• Etsy launches image-based search on mobile, letting users find similar items via photos. • Feature uses machine‑learning embeddings to represent 100M+ listings in low‑dimensional space. • Query image converted to embedding, then nearest‑neighbor search retrieves visually similar results instantly. • Embeddings cluster by category (e.g., bags, craft supplies, home goods) for intuitive similarity. • System built on approximate nearest‑neighbor (ANN) indexing for sub‑second performance. • Adoption aligns with e‑commerce trend, enhancing discovery for unique handmade products. • Engineering focus on scaling embeddings, efficient ANN, and rapid query response across millions of items. • Future work includes multi‑task modeling to improve relevance and personalization.

Article Summaries:

  • Etsy has launched a new image‑based discovery feature on its mobile apps, letting shoppers search the marketplace by uploading a photo. The system converts every listing image into a dense embedding using a fine‑tuned EfficientNet CNN, then stores these vectors in an approximate nearest‑neighbor index. When a user submits a query image, the model generates its embedding and retrieves the most visually similar listings in milliseconds. The approach leverages transfer learning-freezing most pretrained layers and training only the final layers and classification head-while using a classification proxy to learn useful embeddings. The feature is already gaining traction across e‑commerce platforms.
  • Etsy has launched a new image‑based search feature on its mobile apps, letting shoppers upload a photo to find visually similar listings from its nearly 100 million‑item catalog. The system converts every listing image into a dense embedding using a transfer‑learned EfficientNet CNN, then indexes these vectors in an approximate nearest‑neighbor (ANN) engine for millisecond retrieval. The embeddings are generated by training the model on Etsy’s own product data, with the final classification head removed to produce a reusable feature vector. The rollout demonstrates Etsy’s move toward multimodal search, leveraging machine‑learning to match user‑submitted images with relevant items.
  • Etsy has launched a new image‑based search feature on its mobile apps, letting shoppers upload a photo to find visually similar listings from its 100 million‑item catalog. The system converts every listing image into a dense embedding using a transfer‑learned EfficientNet CNN, then indexes these vectors in an approximate nearest‑neighbor (ANN) engine for millisecond retrieval. Etsy trained the model on its own data, replacing the pre‑trained head with a task‑specific classifier and fine‑tuning only the final layers. The approach leverages embeddings learned via a classification proxy, enabling rapid, scalable visual search across the platform.
  • Etsy has launched a new image‑based discovery feature on its mobile apps, letting shoppers search the marketplace by uploading a photo. The system converts every listing image into a dense embedding using a fine‑tuned EfficientNet CNN; the model’s penultimate layer provides the vector representation. Pre‑computed embeddings for nearly 100 million listings are indexed in an approximate nearest‑neighbor (ANN) search, returning visually similar results in milliseconds. Etsy trained the network on its own catalog, using transfer learning to freeze most pretrained weights and optimize only the final layers and classification head. The rollout demonstrates the company’s shift toward multimodal search and advanced machine‑learning infrastructure.
  • Etsy has launched a new image‑based search feature on its mobile apps, letting shoppers upload a photo to find visually similar listings from its 100 million‑item catalog. The system converts every listing image into a dense embedding using a transfer‑learned EfficientNet CNN; only the final layers are fine‑tuned on Etsy’s own data while earlier layers remain frozen. Query embeddings are matched against pre‑computed embeddings via an approximate nearest‑neighbor (ANN) index, returning results in milliseconds. The approach, trained on a classification proxy task, clusters similar items (e.g., bags, craft supplies) in embedding space, enabling rapid visual retrieval across Etsy’s diverse inventory.

Sources: