07:32
AI image generation
Nano Banana Pro (Gemini 3 Pro Image) is a major leap over the original Nano Banana for product photography, delivering 4K resolution, dramatically improved label-text accuracy, and near photo-realistic results across a range of difficult product-photo tests, according to Thomas's first-impressions testing.
Nano Banana Pro (also called Gemini 3 Pro Image) generates 4K resolution images and, per the video, finally produces accurate label text.
A direct side-by-side test reusing the exact same hot sauce bottle image and prompt from a prior Nano Banana video shows the Pro version has far more fine, crisp label detail.
A hand-holding-the-bottle test shows much greater photorealism than the old generation, though the video notes the smallest label text details still aren't quite correct.
A water tank splash test with a pink Long Drink bottle produces realistic water physics and product integration, with only a minor error: the 'V' under the percentage sign looks 'a bit tired.'
A Feastables chocolate bar test (a graphically complex label) placed in a peanut-butter snowman environment comes back with text and graphics almost perfectly intact, with only a slight error on the fair trade stamp.
A pink skincare product with a full paragraph of white text on a pink background is rendered correctly across multiple generations, which the video calls something that 'would be a nightmare for the models before Nano Banana Pro.'
Adding more bottles/water elements to the skincare scene introduces some small errors, suggesting compositional complexity is where mistakes creep in.
A Versace fragrance bottle test (complex product shape) comes back spot-on, and remains spot-on even when regenerated as a row of five bottles in the same environment.
A YouTube thumbnail generated from four reference images plus a simple text prompt produces a result the video calls mind-blowing.
Nano Banana Pro is available through Google's own Gemini and Google AI Studio, but Thomas personally uses it via the Higgsfield AI platform.
The video is sponsored by Higgsfield, which supplied 10 free Ultimate Plan promo codes given away in the video and is running a promotion for unlimited Nano Banana Pro generations for a year if subscribed before the end of that week, plus a standing 10% discount code.
Errors appear to scale with compositional complexity (adding more bottles or elements to a scene) rather than purely with label complexity, since even the hardest single-label tests (Feastables, skincare paragraph) come back nearly perfect until more objects are added.
Even the video's most impressive results retain small residual errors (the tired 'V', the fair trade stamp, the hand-holding label details), indicating the model is described as vastly improved rather than literally flawless.
The video's methodology of reusing identical prompts and reference images from a prior Nano Banana video is an implicit controlled comparison, isolating model improvement from any change in prompting.
Thomas frames remaining errors as fixable via 'a couple of generations or just a slight touchup,' implying the practical workflow is iterate-and-cherry-pick rather than expecting single-shot perfection.
«Today we're testing out Google's new AI image generation model called Nano Banana Pro that is capable of generating 4K resolution images and finally accurate label text.»
— 00:00
«the visual reasoning of the Gemini 3 Pro model is absolutely insane.»
— 00:31
«This is actually insane.»
— 00:40
«some of the smallest details on the text label isn't quite correct, but the result is very crisp and impressive.»
— 02:28
«looking at this result that is spot on.»
— 03:09
«text should read V. And right now the V looks a bit tired.»
— 03:20
«I have to say this one is next level impressive.»
— 04:03
«these generations, my head is absolutely spinning.»
— 04:58
«holy smokes, this is on a completely new level of detail.»
— 05:24
«they are all spot-on perfect and there are no errors on the product whatsoever.»
— 06:03
«oh my god, the fact that it was able to generate this thumbnail from these four images and a simple text prompt really blows my mind.»
— 06:31
Reception
Viewers appreciate the practical tutorial and engage enthusiastically with the tool, but multiple technical pain points—particularly text rendering quality issues—create a mix of admiration and frustration.
This is a sponsored first-impressions demo rather than an independent benchmark, showcasing curated best-case outputs on hand-picked products with heavy promotion of sponsor Higgsfield throughout; audience comments suggest results are less consistent on other products and labels than the video's examples imply.
07:32