New Google Patents · Filed Jul 15, 2025 · Published Jul 30, 2026 · verified — real USPTO data

Google Patent Applies Distinct Photo Filters According to Recognized Scene Content

Most phone cameras apply the same color and exposure rules to every pixel in a shot. Google's new patent describes a system that first figures out what each pixel actually is, then applies a completely different set of adjustments depending on the answer.

Google Patent: Semantic-Based Image Signal Processing — figure from US 2026/0220741 A1
Figure from the official USPTO publication.
See all 7 drawings from this filing ↓
Publication number US 2026/0220741 A1
Applicant Google LLC
Filing date Jul 15, 2025
Publication date Jul 30, 2026
Inventors Jun Nishimura, Liang-Chun Lin
CPC classification 382/100
Grant likelihood Medium
Examiner CENTRAL, DOCKET (Art Unit OPAP)
Status Docketed New Case - Ready for Examination (May 13, 2026)
Parent application is a National Stage Entry of PCTUS2023024031 (filed 2023-05-31)
Document 21 claims

How Google's camera knows skin from sky before editing

Imagine you take a photo of a person standing in front of a bright sky. Your phone's camera software has to decide how to sharpen the image, adjust colors, and fix the exposure. The problem is it usually applies the same recipe everywhere, which is why faces can look washed out when the background is bright.

What Google is describing here is a system that looks at the photo and labels what's in each part of it, things like skin, sky, foliage, or pavement, before any editing happens. Then it applies a different set of adjustments to each category. Skin tones get one treatment, the sky gets another, and so on, all in a single processing pass.

The clever part is that this labeling and filtering happens through a compact data structure called a bilateral grid, which lets the system do all of this quickly enough to work on a phone without draining the battery. It's the difference between a one-size-fits-all Instagram filter and a professional photo editor who knows what they're looking at.

How the bilateral grid maps pixels to per-class filter settings

The patent describes a pipeline with several distinct stages working together.

First, an ML model analyzes the incoming full-resolution photo and produces a bilateral grid, a compressed, lower-resolution map of "transform coefficients." Think of it as a cheat sheet that records, for each region of the image, what kind of thing is there (a face, the sky, a tree) and how pixel values in that region should be remapped before filters are applied.

Next, a grid slicing step takes that low-resolution cheat sheet and reconstructs a specific coefficient for each individual full-resolution pixel by interpolating (filling in the gaps) from the grid. This means the system avoids storing a huge per-pixel map while still getting per-pixel precision at the end.

The reconstructed coefficient is then used to generate a filter guide value for each pixel. That guide value feeds into the final image filter, and here's the key detail: the filter's parameters change depending on which class the pixel belongs to. A pixel classified as skin might get a gentle smoothing operator; a pixel classified as foliage might get a sharpening operator. The classes are learned from the image itself, not pre-programmed categories.

The result is an output image where every pixel has been processed with settings tuned to what it actually represents, without running a separate full-resolution pass for each category.

What this means for computational photography on Pixel phones

For Pixel phone users, this kind of semantic-aware processing is what separates a technically correct photo from one that looks natural. When your phone's camera already applies things like face retouching or sky enhancement, it's usually doing so with separate, manually tuned systems. A unified pipeline that learns the scene content and adjusts filters accordingly could produce more consistent results across unusual lighting or mixed scenes.

More broadly, this is Google pushing image signal processing (the chip-level work that turns raw sensor data into a finished photo) in a direction where the camera doesn't just capture a scene, it understands it first. That matters for everything from portrait mode accuracy to video processing on future Pixel hardware.

Editorial take

This is a technically solid patent in a space where Google already has real products and competitive pressure from Apple and Samsung. The bilateral grid approach is a known technique in computational photography, and the novelty here is wiring an ML-based scene classifier directly into that pipeline. Whether this ends up in a chip or stays on paper, it reflects a clear direction: Google wants its image processing to be content-aware at the hardware level, not bolted on as a software afterthought.

The drawings

7 drawing sheets from US 2026/0220741 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

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