← Retour au blog
tech 8 June 2026

Playing with Vision Embeddings: A Deep Dive

Embeddings are the native language of neural networks, translating images into machine-understandable data. Let's explore how DINOv3 transforms these representations.

Article inspired by the original source
Playing with Vision Embeddings ↗ prestonbjensen.com

Introduction to Vision Embeddings

Embeddings are the native language of neural networks. They transform images and other complex data into lists of numbers that encapsulate semantic meanings. Yet, these numbers often remain opaque and difficult for humans to interpret. In this article, we explore how the DINOv3 ViT-S model uses embeddings to interpret images.

The DINOv3 ViT-S Model

DINOv3 ViT-S is a vision transformer model that learns to map raw pixels to a rich feature space without prior knowledge. It can compress any image into a single embedding, a list of 384 numbers. One interesting aspect of DINOv3 is that it was trained so that different crops and augmentations of an image have similar embeddings.

How DINOv3 Handles Images

Unlike other models that require prior knowledge, DINOv3 operates without understanding the language or descriptions of images. It focuses on pixel similarity and visual patterns to generate meaningful embeddings.

Generating Images from Embeddings

To translate these lists of 384 numbers into understandable images, we use a full differentiation approach. By maximizing the cosine similarity between the generated image's embedding and a target embedding, we can create visuals that match these points in the embedding space.

Techniques Used

Two key techniques facilitate this conversion:

  1. Full Differentiability: Allows adjusting the image pixels so that the output vector better matches a target vector.
  2. Cropping and Augmentation Strategy: Simulates how DINOv3 was trained, avoiding high-frequency noise that could bias results.

Practical Applications

Vision embeddings have varied practical applications, ranging from object recognition to artistic image generation. Companies can use these technologies to enhance their image recognition systems or create AI-assisted design tools.

Use Case

Consider a graphic design startup using DINOv3 to generate artistic concepts based on minimal textual descriptions. By providing directions in the embedding space, it can explore innovative visual styles.

Conclusion

Vision embeddings offer a powerful and flexible means for neural networks to understand and interpret the visual world. With models like DINOv3, we are taking a new step towards machines capable of autonomous perception and creation.

Let's discuss your project in 15 minutes.

vision embeddings DINOv3 neural networks image generation AI applications
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call