2026-09-01

1First paper to extend SID generative retrieval to image search — Alibaba Pailitao

Pailitao (拍立淘) is a surface where a user submits a single real photo as the query and finds identical or similar products in the Taobao catalog. It is a product that has run for a long time on the Taobao home, alongside text search. Image search is harder than text search because a real photo carries several kinds of information at once. A single frame mixes the search target, such as one piece of clothing, auxiliary evidence, such as a brand logo or tag text in the photo, and irrelevant background, such as a cafe interior or another person's arm. The user wants only one of these. This paper is the first attempt to extend SID generative retrieval to the image modality, and it tries to show that the retrieval backbone can handle cropping and OCR in a unified way.

• Pailitao joined the SID generative retrieval lineage verified — arXiv 2608.26658.
  • All nine authors are affiliated with Alibaba Group Hangzhou.
  • The HTML version explicitly names Pailitao as "Taobao's image search system".

Earlier approaches left the job of separating the mixed information in an image to a preprocessing pipeline. A crop model cut out the search target region and an OCR model extracted the text in the photo, and the search model consumed the results. PailitaoGR removes cropping and OCR from the pipeline and absorbs them into the retrieval model. The authors called these "Zooming without Cropping" and "Reading without OCR".

• PailitaoGR unifies two components under the concept name Latent Think-with-Images verified.
  • The target enhancer does not crop explicitly; it emphasizes the search target region among the visual tokens of the full image and absorbs this through training ("Zooming without Cropping").
  • The auxiliary enhancer selectively extracts and uses relevant text evidence in the image without an OCR pipeline ("Reading without OCR").
• Training combines on-policy distillation and a contrastive strategy verified.
  • The target enhancer and the auxiliary enhancer receive different supervision, but they are aligned so that the generative retrieval model itself internalizes both capabilities.

• The paper states an average +13.8% improvement over the offline baseline verified.
  • The dataset was sampled from Pailitao's real image search logs.
• Production A/B numbers are not released in the abstract or the body verified.
  • From the abstract alone, we cannot confirm whether the model was actually deployed to serving on Pailitao or is still at the offline experiment stage.
  • Given Alibaba's production practice, if the data are production logs and all authors are Alibaba employees, the model may be planned for deployment or partially deployed inferred. However, the abstract does not give evidence for this.

"Generative retrieval has demonstrated strong performance by directly generating product semantic identifiers (SIDs). Extending this paradigm to image search, however, is nontrivial because real-world query images contain diverse information, including the search target, useful auxiliary evidence, and irrelevant visual content." — PailitaoGR (arXiv 2608.26658), 2026-08-27

"we propose PailitaoGR, a Latent Think-with-Images method for generative image retrieval, which internalizes target-focused perception and selective auxiliary-evidence utilization into the generative retrieval model, enabling Zooming without Cropping and Reading without OCR." — PailitaoGR (arXiv 2608.26658), 2026-08-27

• The number of companies grew. This is the first time Alibaba has made a statement that treats SID itself as a serving target.
  • Until now, the companies with production claims in the SID lineage were Kuaishou and Yandex Music: HD-Rec·PushDualGen and Gryphon-v2·Sona, respectively. Counting the experimental lineage broadly, Google (the original TIGER paper) is also included. Alibaba is the fourth company on the SID line.
  • Alibaba has already made several statements on the LLM-native and scaling axes with MetaStrategy/DREAM/TransRetrieval.
  • With this paper, at least three companies on the SID line (Kuaishou·Yandex Music·Alibaba) have used production logs for training.
• The modality axis opened. This is the first statement of using SID generative retrieval on a surface that takes a real photo itself as the query.
  • So far, SID has been applied to text items (news, short video titles, music tracks), e-commerce product IDs, and user history sequences.
  • If the claim that cropping and OCR are absorbed into the backbone holds, it suggests that the entire image search stack could be compressed into SID and a single retrieval model.
  • However, because there are only offline metrics, it is not confirmed whether this compression is actually a gain in terms of production latency and cost.
• The strength is limited because the abstract has no production A/B numbers. The paper states that it trained and validated on real Pailitao logs, but it has no serving deployment claim.
  • The sub-bullets that currently carry weight in the thread are those that also released production metrics (PushDualGen's effective play rate +8.50%, Gryphon-v2's active users +1.41%). So we did not treat PailitaoGR as the same grade.
  • Within the thread, we record it only as a signal that "the image search modality has started to be incorporated into the SID lineage".
• For the tracked services X, Meta, and Netflix, there is no counterpart to fold in immediately.
  • X has no image search.
  • Meta has long operated visual search in Marketplace, but no SID-family claim is public.
  • Netflix has no image search surface itself. The artwork and preview personalization in MAPS deals with the image modality, but it is personalization rather than search, so it does not overlap directly.

Corrections

(none)

Radar

• Alibaba PailitaoGR (arXiv 2608.26658) extended SID generative retrieval to Pailitao, Taobao's image search surface. It unifies Zooming without cropping and Reading without OCR in a single backbone. It released only +13.8% offline and has no production A/B numbers. Added to the generative retrieval / semantic ID thread in radar.md.

What to check next