is the upstream artifact that several already-filed papers in the wiki treat as a fixed point. V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning explicitly benchmarks against DINOv3 ViT-7B on Ego4D STA (V-JEPA 2.1 ViT-G beats it 7.71 vs 5.68 mAP despite being smaller) — positioning V-JEPA as the predictive-latent alternative to DINOv3’s contrastive/SSL recipe in the World Foundation Models cluster. OneVision-Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence uses DINOv3 as the strongest comparison encoder under attentive probing (OV-Encoder beats it by 8.1% Top-1 on Diving-48). Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis surfaces the counter-result that stronger DINOv3 encoders (B → H+) monotonically hurt REPA-style external alignment for generation — a sharp caveat to the “bigger encoder is better” framing. And UniTok: A Unified Tokenizer for Visual Generation and Understanding / Improved Baselines with Representation Autoencoders (RAEv2) build directly on frozen DINOv3 as the encoder half of a unified tokenizer. Filing the source artifact makes those downstream comparisons properly anchored instead of dangling references.