Independent hackathon demo. Not affiliated with Meta or Facebook Marketplace.Independent demo. Not affiliated with Meta.
Sellers choose a category, tap attributes or write their claims, and take the photos those claims need. Meta's Muse Spark describes each photo without seeing the claims, and fixed rules show buyers what's supported, what's contradicted, and what photos can't settle.
From our 12-case furniture evaluation with Muse Spark and one live check. How we measured
Uses
The model describes; it never decides. Every verdict comes from plain rules you can read, applied to what was actually in the photos.
Each claim gets a type: a piece count, a colour, no visible damage, no visible damage on a specific part, label text, or other. Other claims stay on the report, marked as something photos can't show.
A count needs the front and side; damage adds a close-up. The listing's code goes on paper in the front photo.
Muse Spark gets one photo at a time and the nouns to count, never the claims. SAM 3.1 counts again and boxes the damage.
The report is fingerprinted with its photos. Editing the listing voids it, the seller can withdraw it, and it expires in 30 days.
This is what the sample produces: recorded Muse Spark observations of our own test photos of a dining set.
Evidence report · sample
No scratches or stains
Seen: peeling in the detail photo (tabletop near edge).
Set of 4 chairs
The front photo shows all 4 chairs.
White finish
The item in the photo is white.
Solid oak, not veneer
Photos can't show this.
We ran 12 labelled furniture cases two ways with Muse Spark 1.3: blind observation followed by the rules, and the model shown the claim next to the photos and asked for a verdict.
Synthetic furniture photos; labels written before the first run. The new categories and part/label rules are not covered by this evaluation. A sanity check, not a benchmark: the two set-ups also differ in their instructions. Muse Glimmer 30B gave the same split. The whole run cost $0.22.
"Three seat cushions", next to a photo of a sofa too dark to make out.
It contrasts claims with photo evidence and says what stays inconclusive. It never certifies the item or the seller.
The sample replays recorded Muse Spark observations of our test photos, so it needs no photos of yours and makes no API call. It starts without the close-up: check, see Spark ask for it, add it and check again.