GPT Image 2: When AI Image Generation Switches from "Art Class" to "Language Class"

I. Architectural Revolution: When Image Generation Learns to "Think"

Let's lead with the conclusion: OpenAI has likely already moved off the "pure diffusion model" track. They've switched image generation from "art class" to "language class" — using an LLM that can read instructions, retain context, and understand object relationships to handle semantic planning, with the final pixel-level rendering handled by a diffusion component or other decoder. And that LLM is, in all likelihood, GPT-4o.

Why Diffusion Models Can't Spell

Over the past two years, the AI image generation field has had one unwritten law: you can get AI to render the translucent texture of a backlit Maine Coon's fur, but you can't get it to correctly spell the six letters of "Coffee" on a shop sign.

The reason lies in how diffusion models work. At its core, it's a "sculptor" — starting from random noise, each step uses a U-Net to predict and erase noise, and after dozens of iterations, it "carves out" fur, irises, and light effects. This process excels at reconstructing continuous textures — the fur can be a little stiffer or softer, the color 5% warmer, no harm done.

But text is made of discrete symbols. There's no "looks like" — there's only "is or isn't." The letter A is A. You can't add 15% of B and 8% of C and still expect it to be A. Every denoising step in a diffusion model is a tiny "estimation." Applied to textures, that's style. Applied to text, it turns "O" into "0" or produces "WElcOm e."

The Autoregressive Solution: Pixels as Tokens

GPT Image 2's approach: treat images as language.

A tokenizer slices an image into a grid and assigns each cell an ID — this is image tokenization. From CLIP to DALL·E to GPT-4o, OpenAI has gradually built a semantic representation system capable of mapping between vision and language. Images and text are projected into the same aligned semantic embedding space.

In the LLM's eyes, the phrase "a backlit Maine Coon" and a photo of a backlit Maine Coon are simply two sets of coordinates in the same semantic space. Generating an image is no different from generating text — it's just tweaking a few tokens in the encoded language system.

That's why text suddenly renders correctly. Because to the LLM, writing a "W" and writing a "我" are fundamentally the same operation.

Autoregressive Sets the Tone + Diffusion Polishes

If images were generated entirely through autoregressive token generation and then mapped pixel-by-pixel, the visual quality would be disastrous — autoregressive models are exceptionally good at deciding what to draw, but not particularly skilled at making it look good. Diffusion models, meanwhile, are the exact opposite: their textures and lighting are indistinguishable from reality, but they often have no idea what they're drawing.

A highly coherent inference, therefore, is this: the autoregressive component sets the tone — generating semantic tokens from the prompt, nailing down what's in the scene, spatial relationships, and overall composition; the diffusion component handles the polish — receiving the semantic tokens and filling in high-fidelity pixels.

Google published the Transfusion paper. Meta built Chameleon. They're on similar paths. Is OpenAI using this approach? At the April 2026 press event, OpenAI refused to answer any questions about model architecture. The refusal itself is a signal.

Experience the new architecture: GPT Image v2 Text to Image, GPT Image v2 Image to Image.