Can an AI learn to read fashion photographs?

A small one can, for the part that matters to AI shopping: naming the house, season and year of a runway photo. Theia, a research project of ours, gets all three right on 56.4% of 287 runway photos, ahead of eight other models.

Theia is part of ArchiveAgent, a project inside our research on how AI sees fashion. It is a small open model fine-tuned on 5,025 examples for $18.35. It is a research model, not a product, and access is by invitation.

Fig. · Runway Test ranking · N = 287 photos · house, season and year all correct · October 2026 · Theia 56.4%, Kimi K3 44.9%, Claude Opus 5.5 28.6%, GPT-6.1 Sol 26.1%, Gemini 3.1 Pro 23.0%, Claude Sonnet 5.5 21.6%, Kimi K2.6 12.2%, Qwen3.5-397B 5.2%, Grok 4.7 1.4%

Theia56.4%

Kimi K344.9%

Claude Opus 5.528.6%

GPT-6.1 Sol26.1%

Gemini 3.1 Pro23.0%

Claude Sonnet 5.521.6%

Kimi K2.612.2%

Qwen3.5-397B5.2%

Grok 4.71.4%

What this does not show

Theia lost the exact product test (50.7% against 58.7% for Claude Opus 5.5) and drops to 57% on houses it never trained on. Each model ran once and we scored it ourselves, so treat this as a first result, not a verdict.

Source: ArchiveAgent, Theia version one (archiveagent.si), October 2026.

See the rest of our research

Studies on what AI shopping systems show, link to and recommend.

Read our research