The newest image models are marketed on the two things they still get wrong

Everyone is suddenly selling text

Two years ago, garbled lettering was the giveaway that an image was generated. Now text rendering is the headline feature of every flagship. Ideogram 3.0 published a human-preference ELO of 1132, ahead of Imagen 3 (1023), FLUX Pro (998), Recraft V3 (937) and DALL·E 3 (910), and sells on “Photorealism. Legible text.” Seedream 4.0 pushed resolution from 2K to 4K and claims it now renders text “correctly and clearly” alongside formulas and tables. Google’s Nano Banana Pro went further and simply named itself the best model in the world for in-image text.

The honest outlier is Black Forest Labs’ FLUX.1 Kontext, whose own announcement describes its typography as merely “competitive” — and even OpenAI’s gpt-image-1 makes the softer claim that it “reliably renders text within images,” pitched at storybooks and educational material rather than dense marketing layouts. When the whole industry’s marketing points at one feature, that feature is the one that used to be broken — and “better” is not the same as “done.”

The text problem moved; it did not leave

Independent benchmarks are less generous than the launch posts. On OCRGenBench, a visual-text benchmark of 19 state-of-the-art image models, only two — Nano Banana Pro (77.19) and Flux.2-dev (70.19) — clear 70 out of 100, and most score below 60. The best in-image text in the world lands at roughly 77%. That is fine for a three-word slogan and a coin-flip for a nutrition panel, a phone number or a paragraph of legal copy.

The STRICT stress test shows where the wall is. The strongest models, GPT-4o and Gemini, “maintain strong performance up to approximately 800 characters, beyond which accuracy begins to decrease”, while most diffusion models break down past about 200. And the failure is not always visible garble: “some models, especially FLUX 1.1 pro, tend not to follow instructions as the number of characters increases,” quietly dropping the requested text and painting a clean, wordless image instead. Non-Latin scripts are worst of all — the same study found “English generally yields the highest accuracy … while Chinese exhibits the lowest.” The cause is structural, not a maturity problem the next version outgrows: a locality bias in how diffusion models build an image limits their grip on long-range spatial structure like a line of text.

So “legible text” is true for a headline and false for a paragraph, a serial number, or a logotype in a second language — which is exactly the distinction a human skimming a batch of 300 product images will miss.

Hands and physics did not get a keynote

The failure classes nobody puts on a launch slide are still measurable. A 2026 anatomy benchmark scores current models with an Anatomical Error Rate and gives FLUX.1-dev 0.417 and SDXL 0.873 before correction; a comprehensive GPT-4o evaluation flags “abnormal human poses or anatomical structures” and “spatially implausible object overlaps” as recurring weaknesses. Physics is worse: on PhyBench the best text-to-image model scored only 1.71 out of 3 on mechanics, and DALL·E 3 “generates a balanced seesaw with an elephant and a mouse on either side.” A January 2026 benchmark of 21 state-of-the-art models concludes that “higher-order spatial reasoning remains a primary bottleneck,” and T2I-CompBench finds spatial relationships the single hardest sub-category, with counting still under ~0.60.

The most telling sign is that the evaluations themselves keep being rebuilt because the models saturate them. GenEval, once well-aligned with human judgment, has “drifted far” from it — an absolute error “as much as 17.7% for current models” — and its harder successor GenEval 2 puts the top model at just 35.8% prompt-level accuracy, collapsing toward zero as prompts get more compositional. The scores went up because the test got easier relative to the models, not because hands and physics got solved.

Editing models bring their own defects

The real 2025–26 shift is from generating to editing — instruction models like FLUX.1 Kontext and Nano Banana that change one thing in an existing image. That adds a new failure class. FLUX.1 Kontext’s own paper states that “excessive multi-turn editing can introduce visual artifacts that degrade image quality,” that the model “occasionally fails to follow instructions … ignoring specific prompt requirements,” and that even its distillation step “can introduce visual artifacts.” Google, launching Gemini 2.5 Flash Image, listed the open problems out loud: it is “actively working to improve long-form text rendering, even more reliable character consistency, and factual representation.”

And the provenance mark that ships with these models is not a quality signal. Every image from Gemini carries an invisible SynthID watermark so it can be identified as AI — which tells you the image is generated, not whether the hand has five fingers or the logo is spelled right. The failure classes in 2026 are the same ones from three model generations ago; they are just rarer and better hidden, which makes catching them by eye harder, not easier. You cannot audit the model, so you have to audit the output — one scan per creative before it ships.