Semantic Visual Search

Search for images and videos using natural language descriptions. WISE interprets the visual content of media files directly.

Examples:

• man at a train station
• horse in an airplane
• man with a flower
• pirate with a pistol

Queries can be written in multiple languages. Try queries on your own language. For example:

• Hindi: एक व्यक्ति रेलवे स्टेशन पर (means "man at a train station")
• Telugu: విమానంలో గుర్రం (means "horse in an airplane")

Face Search

Locate a specific person across images and videos using a reference image.

Audio Search

Search within audio files to locate relevant segments based on spoken content or sound.

Note: Coverage is currently limited and under active development.

Support

1.  Support The tool improves content discovery and fills an important gap in Commons search. Gopavasanth (talk) 23:12, 2 May 2026 (UTC)
2.  Support can't wait to try it with all pictures Nat (WDU) (talk) 09:09, 3 May 2026 (UTC)
3.  Support obviously very useful. Not sure what this vote is for though. Also doubtful whether facial search is a good idea to include. Only truly useful if covers nearly all files (or videos) on Commons. --Prototyperspective (talk) 17:13, 4 May 2026 (UTC)

   @Prototyperspective, I’ve updated the section to clarify why we are requesting support and interest. Gopavasanth (talk) 08:14, 5 May 2026 (UTC)

4.  Support First of all it's already pretty cool to use and quite impressive, in particular the frame-wise overview of the matching video stills! Being able to easier find relevant scenes in video fames is super useful, considering videos are often only described as a whole, but not scene-wise. Furhermore, it could be quite useful for suggesting categories/depicts for media that are either uncategorized or poorly categorized, or for new uploads. Also it could possibly help people identify locations/art/objects via the image-input, by returning depicted items (P180) of the matching images (assuming it is acurate enough). Ultimately this would be great to have available as an API-action as well for the same reasons. It could be useful, if technical feasible, to include structured data (P180, P1071) in the training or otherwise correlate them, so that searching by wikidata item would be possible. I can't say anything about the involved costs and if it would be worth it, but I would find the feature definetely useful and interesting to have. Nylki (talk) 18:22, 5 May 2026 (UTC)
5.  Support interesting project Amrit Sufi (talk) 06:16, 6 May 2026 (UTC)

Oppose

Neutral / Questions

1. Perhaps it would be a good idea to make the project— and this page too —multilingual, so we can gather a wider range of opinions and raise awareness of the project.

P.S. This text was generated using deepl.com. AnBuKu (talk) 12:47, 5 May 2026 (UTC)

General feedback

I must say I'm quite impressed. I tried to come up with some intentionally weird queries (e.g. "window in the ceiling") and got surprisingly good results. That said, I noticed a few things that can be improved.

• I cannot link to search results. https://wise.wmcloud.org/?q=example for example doesn't work.
• My browser history is also broken. I cannot go back to a previous search result.
• From the perspective of a Wikipedia editor my first idea would be to find good illustrations and photos for articles. Not videos. While being able to visually search video content is an absolutely amazing feature, I think it's not as important as visually searching the 120 million photos we have.
• In some of my queries (e.g. "the oldest and most weird woodworking tool ever") I noticed a lot of frames that are taken in the exact middle of a cross-fade. This is arguable not terribly useful as such blended frames are not really part of the "content" of a video but merely a visual glitch as a result of how the content is presented. I believe it should be possible to detect cross-fades and exclude them.

--Thiemo Kreuz (WMDE) (talk) 09:49, 13 May 2026 (UTC)