00:00Google just launched Nano Banana 2, and they're making some big claims.
00:03Better world knowledge, text that actually works, and character consistency.
00:07Instead of just reading you the blog post, I went through every claim,
00:11tested it, and I'm going to show you what's improved.
00:13Also, how it compares to GPT-1.5, Nano Banana Pro, and the original model,
00:17and give you some of the best use cases for it. Let's get into it.
00:28Nano Banana 2 is the successor to the original model that launched last August
00:33and went on viral for its quality and ease of use natively inside Gemini.
00:37Then came Nano Banana Pro with even better quality, but as you'd expect, slower and pricier output.
00:43With this new release, we are in a rare and exciting situation.
00:46My testing shows that Nano Banana 2 delivers pro-level quality, sometimes even beating it,
00:52but a much cheaper price and faster speed.
00:54Usually, in the AI space, paying more equals better results,
00:58for example, using Claude Opus for complex coding.
01:01But Nano Banana 2 gives us state-of-the-art image generation at a fraction of the cost,
01:07making it the new default for most use cases.
01:09Google is seriously flexing its deep pockets here.
01:12It's a genuine two-way race right now between GPT-Image and the new Nano Banana 2,
01:17at least until we see GPT-Image 2 and Nano Banana 2 Pro in the coming months.
01:21Let's test those claims, one by one.
01:24First off, Google says Nano Banana 2 can tap into web search results to ground images in reality,
01:29just like Nano Banana Pro.
01:31Just a quick note, for these tests, I'll be using Google Gemini and Google AI Studio
01:36to compare the new model to the older versions of Nano Banana.
01:39Let's start with a globally recognized landmark.
01:41So, I asked Google Gemini to generate an image of Taipei 101
01:46with the surrounding buildings in current weather conditions.
01:50Now, for older models, including Nano Banana, that wouldn't be really possible.
01:55The AI would just hallucinate the weather conditions,
01:57because it is not able to look up the factual information.
02:01However, with Nano Banana 2, just like with the Pro model,
02:03we can ask for real-time information, and that should be represented directly in the visual.
02:08In this prompt, I even specified that it should use available images online
02:12and real-time weather data to generate an accurate view of Taipei 101.
02:18We can already see loading Nano Banana 2,
02:21and that means that we are using the latest model.
02:24Beep, boop, boop, boop, boop, boop, boop, boop, boop, boop, boop.
02:26Okay, let's see what we got.
02:28So, I see that the image here shows Taipei in a very warm 24 degree Celsius,
02:34and a weather forecast that doesn't seem right to me.
02:37I need to double-check it against an actual weather forecast.
02:40And yeah, this does not seem right.
02:42In reality, it is 19 degrees and cloudy with a light rain.
02:47However, Gemini hallucinated the weather in Taipei.
02:50So, what has caused this hallucination?
02:53Well, we're using the fast model,
02:55which means Google Gemini is generating an image right away
02:58without spending any time to look up actual factual information.
03:02To solve this, we would need to switch to the thinking model.
03:05So, if I'm in a new conversation,
03:06again, I use the same prompt,
03:10but I switch to the thinking model.
03:13Now, it has actually looked up the real weather forecast,
03:16and we can double-check with the real forecast.
03:19Let's try out the same prompt in NanoBanana,
03:22NanoBanana Pro, and Chagipity.
03:24If I try the same exact prompt with the original NanoBanana,
03:28I get a refusal from the model,
03:30because it cannot look up this sort of information online.
03:33And I get a very generic image,
03:35where the model is essentially hallucinating the weather conditions,
03:39and even hallucinating the source of that information.
03:41If we run the same prompt through NanoBanana Pro,
03:44we get something very similar to NanoBanana 2.
03:47If we run the same prompt through Chagipity,
03:50it writes a Python script,
03:53two Python scripts,
03:54and creates an image.
03:56Okay, let's download the image.
03:57Okay, so it took an actual image of Taipei 101,
04:01it found online,
04:02and it added the weather information in the top right corner.
04:06The information is accurate, so that's good,
04:08but the image does not match the weather conditions.
04:11So, it's not really what we were shooting for.
04:13I'll give Chagipity one more try,
04:15but I will make it explicit that I want it to generate an image.
04:19Okay, Chagipity, on its second attempt,
04:22got something quite a bit better.
04:24However, the seven-day forecast is not actually for the next seven days,
04:29so I'm starting to doubt whether it really looks up the real information,
04:33or hallucinated some parts of it.
04:35Either way, NanoBanana Pro and NanoBanana 2
04:37are the winning models in this category.
04:39If you're enjoying the breakdown,
04:41and would like to see more videos on tools like this,
04:43subscribe to the channel for weekly releases.
04:46Let's give these models a real challenge.
04:48Now, they will need to perform a web search
04:50to look up latest news online,
04:52and generate a magazine spread type of image
04:55showing the latest headlines.
04:57Let's see how these models handle that.
04:59First, I started with Chagipity,
05:01and the results were interesting, to say the least.
05:04Chagipity did the usual AI hallucination type of mistake,
05:08where it generated very generic news headlines,
05:12clearly not something that it looked up online.
05:15So, I gave Chagipity a few more tries.
05:18In its second attempt, I switched to Chagipity's thinking model,
05:21and I told the model that it has plenty of time
05:23to look up the information and get this right.
05:26However, the generated image is again,
05:28extremely generic, no actual news being covered.
05:31A fail for me.
05:32Then, I provided the same prompt to Nano Banana,
05:35the older model from last year,
05:37and it failed in a different way.
05:39It was sufficiently self-aware that it would not be able
05:41to fulfill this request, so it refused it outright.
05:45Let's move to Nano Banana Pro.
05:46I think it should do a much better job with this task.
05:49So, we can see that we have headlines such as
05:51US-Iran tensions escalate,
05:53and Supreme Court strikes down tariffs.
05:55Clearly, recent events and very well prioritized.
05:58Let's move on to Nano Banana 2 and see how it handles this task.
06:02If I provide the same prompt to it in Google's Gemini,
06:05I get a pretty good image with seemingly relevant news.
06:08I need to squint my eyes to see these headlines,
06:11but these are actual headlines.
06:13If you don't look too close, it looks pretty good.
06:15However, it is at a very low resolution.
06:18Nano Banana 2 now supports everything from ultra-wide panoramic
06:21to ultra-tall vertical formats and resolutions from 512 pixels up to 4K for print.
06:27The low-resolution mode is great for rapid iteration,
06:30where you don't need high-res until you have nailed the concept.
06:33However, you cannot choose the resolution yourself.
06:37At least, not through their Gemini user interface.
06:40Here's a pro tip.
06:41You would need to switch to something like Google AI Studio
06:44or third-party services like Replicate or File AI
06:47to control the resolution, aspect ratio, and the like.
06:51I switched over to Google AI Studio
06:52so I could generate high-resolution images with Nano Banana 2.
06:56This time, it created images which are much high-resolution
06:59and they have real headlines, real news events.
07:02However, in text quality, it is very comparable to Nano Banana Pro.
07:07So if you zoom in up close,
07:09you can still see imperfect text rendering and artifacts.
07:12In this attempt, we can see that it used web search to figure out latest news.
07:17However, it failed to correctly represent the current US president.
07:20In the US State of the Union section, it shows a picture of Joe Biden,
07:26which I'm pretty sure is no longer the US president.
07:29That's because, to some degree, it lives in the past.
07:31It has a knowledge cut off of January 2025.
07:34This is getting closer and closer to being production-ready.
07:37We really push these models to the limits.
07:39Both Pro and the second version are clearly better than ChatGPT and the older Nano Banana model.
07:44And if you know what you're doing,
07:46with some prompting and review loops, it is definitely usable.
07:48Just don't expect miracles.
07:50Nano Banana 2 is a highly practical upgrade,
07:53especially for anyone creating content across different platforms.
07:57You can see this across all the examples I showed.
07:59Nano Banana 2 matches Nano Banana Pro in image quality.
08:03The tech magazine spread looked like a real magazine
08:06and the Type A example looked like a photo.
08:09With some prompt engineering and a few more tries,
08:12we could get even better results.
08:13And this is running at flash speed,
08:15meaning you can get these results in seconds, not minutes.
08:19There is no contest between Nano Banana and Nano Banana 2.
08:22The surprising part is that we are getting a state-of-the-art model
08:26for half the price of Nano Banana Pro.
08:29Let's get back to the examples I have shown previously.
08:32There are a few tips.
08:34First, if the model needs to do any thinking about the scene
08:36or look up real information,
08:38choose the Gemini's Thinking or Pro model.
08:40I also recommend mentioning you want the model
08:43to look up up-to-date information online
08:45before generating a visual.
08:47Second, the model now has a dynamic thinking effort setting,
08:50but it is not available through the Gemini interface.
08:53It might be a bit reluctant to perform web searches
08:56and to think through the prompt.
08:58I recommend mentioning that this is an important task
09:00and that the model has plenty of time to get it right.
09:03This classic prompt engineering tip still applies.
09:06Third, if you want a vertical or horizontal image,
09:08specify an exact aspect ratio in your prompt,
09:12such as 16 by 9, 4 by 3,
09:14or one of the new and more extreme ratios,
09:16such as 1 by 4 or even 8 to 1.
09:20Fourth, we cannot control the resolution
09:22through the Gemini chat interface.
09:23So for real production use,
09:25I recommend using Gemini AI Studio
09:27or services like Replicate and File AI
09:30where you can pay based on your usage.
09:32And lastly, the model is better at real-life places
09:36and real-world objects
09:37because now it has access to image search.
09:40Nano Banana Pro had access only to web search.
09:43So here's my take.
09:44Nano Banana 2 is a slight upgrade in quality,
09:47but a massive upgrade considering its price and speed.
09:51The fact that the cheapest model is the best model
09:54just blows my mind.
09:55Previously, I would need to use a cheap model
09:57for a quick iteration
09:58and switch to a better model for the final version.
10:01Now, I can't just use the same model.
10:04The fact that it has web and image search
10:06makes it my default model for image generation.
10:09Does that mean that it's the best model for every task?
10:12Well, nearly.
10:14In my experience, GPT Image 1.5
10:16is still a bit better at style transfer tasks
10:19where it needs to exactly match a provided visual style.
10:22Nano Banana 2 is aggressively closing the gap
10:25in almost every other category
10:26and doing it at a fraction of the cost.
10:29That leaves me wondering,
10:30what will the next generation of models bring?
10:33I expect GPT Image 2 and Nano Banana 2 Pro
10:36to be released in a month or two.
10:39So subscribe if you don't want to miss out.
10:41Meanwhile, the model is live right now.
10:44So go try it out.
10:45Let us know your results in the comments.
10:47Another tool that is exploding in popularity right now
10:50is Cloud Code.
10:51Anthropic recently released Cloud Opus 4.6
10:54and introduced Agent Teams,
10:55which is completely changing people's workflows
10:57and will likely be the new industry standards for coding.
11:00We have made a full video on it over here.
11:03So go check it out.
11:05See you in the next one.
Comments