WEBVTT 00:00:00.000 --> 00:00:03.060 So there's this new AI image editing tool, right? 00:00:03.259 --> 00:00:05.719 And it's just, it's blowing everything else away 00:00:05.719 --> 00:00:07.719 in blind tests. Yeah, it's really something. 00:00:08.199 --> 00:00:11.400 We're talking over 200 ELO points ahead on Elma 00:00:11.400 --> 00:00:14.050 Arena. That's a massive gap. I mean, that really 00:00:14.050 --> 00:00:15.810 grabs your attention, doesn't it? It changes 00:00:15.810 --> 00:00:18.250 the whole game. It absolutely does. This thing, 00:00:18.250 --> 00:00:20.530 we've been calling it Nano Banana, just kind 00:00:20.530 --> 00:00:23.429 of internally. But the official name is Gemini 00:00:23.429 --> 00:00:26.989 2 .5 Flash Image. It's really shifting how we 00:00:26.989 --> 00:00:28.929 think about, you know, accurate photo editing. 00:00:29.070 --> 00:00:31.910 Whole new benchmark. Exactly. Welcome to the 00:00:31.910 --> 00:00:33.590 deep dive. Today, we're going to cut through 00:00:33.590 --> 00:00:36.909 all the hype around this new Google tool. Our 00:00:36.909 --> 00:00:40.829 mission really is to figure out why. Why is it 00:00:40.829 --> 00:00:43.070 number one? Where's that power coming from? And, 00:00:43.090 --> 00:00:46.570 you know, crucially, how can you use it for fixing 00:00:46.570 --> 00:00:48.929 old photos, maybe, or even professional character 00:00:48.929 --> 00:00:51.530 design? Yeah, we'll dig into the tech behind 00:00:51.530 --> 00:00:53.109 it first, then walk through the basics, some 00:00:53.109 --> 00:00:54.890 of the editing features. We need to talk about 00:00:54.890 --> 00:00:56.850 the consistency stuff. That's huge. But also 00:00:56.850 --> 00:00:59.450 look at where it kind of hits its limits creatively. 00:00:59.810 --> 00:01:02.549 OK, good. And then we'll wrap up with some solid 00:01:02.549 --> 00:01:05.870 tips, right? How to actually write prompts that 00:01:05.870 --> 00:01:07.730 work consistently. So let's start with that score 00:01:07.730 --> 00:01:10.370 gap. Okay, the numbers. They're pretty wild. 00:01:11.030 --> 00:01:14.069 If you look at Ella Marina, most of the top editors, 00:01:14.290 --> 00:01:18.290 they hover around 1 ,100 ELO. This tool, Mano 00:01:18.290 --> 00:01:21.069 Banana, it's pushing past 1 ,300. Okay, wait. 00:01:21.069 --> 00:01:23.250 For someone listening who doesn't follow these 00:01:23.250 --> 00:01:25.370 leaderboards religiously, what does a 200 ELO 00:01:25.370 --> 00:01:27.790 gap actually mean? It's a really good question. 00:01:28.090 --> 00:01:31.409 Think about chess. If a player has a 200 ELO 00:01:31.409 --> 00:01:34.430 advantage, they're expected to win about, what, 00:01:34.569 --> 00:01:36.870 76 % of the time against the lower rated player? 00:01:37.129 --> 00:01:39.969 It's a significant edge. So in these blind tests 00:01:39.969 --> 00:01:41.909 where people are just picking which image looks 00:01:41.909 --> 00:01:45.290 better or more accurate, this new model is just 00:01:45.290 --> 00:01:47.469 winning, like three out of four times against 00:01:47.469 --> 00:01:49.469 the best competitors out there. Pretty much, 00:01:49.609 --> 00:01:52.390 yeah. Consistently, overwhelmingly chosen. That's 00:01:52.390 --> 00:01:54.689 the advantage we need to unpack. OK, so if it's 00:01:54.689 --> 00:01:56.430 that much better, the sources must be saying 00:01:56.430 --> 00:01:59.170 it's not just slightly more data or better code. 00:01:59.250 --> 00:02:00.950 It has to be something different at the core. 00:02:01.170 --> 00:02:04.390 It is. The secret sauce is the foundation. Gemini 00:02:04.390 --> 00:02:08.009 2 .5 Flash. That's Google's special model. And 00:02:08.009 --> 00:02:11.750 the key thing is, it's multimodal. Multimodal. 00:02:11.909 --> 00:02:14.009 Meaning it gets different kinds of info at once. 00:02:14.569 --> 00:02:16.990 Words, pictures. Exactly. Words, pictures, sounds, 00:02:17.210 --> 00:02:19.530 even video concepts. It understands them all 00:02:19.530 --> 00:02:22.530 together. Ah, okay. Like stacking different kinds 00:02:22.530 --> 00:02:25.189 of Lego blocks of data, maybe? That's a great 00:02:25.189 --> 00:02:28.169 way to put it. Imagine editing a Lego build. 00:02:28.550 --> 00:02:30.930 A normal tool might just see, I don't know, the 00:02:30.930 --> 00:02:33.889 color of a brick. A multimodal one sees the color, 00:02:33.969 --> 00:02:35.689 the shape, where it sits next to other bricks, 00:02:36.169 --> 00:02:37.949 and it understands the written instructions to 00:02:37.949 --> 00:02:40.409 the final thing. Oh, okay. So it has a deeper 00:02:40.409 --> 00:02:43.370 context. Right. So because of that, NanoBanana 00:02:43.370 --> 00:02:46.430 isn't just making images, it's editing them super 00:02:46.430 --> 00:02:48.509 accurately, especially keeping all those little 00:02:48.509 --> 00:02:51.289 details right. That's stability. That makes sense. 00:02:52.039 --> 00:02:54.300 If the power comes from being multimodal, how 00:02:54.300 --> 00:02:57.020 does that, like, technical thing translate into 00:02:57.020 --> 00:02:58.939 a real benefit for someone just trying to fix 00:02:58.939 --> 00:03:02.099 an old photo? Well, that reliability, you know? 00:03:02.219 --> 00:03:05.280 Keeping tiny details consistent, like a specific 00:03:05.280 --> 00:03:07.919 glint in someone's eye or the weave of fabric. 00:03:08.199 --> 00:03:10.400 It just makes the tool feel much more trustworthy. 00:03:10.659 --> 00:03:13.240 Right. It stops those weird, jarring changes 00:03:13.240 --> 00:03:15.439 you sometimes see where the AI gets confused. 00:03:15.699 --> 00:03:18.139 Exactly. It avoids that stuff. OK, so we get 00:03:18.139 --> 00:03:21.189 the why. It's that multimodal foundation. Now, 00:03:21.189 --> 00:03:23.229 for someone listening who's thinking, OK, I want 00:03:23.229 --> 00:03:26.610 to try this, where do they actually go? Two main 00:03:26.610 --> 00:03:30.009 places. For beginners, definitely Google AI Studio. 00:03:30.090 --> 00:03:32.750 It's free. The interface is super simple, really 00:03:32.750 --> 00:03:35.550 straightforward. You just upload your picture, 00:03:35.969 --> 00:03:37.969 type what you want changed in the text box, be 00:03:37.969 --> 00:03:40.789 specific, click Run, and then you can download 00:03:40.789 --> 00:03:43.789 it. And Ella Marina is still there for the comparisons 00:03:43.789 --> 00:03:45.849 if you want to really nerd out on the scores. 00:03:46.009 --> 00:03:48.629 Right, for testing. But for actually making stuff, 00:03:48.849 --> 00:03:51.560 AI Studio is the way to go. OK. But there's a 00:03:51.560 --> 00:03:53.379 catch, right? The free use. There is, yeah. It's 00:03:53.379 --> 00:03:55.520 limited right now. Usually like three to five 00:03:55.520 --> 00:03:57.680 free edits before you hit a wall. So planning 00:03:57.680 --> 00:04:00.000 your edits is really important. If you have a 00:04:00.000 --> 00:04:02.780 ton of photos, the sources suggest maybe using 00:04:02.780 --> 00:04:06.060 different Google accounts. But yeah, treat those 00:04:06.060 --> 00:04:08.719 first few tries carefully. Like gold. Got it. 00:04:09.300 --> 00:04:12.009 Okay, speaking of valuable stuff. Photo restoration. 00:04:12.550 --> 00:04:15.270 The source has really highlighted this. Fixing 00:04:15.270 --> 00:04:17.389 old family photos. That's a massive strength, 00:04:17.430 --> 00:04:19.329 again, because of that detail retention. You 00:04:19.329 --> 00:04:22.129 can give it an old, faded, scratched up black 00:04:22.129 --> 00:04:25.730 and white photo and just ask simply, like, color 00:04:25.730 --> 00:04:28.550 this photo naturally and fix the torn parts. 00:04:29.129 --> 00:04:32.050 Doesn't just slap color on. It actually rebuilds 00:04:32.050 --> 00:04:34.129 the missing bits, smooths out the damage, gets 00:04:34.129 --> 00:04:35.930 the color balance right. And crucially, like 00:04:35.930 --> 00:04:37.790 we said, it doesn't mess up the faces, does it? 00:04:38.110 --> 00:04:40.529 Older tools sometimes tweak the expression or 00:04:40.529 --> 00:04:43.089 the features when fixing damage. Right. This 00:04:43.089 --> 00:04:45.589 one keeps those original features almost perfectly 00:04:45.589 --> 00:04:48.550 preserved. That's huge for anyone trying to digitize 00:04:48.550 --> 00:04:50.790 family albums or for historical stuff. Keeps 00:04:50.790 --> 00:04:53.930 the soul of the picture. Exactly. So beyond fixing 00:04:53.930 --> 00:04:56.279 old damage, how about just basic cleanup? Getting 00:04:56.279 --> 00:04:58.160 rid of photobombers, that kind of thing. Yeah, 00:04:58.220 --> 00:05:00.660 it's great at that, too, because it's precise. 00:05:01.060 --> 00:05:03.379 You can tell it very specifically. Remove the 00:05:03.379 --> 00:05:05.819 person in the background on the left, and it 00:05:05.819 --> 00:05:08.000 intelligently takes them out without messing 00:05:08.000 --> 00:05:10.759 up the rest. OK, now this is where, for me, it 00:05:10.759 --> 00:05:13.100 gets really exciting, especially for creative 00:05:13.100 --> 00:05:15.639 work. Advanced consistency, that's always been 00:05:15.639 --> 00:05:17.980 the hard part for AI image tools. Oh, absolutely. 00:05:18.459 --> 00:05:21.779 This is where that 200 ELO lead really feels 00:05:21.779 --> 00:05:24.720 earned. It's ability to keep a person looking 00:05:24.720 --> 00:05:28.579 like the same person across really drastic changes. 00:05:29.079 --> 00:05:31.019 That's the magic. So you can give it one picture. 00:05:31.220 --> 00:05:33.939 Yep, one reference photo. And then ask for huge 00:05:33.939 --> 00:05:35.860 changes. Like put them in a totally different 00:05:35.860 --> 00:05:38.180 outfit, a beige wool jacket, maybe then a long 00:05:38.180 --> 00:05:40.420 green dress. Change the whole background, move 00:05:40.420 --> 00:05:42.980 them from inside an office to like an autumn 00:05:42.980 --> 00:05:45.839 garden. And the face. It just stays the same. 00:05:46.110 --> 00:05:49.009 perfectly consistent through all that. Undeniably 00:05:49.009 --> 00:05:52.129 consistent. It's uncanny. Whoa. OK, that is the 00:05:52.129 --> 00:05:54.050 moment of wonder. Imagine how much time that 00:05:54.050 --> 00:05:56.569 saves for, say, game designers making character 00:05:56.569 --> 00:06:00.189 sheets. Exactly. Or concept artists. You need 00:06:00.189 --> 00:06:03.050 consistent looks for characters, whether it's 00:06:03.050 --> 00:06:06.069 anime or detailed robots. Just use text prompts 00:06:06.069 --> 00:06:08.670 for the changes and the core character doesn't 00:06:08.670 --> 00:06:12.009 drift. It's amazing. It shrinks that whole tedious 00:06:12.009 --> 00:06:14.310 process down, right? Ensuring lighting is the 00:06:14.310 --> 00:06:16.930 same, features are the same. from potentially 00:06:16.930 --> 00:06:19.850 hours to just seconds of prompting. The scalability 00:06:19.850 --> 00:06:22.310 is, yeah, it's pretty mind -blowing. And this 00:06:22.310 --> 00:06:24.709 consistency, does it carry over into things like 00:06:24.709 --> 00:06:27.009 virtual try -on? The sources mentioned that was 00:06:27.009 --> 00:06:29.769 a big plus. Totally. That's another area where 00:06:29.769 --> 00:06:31.529 it really pulls ahead of other apps right now. 00:06:31.529 --> 00:06:33.629 OK. And we're not just talking simple stuff like 00:06:33.629 --> 00:06:36.170 putting one shirt on one person. It handles more 00:06:36.170 --> 00:06:38.970 complex swaps, too. Yeah. Especially in group 00:06:38.970 --> 00:06:41.389 photos. Give me an example of a complex swap. 00:06:41.430 --> 00:06:43.769 What does that look like? OK. Picture a photo 00:06:43.769 --> 00:06:45.779 with, say, three or four people. You can tell 00:06:45.779 --> 00:06:48.939 the AI. Swap the entire outfit of the man in 00:06:48.939 --> 00:06:50.779 the blue shirt with the outfit of the woman in 00:06:50.779 --> 00:06:53.240 the red shirt. But keep everyone else exactly 00:06:53.240 --> 00:06:55.660 the same. And it does it. It transfers the clothes, 00:06:55.759 --> 00:06:57.759 the style, the fit, the texture between those 00:06:57.759 --> 00:07:00.360 two people, keeps their faces right, and doesn't 00:07:00.360 --> 00:07:02.079 make a mess of the background. That's impressive. 00:07:02.920 --> 00:07:05.480 OK, so if it's that good with characters and 00:07:05.480 --> 00:07:07.740 details, where do things start to break down? 00:07:08.100 --> 00:07:09.980 When are you pushing it too far? What's the first 00:07:09.980 --> 00:07:13.100 sign? Usually style changes. That's where the 00:07:13.100 --> 00:07:16.759 cracks appear first. Especially complex artistic 00:07:16.759 --> 00:07:19.540 requests, things far away from just realism or 00:07:19.540 --> 00:07:22.139 simple photo effects. Right. Even with that amazing 00:07:22.139 --> 00:07:24.959 consistency, the sources mentioned this art gap, 00:07:25.500 --> 00:07:27.720 a limit we need to be aware of. Correct. It's 00:07:27.720 --> 00:07:30.379 great at realistic style changes. Making a photo 00:07:30.379 --> 00:07:33.100 look like it was shot on old film, easy. Turning 00:07:33.100 --> 00:07:35.610 it into a realistic -looking oil painting. does 00:07:35.610 --> 00:07:37.930 that well. Adjusting mood and lighting too. Yeah, 00:07:38.170 --> 00:07:40.670 excellent that. Warm sunset light, gloomy rainy 00:07:40.670 --> 00:07:43.050 day effects, no problem. But the really artsy 00:07:43.050 --> 00:07:45.370 stuff is harder. That's where it struggles. If 00:07:45.370 --> 00:07:48.089 you say make this real photo look like true cell 00:07:48.089 --> 00:07:51.290 shaded anime, the results. They often look more 00:07:51.290 --> 00:07:53.670 like a detailed pencil drawing, maybe with anime 00:07:53.670 --> 00:07:56.870 colors, but not the actual style. And mixing 00:07:56.870 --> 00:07:59.170 styles, that's a big failure point currently. 00:07:59.850 --> 00:08:03.269 Like our Lego analogy, asking for left half Lego 00:08:03.269 --> 00:08:06.569 style, right half Pixar style, is usually too 00:08:06.569 --> 00:08:08.769 much. Too confusing for it. Yeah, it's like trying 00:08:08.769 --> 00:08:11.910 to combine incompatible instructions. Too many 00:08:11.910 --> 00:08:14.709 style words at once. Watercolor, vintage, pop 00:08:14.709 --> 00:08:18.769 art just leads to messy, weird results. And complex 00:08:18.769 --> 00:08:21.389 scene changes too, like changing multiple objects 00:08:21.389 --> 00:08:23.290 in a big way. Right, the source has had that 00:08:23.290 --> 00:08:25.819 example. turn all the girls into cats and boys 00:08:25.819 --> 00:08:28.920 into dogs in a group photo. That's just asking 00:08:28.920 --> 00:08:31.279 way too much right now. Too many individual changes 00:08:31.279 --> 00:08:33.919 at once. Yeah. I still wrestle with prompt drift 00:08:33.919 --> 00:08:35.779 myself sometimes. You try to combine too many 00:08:35.779 --> 00:08:38.019 things. Oh, yeah. I tried to get it to put a 00:08:38.019 --> 00:08:40.399 tiny Viking helmet on my dog and give him a glowing 00:08:40.399 --> 00:08:42.919 laser sword in his mouth in the same edit. Huh. 00:08:43.100 --> 00:08:45.820 How did that turn out? It was just... A mess. 00:08:46.240 --> 00:08:48.919 A melted, weird, inconsistent blob. It's easy 00:08:48.919 --> 00:08:50.860 to overload it. That's actually really helpful 00:08:50.860 --> 00:08:53.580 to hear because it shows it's not just user error 00:08:53.580 --> 00:08:56.639 sometimes. The tool has limits. Definitely. So 00:08:56.639 --> 00:08:58.960 if we know it struggles with those really complex 00:08:58.960 --> 00:09:01.480 multi -part edits, what's the best strategy? 00:09:01.620 --> 00:09:03.440 How do you get the result you want? You got to 00:09:03.440 --> 00:09:06.259 break it down. Always. One simple change at a 00:09:06.259 --> 00:09:09.500 time. Edit the main object first, then change 00:09:09.500 --> 00:09:11.820 the lighting, then maybe add a simple style. 00:09:12.000 --> 00:09:14.440 Atomic steps. OK, one step at a time. Makes sense. 00:09:14.740 --> 00:09:17.799 So going back from the limits to the strengths 00:09:17.799 --> 00:09:22.179 micro editing, tiny, precise changes. Yes. This 00:09:22.179 --> 00:09:24.440 is where that Gemini flash precision really comes 00:09:24.440 --> 00:09:26.379 through. Things like just changing eye color 00:09:26.379 --> 00:09:29.240 or adding just a little smile, stuff that other 00:09:29.240 --> 00:09:31.879 tools might use as an excuse to redo the whole 00:09:31.879 --> 00:09:35.179 face. Exactly. Nano Banana handles that delicately. 00:09:35.460 --> 00:09:37.519 And the consistency holds up so well, you can 00:09:37.519 --> 00:09:40.139 actually generate like a grid of nine different 00:09:40.139 --> 00:09:43.240 facial expressions. Happy, sad, angry, surprised 00:09:43.240 --> 00:09:45.960 from one starting photo. Wow. And the person 00:09:45.960 --> 00:09:47.919 still looks like the same person underneath with 00:09:47.919 --> 00:09:49.940 the same features, same lighting. That's super 00:09:49.940 --> 00:09:52.100 valuable for professional uses. How about text 00:09:52.100 --> 00:09:54.240 and images? That's usually a nightmare for AI. 00:09:54.759 --> 00:09:58.259 It's surprisingly decent here, though language 00:09:58.259 --> 00:10:00.919 translation can be hit or miss. But simple replacement 00:10:00.919 --> 00:10:03.419 works pretty well. changing a sign from welcome 00:10:03.419 --> 00:10:06.259 home to happy birthday. It can often do that 00:10:06.259 --> 00:10:08.639 while keeping the original font style in perspective. 00:10:09.320 --> 00:10:11.700 It seems to treat the text like an object that 00:10:11.700 --> 00:10:13.899 needs to preserve or change carefully. OK, so 00:10:13.899 --> 00:10:15.860 to really leverage all this power, you need good 00:10:15.860 --> 00:10:18.580 prompts. The sources laid out a structure, a 00:10:18.580 --> 00:10:20.639 way to get better results, especially with those 00:10:20.639 --> 00:10:23.059 limited free uses. Yeah, a simple three -part 00:10:23.059 --> 00:10:27.080 structure helps you think clearly. First, the 00:10:27.080 --> 00:10:29.500 action. What are you changing? Change the car 00:10:29.500 --> 00:10:33.980 color to red. OK. Second, the constraint. What 00:10:33.980 --> 00:10:35.980 needs to stay the same? Keep the background and 00:10:35.980 --> 00:10:38.259 the driver exactly the same. Constraint. Got 00:10:38.259 --> 00:10:41.320 it. And third, the style. What's the desired 00:10:41.320 --> 00:10:43.919 quality or feel? Make it look like a professional 00:10:43.919 --> 00:10:47.320 car ad. Action, constraint, style. That structure 00:10:47.320 --> 00:10:49.860 helps avoid those accidental changes. The drift. 00:10:50.159 --> 00:10:52.559 Exactly. It forces clarity. And since those free 00:10:52.559 --> 00:10:56.080 turns are precious, the advice is plan your edits. 00:10:56.350 --> 00:10:58.809 Use clear, simple language, and definitely save 00:10:58.809 --> 00:11:00.769 prompts that work well so you can reuse them. 00:11:01.190 --> 00:11:03.649 Smart. And always double check the output. Absolutely. 00:11:03.990 --> 00:11:07.490 Final QC is crucial. Check the face consistency 00:11:07.490 --> 00:11:09.789 again. Did it actually make all the changes you 00:11:09.789 --> 00:11:12.049 asked for? Sometimes it misses one. And look 00:11:12.049 --> 00:11:14.409 carefully near the edges of your edit. Did anything 00:11:14.409 --> 00:11:15.990 weird happen in the background that you didn't 00:11:15.990 --> 00:11:18.970 intend? Good checklist. So this amazing tech 00:11:18.970 --> 00:11:22.169 is in Google AI Studio now, but the sources were 00:11:22.169 --> 00:11:24.690 clear. This is just the flash version. Right. 00:11:24.809 --> 00:11:27.009 That's the key point. This is the fast accessible 00:11:27.009 --> 00:11:29.190 version, the one winning the ELO scores. But 00:11:29.190 --> 00:11:31.610 there's likely a more powerful full version of 00:11:31.610 --> 00:11:34.730 Gemini behind it, maybe still to come. And developers 00:11:34.730 --> 00:11:37.149 can already tap into this? Yep. The API is available 00:11:37.149 --> 00:11:39.350 so people can build tools that integrate this, 00:11:39.789 --> 00:11:42.450 allow editing through, like chatting with the 00:11:42.450 --> 00:11:45.250 AI, or combine multiple images in really sophisticated 00:11:45.250 --> 00:11:47.950 ways. The potential is huge. OK, so let's boil 00:11:47.950 --> 00:11:50.330 it down. The big takeaway here. I'd say it's 00:11:50.330 --> 00:11:52.990 that NanoBanana's multimodal core gives it just 00:11:53.070 --> 00:11:56.970 unmatched consistency, especially for tiny edits 00:11:56.970 --> 00:11:59.750 and keeping characters the same across big changes. 00:11:59.909 --> 00:12:02.809 Right, but the challenge, the thing users need 00:12:02.809 --> 00:12:05.210 to remember... You have to respect its limits, 00:12:05.669 --> 00:12:08.950 especially with complex art styles, and always, 00:12:09.190 --> 00:12:12.629 always break down big ideas into small, simple 00:12:12.629 --> 00:12:14.870 steps. Don't try to do everything at once. So 00:12:14.870 --> 00:12:17.169 the call to action is... Pretty clear. Yeah. 00:12:17.250 --> 00:12:19.590 Head over to Google AI Studio. Try some simple 00:12:19.590 --> 00:12:22.289 edits first. Just feel that consistency. It really 00:12:22.289 --> 00:12:24.909 does change how you think about what AI can do 00:12:24.909 --> 00:12:27.590 with images. And maybe final thought to leave 00:12:27.590 --> 00:12:31.539 people with, that 200 ELO gap. It suggests this 00:12:31.539 --> 00:12:34.240 isn't just a small step forward. It feels like 00:12:34.240 --> 00:12:36.899 a fundamental shift in defining what realistic 00:12:36.899 --> 00:12:39.539 even means for digital images, right? It really 00:12:39.539 --> 00:12:42.179 does. What other fields, you know, beyond photos, 00:12:42.460 --> 00:12:44.600 architecture, product design, might get completely 00:12:44.600 --> 00:12:47.539 reshaped by this kind of seamless, super precise 00:12:47.539 --> 00:12:50.379 AI integration, something to think about. Definitely 00:12:50.379 --> 00:12:52.360 something to ponder. Thank you for sharing your 00:12:52.360 --> 00:12:54.379 sources with us today. We really hope this deep 00:12:54.379 --> 00:12:56.019 dive was useful for you. We'll catch you on the 00:12:56.019 --> 00:12:56.360 next one.