Nano Banana Pro claims usable white-box product shots in Ask HN commerce tests
An Ask HN commenter reports Nano Banana Pro turned iPhone product photos into usable white-box images and Amazon-style infographics. The evidence is anecdotal, but it suggests ecommerce asset generation may be near replacement-level quality for some tasks.
TL;DR
- In a high-signal Ask HN thread about GenAI "oh shit" moments, one commenter said Nano Banana Pro turned a bad iPhone product photo into a usable white-box shot and Amazon-style infographic images, enough to replace a photographer and designer for that narrow ecommerce task.
- That claim is anecdotal, but Google's GA announcement and Gemini's image-generation docs do position Nano Banana Pro as a professional asset model for generation and editing, with complex instruction following and high-fidelity text.
- The timing matters because the ChatGPT Images 2.0 HN thread and OpenAI's gpt-image-2 launch post show the same race moving toward infographics, layouts, comics, and other text-heavy production visuals.
- According to the discussion digest from the same Ask HN thread, the broader thread was full of similar threshold-crossing stories in coding and technical reasoning, which makes the product-photo anecdote feel less like an isolated flex and more like part of a wider shift.
The interesting bit here is not a benchmark chart. It is that a random HN comment about image generation for commerce describes a workflow creative teams actually pay for. Google now says Nano Banana Pro is built for professional asset production, while OpenAI is pitching gpt-image-2 for ads, product flows, and infographics. The weirdly specific white-box-photo test is exactly the kind of job that exposes whether these models are still toy generators or something closer to a usable production layer.
White-box product shots
Ask HN: What was your "oh shit" moment with GenAI?
For creatives, the standout signal is that image models are crossing from demo quality into production-adjacent usefulness. One commenter described using AI-generated product photography and ecommerce infographics well enough to replace work that would otherwise have gone to a photographer and designer.
The strongest claim in the evidence pool comes from that Ask HN thread, where one commenter said Nano Banana Pro was the first image model that could turn a "crappy iPhone product photo" into a usable white-box shot. For ecommerce teams, that is a much tougher claim than "pretty images," because catalog photos fail on small errors: warped geometry, fake reflections, bad edges, wrong shadows, or packaging details that drift.
Google's own framing is narrower but directionally consistent. Its GA announcement says Nano Banana Pro is for integrating high-quality image generation and editing into production workflows, while the Gemini API docs call it the Gemini 3 Pro Image model for professional asset production.
Amazon-style infographics
ChatGPT Images 2.0
For creatives, the update is about a stronger image generator that handles text, layout, multilingual content, comics, and other structured visuals better than before. The discussion also touches on authenticity and style appropriation concerns, which may matter if you use these tools in production or publish AI-generated work.
The same HN commerce comment did not stop at cutout photos. It also said Nano Banana Pro could produce Amazon-style infographic images well enough to replace design labor for that task. That matters because ecommerce infographics combine several failure modes in one asset:
- product fidelity
- readable text
- structured layout
- visual hierarchy
- brand-safe composition
Those are exactly the areas the current model wave is trying to close. The Gemini docs say Nano Banana Pro uses advanced reasoning to follow complex instructions and render high-fidelity text. On the other side, the ChatGPT Images 2.0 HN thread links to OpenAI's rollout, where the gpt-image-2 announcement promises better layouts, stronger multilingual text rendering, and structured outputs like diagrams, infographics, charts, posters, and comics.
The useful read is simple: text inside images is no longer a side quest. It is becoming the feature that determines whether an image model can move from moodboards into listings, ads, packaging comps, and storefront assets.
The anecdote fits a broader threshold
Discussion around Ask HN: What was your "oh shit" moment with GenAI?
Thread discussion highlights: - dang on workflow acceleration: Watching it do log file analysis in seconds that would have taken me hours, helping with optimizations I’d been putting off for years, tracking down concurrency bugs, and finding information I couldn’t locate with Google. - bluejay2387 on local model coding: A locally hosted model wrote its own semantic search system over 250,000 files and then built a working game mod in under 4 hours, which was enough to freak the author out. - rerdavies on technical reasoning: Claude implemented not just the cited equation from a Spice manual, but also a more complex Lagrangian calculation and symbolic partial derivatives that the source text did not spell out.
According to the discussion summary, the same Ask HN thread also surfaced people using models for log-file analysis, optimization work, semantic search over 250,000 files, game-mod building, and symbolic math that went beyond the cited source material. The common pattern was not perfection. It was that users kept describing moments where a model became good enough to substitute for a chunk of paid or delayed work.
That makes the Nano Banana Pro commerce anecdote more interesting than a one-off product boast. It showed up in a thread where people were reporting capability crossings in very different domains, and the commerce comment framed image generation the same way: not as inspiration, but as replacement-level output for a specific production job.
One caveat still holds. The evidence here is a single commenter, not a controlled test. But because Google's own materials already claim editing, professional asset production, and high-fidelity text in the Cloud post and the Gemini docs, the HN report lands as a plausible early field test, not a category error.