Google Patents a Step-by-Step AI Process for Fixing Blurry or Damaged Photos
Google has filed a patent for an AI system that repairs low-quality photos by running them through a loop of small, guided corrections, rather than trying to fix everything in one shot.
What Google's iterative photo-repair system actually does
Imagine you take a photo in a dark room and it comes out blurry and grainy. You'd love to recover the sharp, clear image that should have been there. That's exactly the problem Google's patent is trying to solve, and it approaches it in an unusually methodical way.
Instead of feeding the bad photo into an AI and hoping for a single clean result, the system described here runs the image through a model many times in a row. Each pass, the model makes a prediction about what the final clean photo should look like, then blends that prediction back with the current image and a small dose of controlled noise. Over several of these rounds, the image is gradually nudged toward something sharp and realistic.
The training side is equally practical: the AI learns by looking at pairs of low-quality originals and their high-quality counterparts. That means it's learning from real before-and-after examples, not from abstract rules about what a good photo should look like.
inputting the current image into the image restoration model; receiving, from the image restoration model, a prediction of a final image based on the input current image; combining the current image, the prediction of the final image, and a noise sample into an output image …
Translation: The AI repeatedly processes the picture and adds noise to help rebuild missing details.
How the diffusion loop rebuilds a degraded image step by step
The patent describes a computer-implemented method that uses a technique borrowed from diffusion models (the same family of AI behind image generators like Stable Diffusion) and applies it specifically to image restoration, not generation from scratch.
Here's how the loop works at a technical level:
- The degraded photo is loaded in as the starting point, called the current image.
- At each step, the current image is fed into a trained neural network, which outputs a prediction of what the final clean image should look like.
- That prediction is then blended with the current image and a noise sample (a small random perturbation, which keeps the process from collapsing to a single flat answer too quickly) to produce a new version of the image.
- This new version becomes the input for the next step, and the cycle repeats for a fixed number of steps.
- After all steps are done, the final version of the current image is returned as the restored photo.
The noise injection at each step is a key borrowed idea from diffusion models. In standard image generation, noise is added and removed to create diversity. Here, it serves as a stabilizer that lets the model make incremental, confident corrections rather than one large, potentially wrong leap.
Training uses paired datasets of degraded and clean images, so the model learns directly from real-world examples of the kind of damage (blur, noise, compression artifacts) it will need to fix.
Provided is an iterative image restoration process using low-quality images paired with the desired high-quality restored image as training data for the one or more image restoration models.
Translation: The system learns to fix bad photos by studying pairs of blurry images and their clean versions.
What this means for photo editing and AI image quality
For everyday users, this kind of system could eventually mean that old, blurry, or poorly lit photos get a second chance without the artificial, over-smoothed look that plagues many current AI photo-enhancement tools. The iterative approach is designed to preserve realistic texture and detail rather than painting over problems with a generic idea of what a face or landscape should look like.
Google has been filing steadily around diffusion-based image processing for a while now. For the photo-editing software industry, a well-executed patent in this space could shape how restoration tools are built across the board, from phone camera processing to professional archival work.
This is the 48th Google filing we've tracked since May in AI photo editing, after earlier applications on copying camera angles across places and fixing light in 3D scenes.
Claim 1 covers a broad method: take a bad image, run it through a model repeatedly, mix the prediction with noise at each step, and output the result. That's a fairly general description of the iterative diffusion approach applied to restoration. It doesn't specify the architecture of the model, the type of image degradation, the number of steps, or how the noise is sampled.
That breadth is a double-edged thing. On one hand, if granted as written, this claim could read on a wide range of implementations, not just Google's specific version. Any system that follows the basic loop of predict, blend, noise, repeat for restoration purposes might fall within its scope. That's a wide net.
On the other hand, diffusion-based restoration is a fairly active research area, and prior art is dense. The patent office will likely push back on how much of this is genuinely new versus a straightforward application of known diffusion techniques to a restoration task. The real question is whether the specific combination, using the degraded image as the start state rather than pure noise, is novel enough to hold up. That's a narrower claim than the current language suggests.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
5 drawing sheets from US 2026/0268453 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →