Samsung Patents a Way to Train Image AI by Hiding Parts of Photos

By Patentlyze Team · Updated Jul 24, 2026

Samsung is teaching its image AI to fill in the blanks, by deliberately cutting pieces out of photos during training and making the model figure out what it's missing.

How Samsung's patch-dropping training trick works

Imagine giving a student a history textbook with random paragraphs blacked out, then testing them on the full text. If they score well, they've learned to think, not just memorize. Samsung's patent applies the same idea to teaching AI how to see.

The system takes an image, chops it into a grid of small tiles, removes one or more of those tiles, and then feeds both the original and the partial version into an AI model. The model is penalized whenever its understanding of the two versions differs too much. Over time, this pushes the AI to build a deep, flexible understanding of what's in a photo rather than memorizing surface-level patterns.

This kind of training can make AI vision models leaner and more reliable, especially on devices like smartphones where you want the model to handle real-world, imperfect images without burning through battery or needing a cloud server.

Inside the loss-function comparison between full and cropped images

The patent describes a training method for vision transformer models (a class of AI that processes images by breaking them into a grid of patches, similar to how a text AI processes words in a sentence).

Here's the core loop:

• Take an original image divided into a grid of patch tiles.
• Remove one or more of those tiles to create a transformed (incomplete) image.
• Run both the original and the incomplete image through the transformer model.
• Calculate the loss (a score measuring how different the model's outputs are for the two versions).
• Update the model's internal weights to minimize that difference, forcing it to produce consistent, meaningful representations even when parts of the image are missing.

This approach is a form of self-supervised learning (training without manually labeled data), and it's closely related to techniques like masked image modeling. The key insight is that a model forced to stay consistent across full and partial views learns richer, more transferable features than one trained only on complete images.

What this means for Samsung's on-device AI cameras

Vision transformers are increasingly the engine behind camera AI features like scene recognition, object detection, and photo enhancement on flagship phones. Training them more efficiently, or with less labeled data, is a real engineering priority for Samsung as it ships more AI features directly on-device rather than routing them through cloud servers.

For you as a user, a model trained this way could be more reliable in messy real-world conditions: partial occlusions, motion blur, or objects at the edge of frame. It's not a feature you'd see advertised, but it's the kind of foundational work that raises the floor on everything from photo search to real-time scene framing in the Galaxy camera app.

Editorial commentary on a publicly published patent application. Not legal advice.