Wearly Research

Updated 3 min readVision

The moat is the garment, not the model

Frontier image models are becoming a commodity. What lasts is knowing which physical garment every image shows. This is the bet behind Wearly Research, and an honest account of what we have built since June.

Frontier labs release new image models every few months, each one more capable and cheaper than the last. For a company that makes product imagery that is good news, because every release lowers our costs. It also tells us where our advantage cannot come from. A model we reach through an API is a model our competitors reach on the same terms.

The asset is the garment

What a general lab cannot reconstruct is the link between an image and the physical garment it shows. Every job in Wearly Studio starts from real photos of a real product, carries the product it belongs to, and ends with a person approving the result or sending it back. Over time that adds up to many images of the same garment, each tied to one SKU, one role and one verdict.

FrontBackDetailOn modelFlat layLabelOne SKU
The garment-identity graph: every image of a product linked to the one physical garment it shows, with its role and whether a person approved it.

We call that structure the garment-identity graph. The masks and tags around it are not the asset, because anyone can re-derive a segmentation. The linkage is. It exists only because real work produced it, and it cannot be scraped from the web or added back after the fact.

What it points to

It points to a narrow model, not a better general one: strong where general models are weak and where this data decides the outcome. Two cases matter. Garment identity means one physical garment held faithfully in a new image. Set consistency means the same SKU kept coherent across every view on a product page.

Measure first

We do not want to train on ambition. The plan is to measure every frontier release on garment identity, and to train only if a gap holds across consecutive releases. If general models close it, we keep buying inference and spend the effort elsewhere. That rule is written down as a stopping condition, not a hope.

Measuring has turned out to be hard in useful ways. When we tested whether frontier models can judge product images the way a client does, the best one matched the client on 20 of 24 images, and none reliably noticed a garment that had turned from brown to olive. That test is written up as a separate note.

Where things stand

  • Released: panda flush, an upscaler that adds resolution without changing the garment, and the first model in the panda family.
  • In use: bamboo forest, where our own models run. It serves panda flush and our garment matting.
  • Done: bamboo bench, the benchmark.
  • In development: the data layer and red panda. Pieces of the data layer run in the studio today; the full graph does not exist yet.
  • Planned: panda and big panda, the models that would learn garment identity. Neither has been trained.
  1. Studio workReal garments, real verdicts
  2. Identity graphImages linked to SKUs
  3. BenchmarkFrontier models measured
  4. ModelsOnly where a gap holds
The loop the program runs on. Better models would feed back into the studio, and the last step happens only if the benchmark says so.