Use a single subject, full-body or upper-body, facing the camera or at a slight angle. Good lighting and a relatively uncluttered background help the person segmentation model separate you from the scene. Avoid heavy motion blur, extreme foreshortening, and cropped limbs—the pose estimator needs visible joints (shoulders, hips, knees) to align the garment warp. Mirror selfies work if the reflection is sharp and your outline is unobstructed.

The pipeline is only as good as the inputs: occlusion (bags, crossed arms), loose hair over the torso, sheer or highly reflective fabrics, and unusual poses can confuse segmentation or warping. Diffusion-based refinement can hallucinate texture or folds that are plausible but not physically measured. Layered outfits, open jackets, and asymmetric draping are harder than a single flat garment layer. We surface a render, not a guarantee of real-world drape or tailoring.

End-to-end latency is typically a few seconds on warm GPU infrastructure: human parsing and pose detection are fast; garment conditioning and the diffusion pass dominate runtime. Queue time during traffic spikes can add delay. We target interactive turnaround for a single look, not batch overnight jobs.

Uploads are encrypted in transit and at rest, processed to generate your try-on, and held in private account-scoped cloud storage only for a limited time. Generated outputs are typically deleted about 14 days after a job finishes; failed jobs sooner (about 72 hours); unfinished provisional uploads typically within about 1 hour. We do not host a cloud browsing library of your photos—optional wardrobe/poses in the app stay on your device. We do not use customer uploads to train or fine-tune our models. Aggregated, non-identifying telemetry (e.g., error rates) may be logged for reliability. See the Privacy Policy for legal detail and regional exceptions.

Fit scores are estimates from visual cues and garment metadata, not a tailor’s measurement. Body shape is inferred from 2D images; depth and true circumference are under-constrained, so sizing recommendations carry uncertainty bands. Use predictions to narrow choices—especially across brands with different grading—not as a substitute for measurements on high-stakes purchases.

Personal and commercial use of renders you generate is allowed under our Terms, subject to rights in the underlying garment imagery (catalog shots you do not own may still carry third-party restrictions). You may not use outputs to misrepresent fit in regulated advertising where substantiation is required, or to create deceptive endorsements. When in doubt, confirm rights to the garment assets and your jurisdiction’s marketing rules.

No. A phone camera is sufficient. A tripod or stable surface reduces blur; even lighting (window or soft lamp) beats harsh overhead spots. You do not need a depth sensor, green screen, or studio setup—those can help marginally but are not required for the core pose + segmentation + try-on stack.