Google Patents an AI That Fills In the Washed-Out Parts of Your Photos

By Patentlyze Team · Updated Jun 5, 2026

Blown-out skies and washed-out faces have been a stubborn problem in phone photography for years. Google is now filing a patent that uses a diffusion AI model — the same class of model behind image generators — specifically trained to hallucinate the missing detail in overexposed regions.

What Google's overexposure fix actually does to your photos

Imagine you snap a photo of your friend standing in front of a bright window. The background sky turns into a featureless white blob — those pixels are "clipped," meaning the camera sensor recorded as much light as it could and still lost the actual detail. Most editing tools can't recover what was never captured.

Google's approach here is to train a diffusion model (the AI architecture that powers tools like Stable Diffusion) on thousands of paired images — one overexposed, one correctly exposed — so it learns how to convincingly fill in what should be there. The model generates a corrected version of only the blown-out areas.

The clever part is the merge step. Rather than replacing your whole image with the AI's output, Google's system calculates blend weights based on how bright each pixel is. Bright, clipped pixels lean heavily on the AI's reconstruction; pixels that were already fine stay mostly original. A final tone-mapping pass then stitches everything into a natural-looking result.

How the diffusion model rebuilds and blends clipped pixels

The pipeline has four distinct stages working in sequence:

• Diffusion model inference: The input photo is fed to a diffusion model (an AI that iteratively refines noisy images into coherent outputs) trained exclusively on overexposed/ground-truth image pairs. It produces an "intermediate image" where clipped regions have been reconstructed with plausible detail.
• Pixel-brightness merge weighting: The system computes per-pixel merge weights derived from the brightness of the original input. Fully clipped (white) pixels get high weight toward the AI output; pixels that were never overexposed get high weight toward the untouched original. This is essentially a soft mask driven by luminance.
• Image blending: The intermediate AI image and the original are composited together using those per-pixel weights, producing a merged image that borrows reconstruction only where it's needed.
• Tone mapping: The merged image is passed through a tone-mapping operator (a process that compresses a wide dynamic range into a displayable range) to produce the final output.

The training data strategy is worth noting: by pairing intentionally overexposed images with ground-truth references, the model learns a highly specific task rather than being a general image-generation model, which likely improves accuracy and reduces hallucination artifacts in non-clipped areas.

What this means for computational photography on Pixel phones

Overexposure recovery has been a known weak spot in computational photography pipelines, particularly in high-contrast scenes like backlit portraits or sunsets. Current HDR approaches typically rely on multiple exposures captured simultaneously (like Google's own HDR+ on Pixel phones). This patent describes a method that could recover detail from a single clipped frame — no bracket shooting required.

For Pixel camera users, this kind of pipeline could quietly improve the worst-case shots that even Night Sight and HDR+ struggle with. It could also be relevant for post-processing in Google Photos, where existing images with blown highlights could be retroactively fixed. The diffusion model approach is more computationally expensive than traditional curve-based recovery, but on-device ML accelerators (like the Tensor chip's image signal processor) are increasingly capable of running exactly this class of workload.

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