Nano Banana 2 - Smaller, Faster, Cheaper

Nano Banana 2 - Smaller, Faster, Cheaper

Machine-readable: Markdown · JSON API · Site index

Поделиться Telegram VK Бот
Транскрипт Скачать .md
Анализ с AI

Оглавление (2 сегментов)

Segment 1 (00:00 - 05:00)

Okay, so Nano Banana 2 has dropped and this is a story about smaller, faster, cheaper, and almost as good quality as Nano Banana Pro. So, let's jump into this. So, what has actually dropped is the Gemini 3. 1 flash image. It's been called Nano Banana 2 basically because the first model that came out was actually a Gemini 2. 5 flash image model. And then of course Nano Banana Pro that we had that came out back in November was Gemini 3 Pro image. And generally the story here is that you're going to get a better quality image model than what you got with the Flash 2. 5. And often that's going to be very close to equal with the Nano Banana Pro model. Now, very quickly, I'm just going to go through some of the key things that it can do. First off, you've basically got a model that does the reasoning over your image. So, this is just like a NA Banana Pro in here. You can see here I've asked it to basically create an image of a cat holding a sign saying on strike, not catching mice today. And one of the things that you will find is that this model is extremely good at text. So if you look at it, it gets all the text there fine. If I ask it then to add in some manunes, it gets that fine. No problem. And these are two of the main sort of features that they've focused on this whole idea of basically getting precise instruction following as well as precise text rendering and translation. So you can see from there we can basically ask it to give the other two cats some signs. We're getting all the text correct here. And we can see that there's a toy mouse in there. When I ask it to make the mouse real and holding a sign saying I support the cats, sure enough, we get a real mouse now holding the sign saying I support the cats. And you can see this is a very nice quality image overall. making use of the translation and stuff like that. I can ask it to translate all the signs into Italian and then from there I can even ask it to translate all the signs into Thai. Now it does seem to me that certain languages have a lot of different fonts and will sort of adjust to the original font that you used more than other languages. So this one we've definitely lost the sort of handwritten elements that we had perhaps in the early English version. Now, that said, I didn't prompt for that in any way, so make of that what you will. If we compare it to the Nano Banana Pro, there are times where it comes off really good, and there are other times where it's not quite as good. So, this is one of the ones that I showed from the original Nano Banana Pro video where I asked it for an unremarkable, unintentional shot, a selfie with a caveman with a T-Rex running behind him. That certainly looks the part there. Ask it for a top view. We certainly get the top view. I ask it for what's going to happen next and it can sort of logically work this out. If you use the exact same prompt for Nano Banana 2 here, we definitely get a lot of the similar kinds of thinking, but the image that it creates perhaps is not quite as good, right? Like this is not from the perspective of the selfie if we can actually see the camera in his hand. And then when we ask for the top down image, we can sort of see like, okay, it's kind of got it the idea right, but now T-Rex is right next to him. And then when I ask it for make a photo of the likely outcome, it does seem like here it's okay. It's definitely much more on track. Now, the look of these two images is very different. And I didn't use any sort of description of the environment or anything like that. So, if you want to get the best results, you definitely want to make sure that you're using that kind of thing. All right. What else do you have in here that actually makes this really good? You've got sort of full aspect ratio support like we did in the Pro model. We've got resolution going down now to 512 pixels. We've got our two thinking levels in here. But the other thing that we've also got is this ability to access tools. So we've got the grounding with Google search tool which Nano Banana Pro has. And then we've got a tool that actually Nano Banana Pro doesn't have which is image search. So you can condition it on getting it to do an image search and then using that in the actual grounding and control in here. Another feature where this model is really good is that it's gotten a lot better at character consistency. So you can see here we've got three input images and then being able to generate multiple sort of images or a story out based on these particular characters. And building on this even more, you can actually now provide up to 14 different reference pictures for your generation. So you can see here we've got a farm. So we've got all these different characters for the farm and some trucks, a tractor, and the actual barn. And it's able to take all of those reference inputs and use them in this

Segment 2 (05:00 - 06:00)

generation. So, this is a big win for if you want to do anything with products. things where you've got multiple references of products in the same product range and you want to actually have all of them in the picture, Nano Banana 2 actually allows you to do that natively. Now, so the model itself is not as cheap as the old version, but certainly not as expensive as Nano Banana Pro. And it's going to be rolled out in a bunch of the different Google apps. So that's everything from like the Gemini app and then also things like anti-gravity AI studio of course and you're going to have full support on Google cloud obviously in Vertex AI etc. It's certainly a win if you were sort of looking at using Nanabanana Pro in production but it was just too expensive to do at scale and stuff like that. This one definitely is a win like that. And also the big feature I'm finding playing with it is just how good the text rendering is that you're able to get all these different generation styles. So obviously with the release of this, it's also got to bode well for the people who are waiting for Gemini 3. 1 Flash model, but I guess that's something we can talk about on another day. Anyway, have a play with the model. Let me know in the comments what you think of it, where you're finding it to be strong, not as usable perhaps as the Nano Banana Pro model. All right, as always, I will talk to you in the next video.

Другие видео автора — Sam Witteveen

Ctrl+V

Экстракт Знаний в Telegram

Экстракты и дистилляты из лучших YouTube-каналов — сразу после публикации.

Подписаться

Дайджест Экстрактов

Лучшие методички за неделю — каждый понедельник