Google Patents a Two-Model AI System That Grades Its Own Image Drafts
Most AI image generators produce a picture in one shot, crossing their fingers. Google's patent describes a system that generates a draft, grades every piece of it, throws out the weak parts, and tries again, all automatically.
How Google's image AI checks its own work
You're generating an AI image and the result comes back looking almost right, but something's off about the hands, or the background is blurry where it should be sharp. There's no easy way to tell the system exactly what went wrong at a granular level, so you regenerate the whole thing from scratch.
Google's patent describes an approach that avoids that brute-force retry. The system breaks an image into small coded chunks (think of them like puzzle pieces), then runs those chunks through a separate "critic" model that scores each one. High-scoring pieces get locked in; low-scoring ones get tossed out and regenerated. This loop can repeat several times until the whole image meets the bar.
The key idea is that the critic is independent from the generator. They're trained separately, so the critic isn't just rubber-stamping the generator's own choices. That separation is what lets it catch errors the generator itself wouldn't flag.
… generating, using a second neural network different than the first neural network, a first plurality of scores based on the first plurality of tokens, each score of the first plurality of scores representing a prediction of whether a token of the first plurality of tokens was generated by a generative model; …
Translation: A second AI checks the work of the first AI by scoring the quality of the generated image pieces.
How the token-critic scores and filters image pieces
The patent describes a two-model pipeline for generating images through iterative refinement rather than a single forward pass.
The first model (a generative neural network) doesn't output pixels directly. Instead, it predicts probability distributions over a codebook of visual tokens, think of tokens as small compressed patches that represent regions of an image. Sampling from those distributions gives you a first draft image encoded as a sequence of tokens.
The second model is a token-critic, a separate neural network trained to do one specific job: for each token in that draft, predict whether it looks like something a good generator would have produced or something that slipped through. Each token gets a score.
The system then selects which tokens to keep based on those scores, locks them in, and feeds them back to the generative model as constraints. The generator fills in the rest, producing a new set of tokens for the positions that were discarded. Because this is non-autoregressive (meaning it doesn't generate tokens one by one in a fixed sequence, but can work on many at once), the process is faster than classic left-to-right generation.
This refine-and-rescore loop can run for multiple passes, progressively locking in confident tokens until the full image is settled.
Systems and methods for iterative non-autoregressive image synthesis using a first generative model and an independent second token-critic model.
Translation: The system builds images through a loop where the main creator model works alongside a separate critic model.
What this means for AI image quality and speed
For anyone using AI image tools, the practical promise here is fewer garbage outputs. Today's generators can fail in subtle ways, a confident-looking result that has structural problems most users can't articulate. A built-in critic that catches weak tokens before the final image is committed could mean fewer retries and more consistent quality out of the box.
Google's run of image-generation filings suggests this is part of a longer effort to close the quality gap with diffusion-based tools like those from competitors. The non-autoregressive approach also has speed advantages: generating many tokens in parallel (rather than one at a time) means the critic-and-refine loop doesn't necessarily add much wall-clock time.
That makes this Google's 16th filing we've tracked since May in our AI models working in teams watchlist, following a voice AI that overrides itself and routing tasks to specialist AIs.
The core technology here is pure software: no new chips, cameras, or physical devices required. Everything described runs on computing infrastructure that already exists inside large AI companies, which puts this closer to a finished product than most patents suggest.
The key design choice is using two separate networks, one that creates images and one that grades them, rather than letting a single system evaluate its own work. A teacher grading their own test tends to miss the same mistakes they made while writing it; a separate grader catches what the first one cannot, and that difference shows up directly in output quality.
The filing goes deep on how the system is trained but says little about how it would reach actual users, so the honest read is that this could become a consumer feature relatively quickly or could serve mainly as a technical foundation others build on later.
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
12 drawing sheets from US 2026/0260391 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →