What I'd love
Muse Vision already finds similar images on the web for one item. I'd like to run it on a whole collection or a multi-selection and get more items with the same vibe. Example: I keep a poster collection and want to find more posters that fit it, without searching image by image.
How it could look
- A setting per collection, "At the end of this collection", with three options: Off (default), Show suggestions, Button only.
- When I scroll to the end, a clear separator appears ("More like this collection"). Below it, suggestions load in batches of about 20, marked as not saved yet (source badge, slightly dimmed).
- If the collection has several sub-vibes, suggestions come in groups ("more like these 6", "more like these 4"), and I can mute a group.
- Each suggestion has three actions: Keep (saves it into this collection), Not this (hides it and steers the next batch away) and Open source. A small hint shows why it was suggested, e.g. "same look, similar palette, matches 'poster'".
- Also available as a command: right-click a collection, select several items, or mark part of an image (e.g. only the typography) and choose "Find more like these".
What it should match on
Not only the look. Muse already knows more about every item:
- Look: the on-device image embedding.
- Colour: the dominant palette, with an optional "keep this palette" switch.
- Topic: tags, titles, notes, text found in images and the collection name.
- Form: portrait or landscape, photo, illustration or type-heavy.
A per-collection control "Match on" with Look, Colour, Topic and Form (toggles or sliders) makes it work for very different boards. A poster wall wants look and form, a clothing wishlist wants topic and colour.
Sources
- The web, through the same path Muse Vision uses today. Those results already carry the crowd knowledge of the web, and Muse re-ranks them locally with the signals above.
- Optional per collection: open archives with public APIs, e.g. Openverse, Internet Archive, Library of Congress and museum open-access collections.
Why it still fits Muse
It's not a feed, and no algorithm decides what I see. It only runs when I turn it on for a collection or ask for it, starts from my own saves, explains each suggestion, and nothing is saved until I keep it. Opt-in like Muse Vision today, with a short note on what gets sent.
Implementation idea
- Group the collection's image embeddings into a few clusters and use a real item per cluster (the medoid, not an average) as the query image. Pinterest's PinnerSage does the same for multi-interest recommendations, because averages drift away from what the user actually saved.
- For keyword-only sources, build short queries from the clusters plus tags, colour words and image text, then re-rank the results locally.
- Score each result as a weighted mix of look, palette distance, topic overlap and form. The weights come from the "Match on" control and get nudged by Keep and Not this.
- Skip anything already in the library, keep the results varied and cache them per collection.
Something like this would be very helpful for me and it would extend on the existing Muse Vision we already got.