ducttapeai.app
LMArena · anonymous image model

ducttape is the codename. Nobody has confirmed what it is.

An unannounced image model has been quietly testing on LMArena under the codename ducttape — appearing as duct-tape 1, duct-tape 2, and duct-tape 3. Text renders cleanly. Layouts hold. Reference images survive. Below is what the community has pieced together about the ducttape ai model — and a way to be first to know the moment the name changes.

One email when the codename is confirmed. No spam, no marketing — we unsubscribe you ourselves if we ever break that rule.

What the community is seeing

Text renders like a designer typed it

Posters, menus, packaging, multi-line captions — letters stay on the baseline, kerning survives, and misspellings are rare. This is the single most commented-on property on LMArena.

Reference images actually follow

Given a reference, ducttape preserves composition, pose, and palette more faithfully than Nano Banana Pro or Seedream 4 in head-to-head pairings.

Multilingual output is stable

Korean, Japanese, and Chinese characters render correctly — not romanised, not mangled. Rare for open Arena models.

How ducttape compares

Blind side-by-side on LMArena against the three image models it runs against most often. Observations are community-sourced and change as variants rotate.

Google · Gemini 2.5 Flash Image

Nano Banana Pro

Strongest at photoreal portraits and subject consistency across successive edits of the same face.

ducttape wins on
ducttape wins on in-image text, multilingual characters, and obeying explicit layout instructions.

ByteDance · Doubao series

Seedream 4

Strongest at stylised illustration, manga/anime aesthetics, and East-Asian design references.

ducttape wins on
ducttape wins on composition prompts ("three panels, third one vertical") and reference-image fidelity.

Black Forest Labs · Flux family

Flux 1.1 Pro

Strongest at open-weight workflows, ultra-realistic micro-detail, and fine-tuned LoRAs.

ducttape wins on
ducttape wins on poster and UI-mockup text, multi-line captions, and packaging copy.

Want the full picture?

Read the longer community primer on what ducttape is, what the duct-tape 1/2/3 variants mean, and what still isn't known.

Read the primer

Frequently asked about ducttape

Plain answers. No guessing, no hype.

What is ducttape?
ducttape is the community codename for an unannounced image-generation model currently testing on LMArena. The name has been in use since late 2025, sometimes appearing as duct-tape 1, duct-tape 2, or duct-tape 3 — different variants of what looks like the same underlying system. Nobody at any major AI lab has officially claimed it yet.
Is ducttape an OpenAI model?
Nobody knows. Popular speculation links ducttape to OpenAI as a possible successor to GPT-Image-1, mostly because of its exceptional in-image text rendering. But there is no statement from OpenAI, Google, Anthropic, or any other lab. Any claim of confirmed lineage is a guess.
What's the difference between duct-tape 1, 2, and 3?
On LMArena the ducttape ai model has appeared under multiple suffixes — duct-tape 1, duct-tape 2, and duct-tape 3 — which the community interprets as different versions or parameter presets. Consensus right now is that duct-tape 2 produces the most balanced output across prompt types, while duct-tape 3 is sometimes more aggressive in stylisation. None of these labels are official.
How is ducttape different from Nano Banana Pro or Seedream 4?
In blind pairings on LMArena, ducttape consistently wins on three axes: text-in-image rendering (posters, menus, signage, packaging), layout fidelity when the prompt specifies a structure, and multilingual character rendering for Korean, Japanese, and Chinese. Nano Banana Pro remains stronger for photoreal portraits; Seedream 4 for stylised illustration. See the comparison above.
Can I use ducttape outside LMArena?
Not yet. There is no public API, no consumer app, and no pricing. The only way to try the ducttape ai model today is through LMArena's blind A/B interface — and even there you don't pick the model: you compare two anonymous outputs and vote. If a ducttape-powered generation product launches on ducttapeai.app, subscribers on this list hear first.
When will ducttape be officially announced?
No timeline is public. For reference: the nanobanana codename appeared on LMArena in mid-2025 and was confirmed as Google's Gemini 2.5 Flash Image within roughly six to eight weeks. ducttape has been visible since late 2025, so a reveal window is arguably open — but we are not guessing a date.
The one email

Be first to know when ducttape resolves into a real name.

Subscribers hear the moment the codename is confirmed and the moment generation goes live here — before the tech press picks it up. No other emails. Ever.