Multimodal 7 min read1 December 2024

Multimodal AI in Commerce: Beyond Image Classification

Joshua Lamerton

Multimodal AI Visual Commerce Image Generation PropTexx E-commerce

For most of AI's commercial history, "multimodal" was a research term. Images and text lived in separate pipelines, separate models, and separate product teams. That's changing — faster than most commerce platforms are prepared for.

The Convergence

Three capabilities are collapsing into a single product layer: image generation, visual search, and spatial commerce (virtual product placement, scene composition, AR previews). The infrastructure to do all three exists. What's lagging is the product architecture to combine them coherently.

What's Working in Production

Virtual staging and AI-enhanced listing imagery are live at scale in real estate — PropTexx is one example. Product visualization in e-commerce (generate this chair in your room, in your color palette) is moving from demo to default. Visual search — find this product from an image — is table stakes for any fashion or home goods platform.

The models are good enough. The latency is manageable. The cost curves are declining.

What's Still Marketing

Real-time, physically accurate AR in mobile browsers. Fully generative product catalogs without human review. Video generation as a commerce tool at scale. These are demos. Impressive ones. But demos.

The discipline is knowing which capabilities to deploy now and which to watch. Multimodal AI is real — the question is sequencing.

Clap