Vision-based AI models for translating catalogue notes - we test Nano Banana, ChatGPT-Image and othe
Is a catalogue note in a foreign language a problem in 2026? Today I will test whether artificial intelligence can replace a translator and translate subtitles from electronic parts specifications into English. Importantly, the whole experiment will be based on screenshots (bitmaps) rather than PDF files, so AI will have no way to make its job easier. Will this form of specification translation be of any use? Let's find out!
The testing methodology will be very simple - I take a screenshot of the catalogue note and try to translate it into English by sending it as an attachment with a prompt to the AI. I will use the free LMArena website to run the AI models:
https://lmarena.ai/
After testing, I will try to subjectively evaluate and group the models according to their results.
Note - I have placed the images as generated by the AI. If something is cropped, it means that such a bitmap was created by the model.
Test 1 - constant-current LED controller
Inputs:
translate to english
seedream-4-high-res-fal
Such a grand hallucination rather rules out this model.
reve-v1.1-fast
No translations.
reve-v1.1
Major unnecessary reworking of document, residual translations.
chatgpt-image-latest (20251216)
Slightly better, occasional typos. Almost usable.
gpt-image-1.5
Slightly better, occasional typos. Almost usable.
flux-1-context-pro
No translations.
flux-2-flex
Virtually no translations, except for the title above the document.
flux-2-flex-20251231
Translation failure - random letters and stamps, mostly no translations.
qwen-image-edit
This model hallucinates a strange background and is able to spoil the document. Useless.
flux-2-max
Vestigial translations, most are a meaningless string of letters. The headline translated.
flux-2-pro
Nonsensical strings of letters, useless result.
flux-2-pro-20251231
Nonsense strings of letters, useless result.
gemini-2.5-flash-image-preview (nano-banana)
Surprisingly the elder Banana did not want to translate anything.
gemini-3-pro-image-preview (nano-banana-pro)
Best result to date. Text almost all correct, occasional errors and typos, only in paragraphs are some words nonsense.
Trial 2 - display controller
translate to english
gpt-image-1-mini
Model reworked image, renamed layout, not translated.
gemini-2.5-flash-image-preview (nano-banana)
Elder Banana could not cope with the translation.
flux-2-pro
The model has attempted a translation but the result is unreadable, virtually only the title is helpful - LED Driver/Keyboard Scan.
flux-2-pro-20251231
Flux 2's primary keywords decrypted, but the rest are useless.
flux-1-context-pro
This model has superimposed the inscription translate to English on the image.
flux-2-flex-20251231
Residual translations.
gpt-image-1.5
At first glance very good, but the introduction from the second/third sentence onwards fell apart.
reve-v1.1
The basics translated, but also damaged the lead-in diagram.
seedream-4-high-res-fal
The title may be translated, but the model has added some strange background.
chatgpt-image-latest (20251216)
Like the second one from OpenAI, it's not bad, only the introduction fell apart afterwards. In addition, I see a slightly damaged lead-in diagram.
gemini-3-pro-image-preview (nano-banana-pro)
Nano Banana Pro has again performed very well.
Trial 3 - synchronous rectifier
This time a trial with a screenshot:
translate to english
qwen-image-edit
Useless result.
chatgpt-image-latest (20251216)
The basic translation is there, but with lots of typos. Synchornous?
gpt-image-1.5
Same as the previous GPT.
seedream-4-high-res-fal
This model has redone the background again....
gpt-image-1
Residual translation. In addition, with another attempt I received a strangely cropped image.
gpt-image-1-mini
And here what happened? A short circuit? And this is between two separate diagrams.... in addition, the model also cropped the picture.
flux-2-flex
Again, residual translation.
gemini-2.5-flash-image-preview (nano-banana)
No translation.
seedream-4.5
It came out slightly better this time, but there are still shortcomings.
flux-1-context-pro
No translation.
flux-2-pro
The letters have been changed, but they don't make sense?
gemini-3-pro-image-preview (nano-banana-pro)
bb558ffb3
Another success story for Nano Banana Pro.
Final ranking of video models
I did additional tests, but did not put any more images in the topic, because the content with several of the same nonsense graphics would be unreadable. In the end, I grouped the models according to my overall feeling, although I noticed that occasionally a particular model might do better or worse - probably the generation has some randomness factor (seed - so called).
Right translations, occasional errors:
- gemini-3-pro-image-preview (nano-banana-pro)
Almost acceptable translations, but problems with some words, blurring of letters:
- chatgpt-image-latest (20251216)
- gpt-image-1.5
Sometimes it explains something, sometimes it hallucinates and creates nonsense:
- reve-v1.1
- gpt-image-1
Bare translation attempts, meaningless letter composition:
- flux-2-max
- flux-2-pro
- flux-2-pro-20251231
Something tries to translate, but hallucinates and rearranges images:
- seedream-4-high-res-fal
Can spoil the image:
- qwen-image-edit
In summary , only the latest Nano Banana Pro seems to give acceptable results in terms of translating images from the catalogue notes, although it still happens to have artefacts. Just behind it is still GPT-Image 1.5 and ChatGPT-Image (20251216), but it is no match for it. The rest of the models are useless, although some of them try to remake the image and some ignore the text completely.
There doesn't seem to be much left to do with AI in this context. It seems to me that as early as 2026 there will be much better models that can handle such translations even better, and even if not, the Nano Banana Pro is still satisfactory.
Do you see a use for artificial intelligence in the role of an image translator? Or do you know of other practical applications for the Nano Banana Pro and similar models?
p.kaczmarek2 wrote 13701 posts with
rating 11511 , helped 623 times.
Been with us since 2014 year.
Comments
And so very good for a machine. It will still take some time before it reaches perfection. Nevertheless, you can read the information you need from such translations, definitely better than Chinese "bushes".... [Read more]
Artificial intelligence is learning by the day and I suspect that by the end of this year there will be a visible difference in this subject. It wasn't so long ago that this was black magic, and now peasants... [Read more]
Mi chatgpt 5.2 still did things differently: https://obrazki.elektroda.pl/7460791500_1767348997_bigthumb.jpg And this one for English and pseudo Polish: https://obrazki.elektroda.pl/8048532500_1767349615_thumb.jpg... [Read more]
Nano banana the best, as usual. With hundreds of pages, however, it is best to do OCR from such a document, i.e. get the Chinese, and then translate yourself with the language model, already without... [Read more]
In my opinion, artificial intelligence will also lead to a standard in China for the use of English in electronic component documentation. By a circuitous route, but nevertheless this is the aspect I see... [Read more]
The idea of a global language has been around for quite a long time, so far nothing has changed. The indicated documentation can be released in English in parallel, now they got a tool where it is done... [Read more]
@fachman1964 my sense is that with them there is often vestigial and untranslated documentation, even when they make it available. In the SDK for Beken and other IoT chips it is similar. @MikeC I think... [Read more]
@gulson It is not a global language, but the equivalent of a technical drawing, i.e. a language that is universal in its assumptions and understood by everyone regardless of the mother tongue used. English... [Read more]
What I am wondering is what is the actual cost of such a reworking of one image by Nano Banana Pro. Say, this situation from my presentation. Has anyone seen such information somewhere? A quick web search... [Read more]
@pkaczmarek2 Any rates, even those actually paid, are not authoritative. Here, the decisive factor is something other than the actual cost. The end-user price is the result of an activity that has as... [Read more]
Something seems to me that in the case of Asian 'bushes', AI rather searches on the manufacturer/other sites for a ready-made English translation...Mainly the complete ones.... [Read more]
It translates from these bushes without any problem ... [Read more]
Test with Polish. translate all to polish https://obrazki.elektroda.pl/7213900200_1767652236_thumb.jpg gpt-image-1.5 https://obrazki.elektroda.pl/1381551100_1767652255_bigthumb.jpg... [Read more]
Why do images two and three have the left and right sections cut off? [Read more]
All in all, I think we've talked about this before, but this is how the AI generates. Below is a short video of what it looks like to me: gpt-image-1.5 cuts off images for me always, no matter if... [Read more]
I encounter such a problem in cases of incompatible settings or incompatible web browsers themselves. Parts of web pages sometimes get cut off as a result, and this happens especially with unusual settings. Can... [Read more]
The website address is the same all the time, as in the first paragraph: Somewhere in the fifteen years I've been doing the frontend and backend, and I haven't encountered the browser truncating... [Read more]
This could be linked to the frame size of the item depending on the screen size and then automatically adjusting the dynamically generated image. Sometimes a slightly unusual web browser or some setting... [Read more]
If the attached images are exactly what AI got, then I'm not surprised. Shit in => Shit out. [Read more]