The moat is the garment, not the model
Recent news
Models
A planned family of domain-specialized image models, trained and evaluated on the garment-identity graph. We publish benchmarks before we train: each model is gated on measured gaps in frontier systems, not on ambition.
A compact model for garment-identity conditioning that holds one specific physical garment faithfully within a single generated image.
The workhorse, built for multiview generation with a shared garment representation that keeps the same SKU coherent across poses, scenes, and try-on.
The same domain structure at the largest scale our data supports, intended for a higher quality tier and for offline batch generation.
A reference-grounded 4K upscaler that takes generation output and grows it to full resolution, preserving garment identity, labels, and logos along the way.
Systems
The annotation layer over every image we process, covering role classification, label detection, garment segmentation, and SKU-level grouping, along with the garment-identity graph it produces.
Self-hosted model infrastructure. Each model is a self-contained, scale-to-zero deployment that can be stood up, evaluated, and retired on its own.
A benchmark for garment identity and multiview consistency, built from the identity graph and calibrated against a human-verified core.