BLOG
GPT Image 2 Leaked: What It Means for Manga Translation
On April 4, 2026, three image models showed up on LM Arena and Design Arena under the codenames maskingtape-alpha, gaffertape-alpha, and packingtape-alpha. Within hours, testers had figured out who they belonged to: OpenAI. The "tapes" were beating Google's Nano Banana Pro — the model behind Gemini 3.1 Flash and most production manga pipelines today — in blind votes.
OpenAI hasn't confirmed a thing. There's no API, no pricing, no announcement. But for anyone editing manga panels, two of the leaked capabilities are worth paying attention to.
What actually leaked
Three Arena models, two production codenames in the leak metadata: Hazelnut (flagship) and Chestnut (mini). Community-reported numbers from blind testing:
- ~99% accuracy on rendered text — including non-Latin scripts.
- A dedicated 4K upscaler.
- Inpainting roughly 4× faster than GPT Image 1.5, with better detail retention around faces.
- Style consistency that holds across multiple sequential edits to the same image.
Adjacent to the leak, OpenAI also confirmed that DALL-E 2 and DALL-E 3 sunset on May 12, 2026. The DALL-E brand is being retired in favor of the GPT Image line — image generation is being folded into the language-model stack instead of living as a separate product.
How it edits images (the part that matters for manga)
Two capabilities decide whether an image model is useful inside a manga pipeline. GPT Image 2 is reportedly stronger on both, but only on these two — everything else is gravy.
1. Text rendered as a native pixel, not pasted on top. Today's models, Gemini included, treat translated text as an overlay. Look closely and it sits on the page. GPT Image 2 generates text as part of the image — correct perspective, lighting, ink weight, slight bleed into the paper texture. That's the difference between a typeset bubble and one that looks drawn in. For SFX (the ドドド and バキ that are the art), it matters even more.
2. Inpainting that doesn't melt the art around it. Pulling out the original Japanese and reconstructing the panel behind it is where most pipelines fail visibly — smeared faces, lost line weight, ghosted backgrounds. The reported 4× speedup is the headline, but the more useful claim is detail preservation: faces and fine line art survive the rebuild.
That's the whole pitch for manga. Style consistency, world knowledge, photorealism — all useful, none load-bearing.
What changes for translation tools
For now, nothing. There's no API to integrate. Tools running on Gemini today — including Inkover — keep running on Gemini, because it ships and GPT Image 2 doesn't.
When it lands, the pragmatic move is multi-model routing: send text-heavy renders to whichever model has cleaner typography that month, keep OCR with whatever's strongest on Japanese. Single-vendor pipelines age fast in this space.
For translators, the only honest claim is less manual cleanup — fewer melted backgrounds, fewer crooked SFX, fewer correction passes on bubble text. The cultural and tone work doesn't get cheaper. That part is still you.
Sources
- LM Arena and Design Arena leak threads — community discussion across X and Reddit, April 4, 2026.
- OpenAI DALL-E sunset announcement — May 12, 2026 deprecation date.
GPT Image 2 has not been officially announced. Every performance number above is from blind community voting on Arena, not OpenAI marketing.
Related reading: