back to indexAI in Action 5 Sept 2025: Google's Nano Banana - Deep Dive & Hands-On Demo

Chapters
0:0 Introduction & Context
3:33 Getting Started with Nano Banana: Logo Generation
9:15 Character Creation: The "Artie" Tests
15:37 Coherence and Consistency
25:35 Advanced Features: Pixel Art and Character Sheets
46:55 Limitations: Image Quality
53:39 Limitations: Safety Settings
55:10 Real-world Applications & Conclusion
00:00:01.840 |
I can't hear you. Hello, hello, hello. How's it going? I like your shirt. 00:00:09.120 |
Yeah, thanks. Oh yeah. Wait, did I? I did. Okay. 00:00:16.240 |
Yeah, I, um, the nano banana stuff is crazy. Okay, I know nothing about it. So if you want to, 00:00:26.640 |
like, tell people what this thing is and all that when we get going, like, yeah, for sure. 00:00:32.640 |
Oh, my goodness. I didn't talk, you know, who was that earlier that mentioned he's, he's a 00:00:40.240 |
presented before Scott. So Scott mentioned earlier about him doing the nano banana thing. I don't know 00:00:45.440 |
if you got a chance to check that out on his page. But yeah, I kind of I can do like a walkthrough of 00:00:52.640 |
what he's done with his team page. And then like in the discord itself, like in the actual AI in 00:00:59.200 |
action, I mean, a latent space discord server. We have a ton of examples of why nano bananas crazy. 00:01:08.480 |
But I guess this could mostly be a nano banana thing. I tried to make an effort to get like us to a good 00:01:17.760 |
place for the the vibe coding a bot a discord bot and it's like extremely complicated. 00:01:24.240 |
But I think I have most of the details worked out minus the server thing and we can kind of like maybe 00:01:30.240 |
talk about that as part of the AI in action this Friday. 00:01:38.000 |
yeah. Okay, you're on the job. You're on the job. 00:01:44.800 |
It's all good. I mean, I'd love to not have not be on the job. Community run. Let's go. 00:01:50.480 |
Yeah, this is that man. Well, this is how I try to run our team, too. I'm just like, 00:01:58.320 |
all right, you got you all tell me what you need to do. 00:02:00.640 |
Right, right. I want to step in as well. Like, okay, let me prompt a couple. 00:02:04.640 |
We don't. Yeah, we don't need to do anything. We just like 00:02:12.000 |
So I guess to give everyone an introduction to what nano banana is, it is a new image generation 00:02:21.600 |
model from Google. And I say new, but I think it's been out for a few weeks or so. How you actually 00:02:27.600 |
access it, there's a couple different ways to access it, how I've been doing it. And I think 00:02:31.520 |
I can show a pretty cool example if I share my screen. 00:02:34.960 |
Oh, that's going to show you guys. If we share, we can move you guys over here. 00:02:43.120 |
So I talked about some of the process I was going through to set up 00:02:54.400 |
the test ding discord server or whatever. And I was like, Oh, this is be a great like 00:02:59.360 |
one of the first steps is like name your bot and give it a picture. 00:03:01.520 |
And I was like, Oh, this would be a great, you know, opportunity for me to see 00:03:04.880 |
how nano banana performs. So I, of course, go to, you know, if I maybe if I start from scratch, 00:03:11.040 |
it'll be better, but AI studio dot Google.com. And again, that's for the transcription, I guess I 00:03:16.240 |
should say AI studio dot Google.com. And then over the on the right hand side, when you load it up, 00:03:21.920 |
oh, they have a what's new button here that you can try nano banana. But if that's gone in the future, 00:03:26.720 |
for anyone watching the recording, if you click over here on the right hand side, and select a new model, 00:03:32.160 |
of course, if you go to all there's like Gemini, Gemini 2.5 pro and flash and all these great models 00:03:36.800 |
that we've gotten to know, but nano banana is their image generation model. And it's technically the slug of it. 00:03:42.720 |
if you're looking to use it as a developer in the API is Gemini dash 2.5 dash flash dash image dash 00:03:51.520 |
preview. And that dash preview simply means that Google is going to screw us over at some point in 00:03:55.440 |
the future, and we'll have to change the slug. So don't get attached to it. But but Gemini 2.5 flash 00:04:01.120 |
has been trained to be an image generation model. 00:04:03.920 |
Wait, so is this the same as nano banana.ai or different? 00:04:07.840 |
I think I think someone's like been opportunistic with that domain name and took it. 00:04:13.520 |
Um, if I'm if I'm not mistaken, nano banana.ai I remember seeing some like, if anyone can do 00:04:21.040 |
do research on this in the background, I want to say looks context. Yeah, I want to say it coming soon. 00:04:30.720 |
I want to say this is like not officially from Google, and that somebody snatched the domain name once they 00:04:37.040 |
figured out that that Google was releasing a new model under the same name, and kind of like ran 00:04:42.000 |
with it. So I believe that's what's happened here is like this is not actually officially from from 00:04:46.080 |
Google, they just kind of like played on the name of the new model. 00:04:49.440 |
Yeah, I see the copyright on, you know, banana 2025, not Google, the alphabet company thing. 00:05:02.320 |
So and that's funny, because like, I'm assuming cable, you like Googled it. And this is the one of the first 00:05:06.720 |
That's what that's what came up their privacy policy says it's done done AI limited from Hong Kong. 00:05:23.520 |
So yeah, I think it was just like, there was some hype around it. And like a lot of the new language 00:05:31.760 |
models nowadays, as far as the rollout goes, as far as I understand, they basically started testing this 00:05:38.000 |
image preview, or this image generation model on other services, like a flux or not a flux a um, 00:05:43.760 |
oh my god, what's the name of the, there's a bunch of companies that do you know, image generation. 00:05:48.880 |
And I think Google tested it with those services first, before they released it on their own 00:05:53.440 |
platform, I think is how word got out, right? Like, so it leaks, and they're like, oh, this is code name, 00:05:59.360 |
nano banana. And then somebody like went and registered the domain name, and kind of like, 00:06:03.200 |
got ahead of the curve, I think is what happened. But officially, officially, okay. 00:06:10.320 |
You broke up a little bit there from me, what would you say? 00:06:28.400 |
Yeah, yeah, no, you're in it. I understand. It's all good. If you can type that in, 00:06:33.920 |
though, because it sounds like you were saying something about the process. But officially, 00:06:38.720 |
what you can do is go to Google, like, you know, Google AI studio, and then find the nano banana 00:06:43.280 |
model, and then type your prompt in. And it's much faster than I expected it to be. And the quality of 00:06:51.200 |
it's been great. So I don't know if you guys have like, if anyone here in the audience has like a 00:06:54.880 |
prompt or whatever that we can kind of throw into nano banana, and see what it comes up with. I just, 00:07:00.720 |
what I ended up doing was, you know, tell it that I was trying to create a testing server 00:07:04.800 |
for a weekly meeting called AI in action. And this is like an offshoot of the latent space podcast. 00:07:11.760 |
And I asked it for a logo, and it actually came out really good compared to other services that I've 00:07:17.600 |
tried because a lot of these super small intricate details end up getting jacked up. Like even here, 00:07:22.480 |
once we zoom in, you can see that there's just like some inefficiencies and like, 00:07:26.880 |
between what a model generates and like what a human would generate, if they're working with like 00:07:31.920 |
vectors or whatever in a program like Adobe Illustrator or Figma or one of these other, 00:07:43.840 |
you know, design applications. But this is like, really, really cool, especially for something that 00:07:49.040 |
was done in like, you know, a couple seconds. So what I what I did was like, it really misunderstood 00:07:54.080 |
the assignment, or maybe I just, I did a terrible job of prompting it, because I wanted a logo for the 00:07:59.280 |
testing discord server and not the official discord server of AI in action. So I followed up with my 00:08:03.520 |
prompt and said, Hey, that's, that's great. But I also need a logo for the AI and testing test 00:08:07.440 |
discord server that I'm setting up. It's meant to host some of the discord bots we were testing as a 00:08:11.680 |
community. So it responded by editing the original image by putting a bug in an in a spy glass here, 00:08:18.400 |
or whatever this is, a magnifying glass, and, and then called it AI and testing about testing server, 00:08:24.240 |
which I thought was crazy. And then so I was like, Well, you know, we're also coming up with this 00:08:30.560 |
fictional character driven by an LL empowered agent named Artie short for artificial. And I asked for 00:08:36.480 |
some help with creating another image for for that character. And it came up with this, which again, 00:08:43.200 |
is like insane, like some of some of the intricate details, if you even six months ago, tried this with 00:08:48.640 |
some of these models. I don't know if you guys have been playing with image generation over the course of 00:08:53.120 |
this, like, you know, AI language model explosion, but like these super small details, like the lines almost 00:08:59.040 |
always end up getting messed up. So the fact that it could get this detailed and have, like, have some 00:09:04.400 |
kind of coherence from logo to logo to logo, it's also unheard of, it's very hard to pull off. As someone 00:09:10.880 |
that's been using image generation models for a long time. But yeah, that that is nano banana. And there's 00:09:16.880 |
a ton of examples in the discord, I can kind of pull up, if you guys have any comments, I can maybe pull up some 00:09:21.680 |
of the cool ones that I saw in the discord. But that in and of itself is nano bananas. It's a new image 00:09:27.680 |
generation model from Google. It's been kind of going viral over the last maybe a few weeks or so. 00:09:32.720 |
Oh, the one example actually can pull up before we maybe turn it loose here is Scott Werner from the 00:09:40.400 |
discord who is also I don't know Scott's on the call, but he has presented many times in an action. He did a 00:09:47.040 |
really cool thing on his team page for his company. So if you go to the sub layer website, sublayer.com/team, 00:09:57.040 |
he basically takes, I'm not sure how he's pulling this part off like what API endpoint he's using to pull 00:10:04.400 |
up today in history. Let me zoom in just a little bit more. But on this day in history, on September 3rd in history, 00:10:13.440 |
it says, in 1967, Sweden's famous Dagen H or H day smoothly switched traffic from driving on the left to the right 00:10:21.440 |
side of the road, a massive national effort that improved road safety. I don't know why this is Lego related, but they, oh, okay, sorry. So this is the actual image generation prompt I've highlighted here to paraphrase. It just says, 00:10:27.440 |
re-imagine this person as a smiley Lego style minifigure traffic officer in 1960s Stockholm with a bunch of other instructions. But what they're doing is like automatically generating both of these, as I understand it, the today in history and the generation prompt, and then feeding that to now Nano Banana to take their original team images and switch it over to, you know, the image that was 00:10:49.440 |
inspired by these prompts. And he said previously they were using the open AI model, but Nano Banana is not only cheaper but faster. And they're playing around with giving this functionality to everyone in their Discord server as like a team. So I thought that was a pretty cool use of Nano Banana and almost like a testimonial because like it's cheaper and faster for them. 00:11:07.440 |
right now I think is what he said in the in the in the Discord server. So I don't know if you guys have any other examples that you guys have seen or want to pull up but that's that's Nano Banana as I understand it. Me and Swix have been going back and forth in the Discord server just like sharing crazy links of like what people are able to pull off. There's some really cool like, 00:11:33.440 |
3D stuff that I've seen when in a banana. But anyways, it's a cheap image generation model is kind of like what I've landed on here. 00:11:43.440 |
And the other thing I guess I could kind of show off, I may need a second or two for this one. 00:11:51.440 |
But if you guys have any, actually, I have not the entire time I haven't checked the chat. So maybe I should check that first. 00:11:57.440 |
So what might be interesting actually would be to do so like take a test example and like there's questions about how, how long can it cohere for how good is that whatever can we like crowdsource with this group like a set of things we might want to do and see coherence on and like. 00:12:19.440 |
See how it does like for example, I'm imagining like let's start with from a logo or something like this, but then say, Okay, you know, now generate me a background for a bunch of social media marketing integrate this integrate that okay now go and you know generate a web. 00:12:37.440 |
You know some some additional assets for web maintaining this style, that type of thing and just see how well it does it at keeping coherence. 00:12:46.440 |
Yeah, I think like so like a almost like an AI in action eval for like image generation and how long it can keep coherence. 00:12:56.440 |
Yeah, just because I'm seeing, I'm seeing questions here about like how well can it keep coherence and I think that is one of the, you know. 00:13:05.440 |
Big areas I think yikes raises a character consistency thing so like if we put a character in there will that character maintain consistency because that's a little bit more detail so yeah I would love to. 00:13:16.440 |
Unless we have other things, but if we want to play with a thing that's a that's a dimension i'd love to test for sure character so like, can we see maybe because like i've actually seen prompts to this let's see if I can download the already had it downloaded. 00:13:31.440 |
Let's go ahead and let's go ahead and throw in because you can add images let me drag and drop from my downloads folder. 00:13:41.440 |
You know something to the effect of this, how about this. 00:13:46.440 |
We're in 2025 so maybe I shouldn't be doing too much. 00:13:50.440 |
This is my character named already short for artificial. 00:13:56.440 |
It's not working testing testing oh it is okay sorry user air. 00:14:05.440 |
This is my character already short for already official, and I want you to put him in a spaceship hovering over the earth. 00:14:19.440 |
It's just reading you as having a southern accent or something. 00:14:45.440 |
Well, and then yeah we should like create a whole storybook or something like that and just see like. 00:15:02.440 |
Take already on an adventure starting in his spaceship and then exploring a new planet. 00:15:22.440 |
So that level of coherence I just never seen in a. 00:15:31.440 |
I would like to see already from a slightly different angle. 00:15:52.440 |
Adobe, I think where they would, you'd like, have a. 00:15:58.440 |
Let you vectorize it and then rotate it and see it. 00:16:07.440 |
It is a different angle, but I don't know about. 00:16:25.440 |
you tried to get like a glass of wine to be completely full. 00:16:32.440 |
And like, no matter what you prompt it like will not do it. 00:16:36.440 |
I think it's better now, but like, this is what's writing. 00:16:39.440 |
Let's show already from the side drinking wine. 00:16:42.440 |
I wonder where the rate limits come in on this model. 00:16:54.440 |
I should have said a full glass of wine missed opportunity. 00:17:18.440 |
That's, that's what some of the examples have showed the most. 00:17:31.440 |
The comments say, if you give me already, I can run one concurrently. 00:17:39.440 |
Um, if you go into the discord server, uh, AI in the AI in action channel. 00:17:44.440 |
I don't know if you post up and actually while I'm doing that, I'm just going to post a link to it in the thing. 00:17:54.440 |
Uh, so, um, if you want to download already yourself, you can go to that link. 00:17:59.440 |
I just posted in the chat and, and like, you know, use, uh, nano banana on your side. 00:18:04.440 |
And then if you look at the top right-hand corner, you'd have to find the nano banana model. 00:18:09.440 |
Um, any other like, uh, test prompts that you think would, would be cool. 00:18:20.440 |
So I'm literally going to, I'm going to type that with no other context. 00:18:30.440 |
I think it was like a fancy name for a different view. 00:18:32.440 |
And I just want to see him like, see if they can like generate a different, completely different view, doing something completely different. 00:18:42.440 |
What, what might be worth the effort here is for me to start over. 00:18:47.440 |
Um, it's, it's, it's, it's a start over and then start with the original image and see if you could do it better. 00:18:57.440 |
Cause they still have them on this random planet that we're now exploring. 00:19:06.440 |
I actually don't know what view isometric is, but it looks, it's. 00:19:13.440 |
The fact that they've kept already intact this whole time is what to me is impressive. 00:19:17.440 |
Cause like before I think that was not a thing. 00:19:49.440 |
Um, so this is prompted by, uh, Oh, don't kill me if I, if I'm not pronouncing this right. 00:20:00.440 |
Cause you knew I was about, I was about to mess that up. 00:20:04.440 |
This character wielding a lightsaber and shooting forest lightning at a sit. 00:20:16.440 |
Just to give a, we were about 20 minutes in now. 00:20:26.440 |
I have created a test discord server that I can like invite everyone to, or I don't know 00:20:33.440 |
This is something I think could maybe help me with. 00:20:35.440 |
Cause I've done zero moderation, actual moderation of, of discord servers. 00:20:39.440 |
But I created a discord server called AI and testing as opposed to AI in action. 00:20:44.440 |
And I think we could do a lot of testing here in the server for the discord bot. 00:20:48.440 |
Like every time we have an idea where we want to vibe code a thing, we could put it in this server. 00:20:54.440 |
Um, I, I, I think there was already one created with the Zach and, um, and I forgot my boy's name. 00:21:00.440 |
Uh, who, who created that the actual original discord bot, but I don't, I don't, I didn't have access to it. 00:21:06.440 |
I'm not sure. I couldn't remember off the top of my head how to find it. 00:21:08.440 |
Um, so I, I figured I'd just, uh, make a new one cause they're fairly simple to make. 00:21:12.440 |
And with this, I've also invited, um, and already bought. 00:21:20.440 |
I think the one thing, um, that I, the one issue that I run into, cause I did figure out how to get, uh, the discord keys. 00:21:30.440 |
You had, you need a bis, a discord bot token and a discord guild ID. 00:21:34.440 |
The one issue that I ran into was like the whole server thing and how I actually hosting the code that runs, um, you know, the services that, that kind of like power the discord bot. 00:21:45.440 |
But I, I would like to do some more testing and figure out like how, how to create a good platform that, uh, the community can have access to, uh, that, um, that will allow us to vibe code on, on a weekly basis. 00:21:57.440 |
Whenever we don't have like a specific topic we want to, we want to talk about. 00:22:01.440 |
Um, so with that said, I, I, I just wanted to kind of show that off and, uh, maybe I'll pull up some of these other, um, uh, images that are coming into the chat. 00:22:22.440 |
Um, so Manuel says, see how coherent it is across styles. 00:22:30.440 |
Uh, I think this is like stretching probably the, the, the context window of, of the model. 00:22:36.440 |
But I would like to see, um, let's see already already in a completely different style with no other context. 00:22:47.440 |
I don't, I don't know what else you're looking for specifically, but, um, while that's being generated. 00:22:53.440 |
I was, uh, gonna ask for a character sheet for pixel graphics. 00:22:59.440 |
Cause you get like four different views or so. 00:23:18.440 |
This is what K-ball actually wants for work for his meetings. 00:23:23.440 |
It's like already to show up and just host, you know, do everything. 00:23:32.440 |
Yeah, Artie's closing the deals, Artie's taking on the, all the responsibility. 00:23:45.440 |
Creating an image of Artie in a meeting with businessmen. 00:23:48.440 |
Okay, let's go back and see if we, that's, I don't, this is not exactly a different style, but it's interesting. 00:23:55.440 |
I want to see, what were the keywords that you said again, Manuel, Artie? 00:24:14.440 |
And while I'm doing that, I should also figure out how to. 00:24:21.440 |
I like the Sith one, that the original that CJ put in there. 00:24:31.440 |
I'll pull them up as I'm trying to figure out how to. 00:24:36.440 |
Well, I mean, it says it's only like, was that roughly 10,000 tokens. 00:24:41.440 |
But I'd imagine like the coherence over the course of a, of a conversation is, is probably 00:25:01.440 |
So if I download it and then click this here, I should be able to show a full screen. 00:25:14.440 |
This is definitely, what are the, prompt golfing at its finest. 00:25:19.440 |
Cause you literally said four words and got it to change the entire style. 00:25:33.440 |
This is, I feel like, so, um, barbaric by typing in these. 00:25:40.440 |
I appreciate the hard work, but like, it's not really necessary. 00:25:47.440 |
That like, I don't know how it works internally. 00:25:50.440 |
But like when I did image model prompting, like two years ago, or even one year ago, you 00:25:55.440 |
had to really like work on your prompts and every word like counted. 00:26:00.440 |
And then open AI does all this like magic transformation. 00:26:06.440 |
Like if you add artist names or whatever, or like all kinds of weird stuff. 00:26:16.440 |
It's been, it's been getting a ton of hype over the last few weeks. 00:26:18.440 |
And I just, I was like, we've got to talk about this in the, uh, 00:26:23.440 |
I think in discord DMS, I should have enough text for it. 00:26:33.440 |
Cause I'm, I got a weekend hack project that I'm doing D and D shit on. 00:26:37.440 |
So I am, uh, we can have, um, uh, uh, uh, we can have, we can replace the person thing. 00:26:49.440 |
The care up character thing up there at the top with already or whatever, but, um, or you could just like dump this and do panel one and then do panel two, et cetera. 00:26:59.440 |
Um, but that this is, I asked Chad GBT to do some storyboarding, to test an image generation model. 00:27:10.440 |
And also I, there's something in the zoom chat too, that I dropped off earlier. 00:27:14.440 |
That's like, uh, let's, let's remix already into, uh, into, uh, that. 00:27:21.440 |
So, so, uh, from, for me, from my understanding, you're saying that this image that you sent originally, 00:27:27.440 |
that the, you are shined renegades, cyber slasher is like an example. 00:27:32.440 |
So that, that what I was going to say is let's feed that to the image and say, make Artie into this guy. 00:27:39.440 |
And then we can take that Artie and we can put him through those scenes. 00:27:52.440 |
I also, there was a thread that I was a part of. 00:27:56.440 |
Not too long ago in the main channel that showed a bunch of crazy examples of. 00:28:05.440 |
Oh, it's still active as like a couple of days ago. 00:28:10.440 |
And there's also another one, but I'm not worried about it. 00:28:15.440 |
High quality character generator puts hangover in the gear. 00:28:21.440 |
Here's a thread and I'll share the links to these threads. 00:28:28.440 |
So you guys can also check them out yourself. 00:28:44.440 |
And then there's another one that I was looking for. 00:29:08.440 |
But between both of these, you should have like more than enough inspiration that you can find here. 00:29:15.440 |
But both of those are threads in the latent space discord server. 00:29:31.440 |
And this can actually go right next to that tab. 00:29:59.440 |
It chose the images it chose for the for the playing cards. 00:30:04.440 |
It's at the cards are totally messed up, right? 00:30:33.440 |
I typed that prompt in and then I go pull up my downloads folder, which has the original already. 00:31:00.440 |
They're not playing cards of him, but he is playing cards. 00:31:29.440 |
It does seem like it left already the same, but then introduced like this other. 00:31:35.440 |
Um, but I'm, I'm curious what the prompt for that one was, Evan. 00:31:44.440 |
It says, um, CJ said to make Artie this character into, into a watercolor sorcerer. 00:31:57.440 |
Um, it rejected, uh, image request related violence, but it doesn't seem to reject referencing 00:32:14.440 |
Um, this is another one of the reasons why I wanted to, um, kind of like give, give, or, you 00:32:21.440 |
know, do some demos on this, uh, as quickly as possible because, uh, yeah, they, they always 00:32:30.440 |
So, um, apparently it has no barriers barrier on referencing the Simpsons. 00:32:37.440 |
I'm just not seeing that image punk rock concert flyer. 00:32:44.440 |
As a copy and pasted your prompt and let's see what happens. 00:32:52.440 |
Anyone tried prompting nano banana with json formatted prompt. 00:32:58.440 |
Oh, I like, so it, it, it gave you, so you gave it one prompt and it gave you multiple images. 00:33:25.440 |
I am noticing also the, the length of time to generate. 00:33:31.440 |
I'm just doing this all in one long thread and each time it takes longer. 00:33:35.440 |
So it's doing something with all the different, uh, images. 00:33:40.440 |
The content, the context as which is, yeah, that's, I have no clue how this works behind 00:33:58.440 |
Well, the multiple generations here, cause you have curry. 00:34:01.440 |
Wait, who is, who, who is this supposed to be? 00:34:12.440 |
So young pictures of him are probably like from a long time ago, but I don't know who that 00:34:31.440 |
What was the, the, the, the, the keywords you told me again, man? 00:34:36.440 |
I wonder if we can get like, get back into that line of like, it was pixel something. 00:34:45.440 |
It's just like, you know, being mindful of the context of like whatever you put in it before. 00:34:52.440 |
I think last time we tried this, we were like 15 images deep or generations deep. 00:34:57.440 |
Um, I want a different, so like, and it definitely takes into context, whatever you've been talking 00:35:02.440 |
about because it took the pixel graphics character sheet and now like put it into the form of this 00:35:07.440 |
punk rock, uh, concert flyer, which is, which is super interesting. 00:35:16.440 |
It doesn't actually understand character sheet the way I want that to understand it. 00:35:21.440 |
It does this like weird mix of RPG character sheet and like, so maybe try animation character sheet. 00:35:32.440 |
Uh, we, oh, I'll start from the top and instead of poker playing cards, I can do, um, animation 00:35:54.440 |
Um, there's a guy named, um, Jeremy Howard, who has talked about this, like editing the, the responses of the LLM to make it more, um, in line with like what you're, what you actually want. 00:36:07.440 |
Um, and I think this is like a case of that, uh, within AI studio and Google. 00:36:29.440 |
I'm trying to, if I could quickly find the other. 00:36:32.440 |
Um, that, that nano banana has a world knowledge. 00:36:36.440 |
So, uh, the prompt it gave, it was like, uh, you are a location based AR, AR experience generator. 00:36:42.440 |
Highlight the point of interest in this image and annotate relevant information about it. 00:36:48.440 |
Um, see if I can zoom in a bit, but it's just taking like real world images and highlighting them, which I thought was pretty cool. 00:37:12.440 |
Um, as far as like 3d goes, but you, it basically prompted nano banana to make an image daytime and isometric with the building only. 00:37:21.440 |
And as you can see, like this image is not exactly high quality. 00:37:24.440 |
It's probably taken on like a somewhat older phone. 00:37:30.440 |
Like the building in the background is not even really the focus. 00:37:34.440 |
Um, but just based on this image, it was able to generate an isometric image of the building, which I think is insane. 00:37:45.440 |
Um, so yeah, there's just like we're using it for fun for the most part in this, uh, you know, like the demo that we've had so far in the first 40 minutes or so. 00:37:55.440 |
I think it can do some really serious stuff, uh, based on what I saw on those, those, um, those threads. 00:38:01.440 |
So I thought this was really cool too, but I'll, I'll skip that one. 00:38:06.440 |
I think, um, speaking of coherence, I don't know if you guys seen this one, but it's a guy talking originally. 00:38:25.440 |
I thought it was, oh, he, he goes into details here. 00:38:28.440 |
He dropped it into nano banana, um, swapped clothes, switch styles, even turned myself into an animated character. 00:38:39.440 |
So free pick is one of the platforms that Google at, we talked about this earlier. 00:38:42.440 |
Google launched this, um, this, this model in stealth on these other platforms. 00:38:48.440 |
And from what I understand, free pick was one of them. 00:38:52.440 |
There's another like AI based platform that I can't think of to save my life. 00:38:55.440 |
But, uh, it, it launched on that platform as well before, before they announced nano banana and said, Hey, come use it in Google AI studio. 00:39:08.440 |
So do you think, did he decompose that into a bunch of frames or did they just, they do that for him under the covers? 00:39:15.440 |
Um, so I, I, I think he used runway cause runway does video. 00:39:19.440 |
So I believe he used runway in this instance. 00:39:21.440 |
Um, but I'll, I'll actually share this link directly in the chat. 00:39:27.440 |
I think here in the, in the rest of the thread, as he's like answering questions and stuff. 00:39:31.440 |
But, um, if I'm not mistaken, he used together as one of the platforms. 00:39:36.440 |
That's not the platform I'm thinking of, but together is another like image generation platform. 00:39:39.440 |
If I remember correctly, where they, they do like hosted models. 00:39:42.440 |
But, uh, I think this is a combination of nano banana and runway for, for this particular video. 00:39:47.440 |
Um, if I remember correctly, I thought this was fire. 00:39:51.440 |
I do a bunch of e-commerce stuff with my, my younger brother. 00:39:59.440 |
So, you know, took a selfie of this guy that I think is a stock photo from what I can tell. 00:40:03.440 |
If not, it's maybe like a high, high quality image. 00:40:06.440 |
And, uh, took, uh, a product image as you can see here. 00:40:10.440 |
That's obviously done in a studio or something like that. 00:40:13.440 |
And he was able to combine both of them into this product placement image, just using a nano banana. 00:40:26.440 |
If you guys, anything else you guys want to cover here in the last 20 minutes, I'm going to kind of open the floor. 00:40:32.440 |
Um, I've yet to try this one, but I really, I saw this one and started freaking out. 00:40:41.440 |
Um, but it basically turned any image into, uh, uh, an AR HUD, uh, because of the world knowledge. 00:40:54.440 |
Apparently this technology has been around since 2017. 00:41:00.440 |
Um, yeah, just a ton of examples, a ton of examples. 00:41:08.440 |
Um, Evan, the, the, the characters that it created is insane. 00:41:21.440 |
David Gutman said, I've been trying to, oh, David, that's what I was trying to think of earlier. 00:41:26.440 |
Um, trying to get nano banana to do something like this. 00:41:30.440 |
This is like the, the sprite thing that you were talking about earlier, right? 00:41:33.440 |
Um, Slono, I mean, well, it's a character sprite sheet. 00:41:45.440 |
So, so I, I don't think this is actually a, a nano banana generation. 00:41:49.440 |
I think David was saying that he's been trying to get it to do something like this. 00:41:58.440 |
This seems, cause I was like, this seems very unrelated. 00:42:00.440 |
Um, Ooh, and CJ got it to do this, which is interesting. 00:42:09.440 |
I like the color palette in the top right hand corner. 00:42:19.440 |
So CJ said that the prompt for this was generate a pixel animation character of this. 00:42:24.440 |
No, a general, a pixel animation character sheet of this character already. 00:42:28.440 |
And this is from, Ooh, pull this character out and have him jumping out of a van and checking 00:42:42.440 |
If you guys have anything else that you want to cover, we can, oh my, no way. 00:42:55.440 |
Um, it's like five minutes later and this thing is, uh, still generating the animation 00:43:03.440 |
So, um, oh my God, that's a great point cable. 00:43:07.440 |
So cable says, uh, they're going to have to neuter it with regard to celebrate, uh, celebs, 00:43:33.440 |
And Keanu Reeves is like, I don't have anything. 00:43:49.440 |
The rock might be realistic, but like, and I don't even know who I'm bad with celebs. 00:43:55.440 |
I just told it generate, you know, five celebrities holding liquid death. 00:43:59.440 |
I can't think of product placement, liquid death, five celebrities. 00:44:05.440 |
Um, actress, if I remember correctly, I just can't remember a name off the top of my head. 00:44:20.440 |
Let me download that specifically into my downloads folder. 00:44:31.440 |
I was able to ask it just like, who are the celebrities? 00:44:48.440 |
Can, can you, can you also like reprompt me on what we were doing with this? 00:44:54.440 |
Cause I, I'm, I'm, I'm in my DM now and I see the prompt and we've got like 15 minutes left. 00:45:02.440 |
Uh, uh, give it, why don't you get, uh, give it a picture of Artie and then give it that first thing in the care, the character block thing at the very top. 00:45:14.440 |
And then, and then prompt it through those scenes, like one through 12 or whatever. 00:45:19.440 |
So the first one will, the first prompt to look something like this. 00:45:26.440 |
And so, but attach Artie to that and be like, make Artie this. 00:45:44.440 |
And then you can, after that, we can test the consistency with, see if, see how it doesn't seem. 00:45:54.440 |
Well, you did use the word stabby, so it might not work. 00:46:10.440 |
See, then try and throw it in scene one and see what happens. 00:46:31.440 |
Let me, let's take a, I think, let me see what that prompt looked like in a sec. 00:46:44.440 |
So we, yeah, we can do like maybe the first two panels and see, see what, you can do one 00:46:53.440 |
You could add, you could try referring to, you could try using that care character block 00:46:59.440 |
thing and see if it will take that as like a variable. 00:47:02.440 |
It could be interesting to see what it remembers. 00:47:06.440 |
And what, so when you say using it as a variable, are you saying to like, cause I, I just removed 00:47:13.440 |
the, like just copy, like copy paste completely from panel one down and don't replace that 00:47:24.440 |
So do I, do I leave the, where it says like little, quite literally prompt, like P R O M P 00:47:45.440 |
And, and I'm curious about what kind of, okay. 00:48:11.440 |
It's a little like stale for me, but the, but I was trying to get it to do stylized earlier 00:48:16.440 |
and it doesn't really, it really like gloms onto the stuff it's already seen. 00:48:33.440 |
Instead of having one little companion on the shoulder, it has two of them. 00:48:45.440 |
And we're changing aspect ratio here too, which is kind of nice. 00:49:00.440 |
I just realized you, you did the aspect ratio thing. 00:49:07.440 |
So it's, it didn't, it looks like it didn't pick up the aspect ratio on this one. 00:49:17.440 |
That's a terrifying grin, but I have to refer back to the chat and see if anyone else has 00:49:30.440 |
any questions or, or has seen some cool examples of, of, so I'm on panel four number, panel number 00:49:36.440 |
four now, but just seeing if anyone has characters, I mean, questions about nano banana or I've seen 00:49:48.440 |
So are we, we, we didn't get a lot of variation on the camera angle here. 00:49:53.440 |
I've noticed that, that it does seem to struggle with camera angles. 00:50:02.440 |
I'm wondering, could it, could you like, I like how, if you look closely at the details right here, the gun fire, like goes over into the next panel, like the bottom panel. 00:50:15.440 |
I wonder if you could tell it, can you ask it, ask it for a wire frame of those three shots. 00:50:28.440 |
No, just like, give me a wire frame of that so I can animate it. 00:50:35.440 |
So far, I really like this character development though. 00:50:53.440 |
I told it to go research the setting and everything. 00:50:59.440 |
The, the one thing I noticed that's missing here in AI studio is like with the, you know, text language models, the text generation models, you can actually, you know, put Google grounding. 00:51:09.440 |
And so it can be like use Google to go search things and add that to the conversation. 00:51:12.440 |
Whereas here, like with the image generation. 00:51:18.440 |
What I was looking for was like a, a little like a skeleton. 00:51:23.440 |
Like you would drag it around in unity or something. 00:51:36.440 |
Like if we fed it something like this, could it create a wire frame from the original image? 00:51:41.440 |
Like if, if it doesn't, cause I, I do feel like it loses. 00:51:44.440 |
Um, like almost intelligence over time because it's trying to keep into context, the entire conversation. 00:51:50.440 |
And that's one, you know, prompt by prompt by prompt. 00:51:57.440 |
The, uh, show the woman wearing the outfit and you just send it a picture of the outfit. 00:52:16.440 |
Cause I feel like every clothing company is now going to like hire a model once and then 00:52:24.440 |
I mean, well, if I was, uh, yeah, if I was doing, if I was in that business, it would be 00:52:38.440 |
Like, I think Google opening it up to everybody for free makes it like the, the barrier super 00:52:43.440 |
And yeah, I do think there, there will be a lot of that stuff. 00:52:52.440 |
So, uh, cable, when you, when you were getting denials on your requests on your, on your prompts, 00:52:56.440 |
do you have the, uh, safety settings turned all the way down? 00:53:03.440 |
The only prompts that the only prompt that it denied was when I asked it to show already 00:53:16.440 |
One of which was, I told the, uh, that original van prompt was like kicking down the door to 00:53:23.440 |
And it didn't, uh, it did not, uh, go for that. 00:53:35.440 |
So, so David, what are your thoughts on this? 00:53:38.440 |
Like this generation when it like did half the job or like didn't do the whole. 00:53:45.440 |
I've been, I mean, I've been trying to get into various, uh, models to, to do the animations, 00:53:52.440 |
I, I don't know if this one is any, any closer, uh, but it's totally possible that with the right 00:53:59.440 |
prompt it, it could be, or doing it frame by frame, uh, would be able to do it, uh, instead 00:54:06.440 |
of trying to get it to do a whole spreadsheet at once. 00:54:12.440 |
Uh, bite probe, um, posted a link to levels.io, uh, tweet and said, uh, prompt a show a women, 00:54:22.440 |
show woman wearing the outfit and, and a picture of this outfit. 00:54:25.440 |
And, uh, we're coming up on the last couple of minutes here, but it didn't output a woman 00:54:30.440 |
wearing like the outfit, uh, that was shown before. 00:54:37.440 |
And then we had a response from CJ saying that this would be great for my e-commerce site. 00:54:41.440 |
I think that it's super powerful for the, for the e-commerce stuff. 00:54:44.440 |
If we're talking like real life use cases that could help make money, you know what I mean? 00:54:49.440 |
So I do want to actually, since we've got only five minutes left, you know, as awesome 00:54:54.440 |
as it is to get flow up here, like jamming on the news of the week, which is great. 00:55:00.440 |
Uh, it would be cool if we got back to some sort of prepared stuff a little bit. 00:55:05.440 |
So does anybody, maybe we can spend a couple of minutes here, like jamming on possible topics. 00:55:10.440 |
Um, we, I don't know, maybe in the AI world, we can count on there being enough news of the 00:55:16.440 |
week that we could just jam on news of the week each week. 00:55:18.440 |
But I, I think maybe we should do a little bit of a call for topics who has something they 00:55:24.440 |
think would be fun for us to do a session on, um, or a session they'd like to see that maybe 00:55:34.440 |
I'm curious, uh, kind of if, uh, if anybody's like taken any episodes of AI in action, or I 00:55:41.440 |
guess you could, you could generalize it more, but I'm curious, like of the stuff that we've 00:55:45.440 |
kind of gone over so far, has anybody like updated their workflow? 00:55:51.440 |
like pretty significantly and like something like significantly impacted them. 00:55:56.440 |
And this is kind of what they're doing with it. 00:55:57.440 |
Kind of thing is, uh, something I'm wondering about, I guess. 00:56:00.440 |
I, I switched to only using GPT-5, so I'm, I'm happy to talk about that. 00:56:21.440 |
It seems like a lot of people have become like Codex CLI fanboys overnight or using Codex and 00:56:29.440 |
Um, I, I tried to prep like for this call so we could like actually vibe code the thing, 00:56:33.440 |
but, but yeah, without having, um, the server or whatever set up. 00:56:37.440 |
It's like, wasn't, wasn't, uh, able to really pull it off. 00:56:40.440 |
But I, I do want to hear more about, I think the most interesting AI in actions for me is 00:56:45.440 |
Um, and I think there've been a couple of cases in the, in the channel of like people 00:56:49.440 |
So I would, any, any, any workflow that could be shared. 00:56:51.440 |
Like Scott gave a good one a couple of weeks ago. 00:56:53.440 |
We have one from Oliver a few weeks ago as well. 00:57:05.440 |
What are you thinking in terms of like what you would want to see is in terms of a prepared. 00:57:21.440 |
So what, what I would be interested in if we could do it. 00:57:26.440 |
So I'm trying to help my wife with some agentic content generation for something. 00:57:31.440 |
And so I actually referenced like Olivia's talk cause she has a very structured thing. 00:57:35.440 |
And I was like, Hey, you should like check this out, build out your sub agent, do whatever. 00:57:38.440 |
She, I got her set up today with Claude code saying where we'll see where that goes. 00:57:42.440 |
Um, but I'd actually be interested if anyone else is doing agentic content. 00:57:49.440 |
Um, creation, you know, something with that, like if somebody would wanted to share their workflow, how they're approaching things, how they've got stuff structured and that sort of thing. 00:57:58.440 |
Right? Like, cause I live in the coding world mostly, but I think, um, you know, content is another very interesting place. 00:58:02.440 |
And the tools that I've seen out there, like she had been trying to just use the Claude web UI and it works for relatively simplistic things. 00:58:11.440 |
But you can, if you have like a very large amount of stuff that where you want to go structured, like it doesn't give you any exposure to sub agents or swatch swapping. 00:58:21.440 |
And so like, we're trying to use coding tools to see if, cause they're based. 00:58:24.440 |
I mean, these coding tools are all trying to solve. 00:58:31.440 |
Well, and that's kind of my project for this weekend is, uh, I want to, uh, uh, with, uh, cause there's the L the LTX video. 00:58:40.440 |
I don't know if there's going to be a video model. 00:58:42.440 |
It's quite fast enough, but basically I want to like do a thing where I can more or less kind of run like a D and D ish game via like a Twitter space and then just get like the transcript and then extract prompts out of that. 00:58:58.440 |
And just start dumping that into a video generation model and see if I can stitch something together where people can like watch their D and D people do shit. 00:59:06.440 |
Um, but I don't know if I'm gonna be able to get that done over the weekend. 00:59:10.440 |
So, um, and I also, and that'll be like a fairly code heavy, but content oriented workflow. 00:59:17.440 |
So, I mean, if we don't, and if we don't end up with anything else, I could just like, you know, the hack on that and then see if I get anything worth showing to anybody. 00:59:28.440 |
Um, as a, as a last resort, I suppose, but it seems like people are still pretty interested in, uh, Codex and CC. 00:59:35.440 |
I know we've done, we've done like a good bit of that. 00:59:37.440 |
It's always good to get more, but it's like, um, I think the content angle is an interesting, uh, uh, 00:59:45.440 |
So my, my CTO does this, I can try to see if he was willing to present next week. 00:59:53.440 |
I'll, I'll, I'll, I can ask him like now so we can have an answer by Monday if people want. 01:00:01.440 |
When you say this, do you mean like those content? 01:00:03.440 |
He generates content algorithmically using AI somehow. 01:00:07.440 |
Uh, I think he has like some level of workflow that maybe is like probably interesting to look at too. 01:00:13.440 |
So I can see if he was willing to present or if he's willing to make me present on it. 01:00:27.440 |
We're gonna, we're gonna sign you up via our discord bot, which now lets me sign someone else up. 01:00:57.440 |
We'll, we'll go with Baruch, uh, Ben or Benjamin, um, for next week. 01:01:05.440 |
You can use Artie, who is still called AI in action bot, but we should change that name to Artie at some point.