Auto Secure LoginResearch
Product Vision moves from one description to an object inventory
The research now separates counting, description and product identity, with an explicit option to withhold an uncertain answer.
Work covered: September 25–October 8, 2026
Describe the objects actually present
Product Vision's expanded workflow builds an inventory of distinct objects in a photograph, then prepares a description for each supported object. Recent evaluation adopted a combined approach that retained the better object-counting behavior while improving descriptions. Outputs remain review-required drafts and do not publish to a storefront.
Recognize a family without guessing an exact listing
Catalog-image, printed-text and listing-title research improved the identification path. An adopted update can return a sufficiently supported product family while withholding an exact listing. That distinction matters when duplicate listings or visually similar variants cannot be resolved from a photograph alone.
Measured progress with real limits
The team kept unsuccessful experiments out of the adopted path and checked selected changes on held-out photographs. The description study still found mistakes in spatial relationships and shapes. Its results concern the evaluated photographs, not every possible catalog. Human review, clear abstention and separate identity evidence remain part of the workflow; this is development progress, not an automatic listing or universal recognition promise.
Explore the product
Prepared from dated development and release records. Status is stated for the recorded milestone; internal tests and review submissions do not establish public availability. Sensitive operational details and customer information are excluded.