Most people wear only about 20% of their clothes, but the rest isn’t useless—it’s untapped potential. Alta Daily, an AI fashion app launched in 2025, turns that insight into a product: users photograph their wardrobe, and the app recommends outfits using natural language prompts, shows them on a personal avatar, and tracks daily wear to avoid repeats.
At the core of Alta is Meta’s Segment Anything Model (SAM), which has segmented and digitized millions of outfits. The technical challenge was removing backgrounds from user-uploaded images to create a clean, editorial-style interface. Fashion images are messy: a white sneaker against a white wall, a blue sweater on a wrinkled blue blanket in bad lighting, thin hangers, reflective surfaces, and human models all break naive segmentation.
Why SAM Won the Evaluation
The Alta team tested multiple segmentation models across eight product categories, from sunglasses to shoes. SAM consistently delivered the best results, handling mirror selfies and items on carpeted floors. Founder and CEO Jenny Wang explained the stakes: “If we knew that every image uploaded was a beautiful model shot, segmentation would be far easier, but because of the nature of user-uploaded content, we need the best possible segmentation.”
SAM 3’s improvements over earlier versions were visible in their tests, enabling the clean interface that makes digital styling feel seamless. The team is already experimenting with Meta’s SAM 3D models for more immersive avatar interactions.
The Cost Math That Changes Decisions
Wang was “shocked” by the cost of external segmentation APIs she first explored—a few cents per image. That adds up fast when users upload constantly. As an early-stage company, Alta had to balance experience quality with budget. Using SAM, they processed more than 20 million images without exorbitant costs.
This is the same tradeoff builders face when choosing between managed APIs and open models. The AgentCore Runtime discussion touched on a similar tension: owning the container versus paying per call. Here, owning the segmentation model—even with engineering overhead—unlocked scale that per-image pricing would have made prohibitive.
What This Means for Your Segmentation Stack
Alta’s experience suggests three practical lessons for product builders:
- Evaluate on your real data, not benchmarks. Fashion’s messy user uploads are a stress test that generic datasets miss.
- Count the per-unit cost early. A few cents per image seems small until you multiply by millions.
- Open models can be a strategic moat. Alta built a massive fashion-specific dataset and runs continuous evals between models, which is only feasible when you control the pipeline.
The app has gained users in the US, France, Germany, Mexico, and the Netherlands, showing that the approach scales globally. Wang’s team plays with everything new, but the foundation is a model choice that balanced quality and cost from day one.
For builders shipping computer vision features, the question isn’t just which model segments best—it’s which model lets you process 20 million images without burning your runway.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.