Latest / AI Fire Daily / #650 Neil: 4 AI Image Models Tested and the Results Surprised Me

Transcript

#650 Neil: 4 AI Image Models Tested and the Results Surprised Me · AI Fire Daily

0:00
We have we have all been there, you know, you
0:02
type out a prompt. Oh, absolutely. A flawless
0:04
masterpiece instantly appears on your screen.
0:06
It feels like magic. But then you zoom in. Right.
0:10
You always have to zoom in. You zoom in and the
0:12
guy in the background has like seven fingers.
0:14
The neon store sign is speaking complete alien
0:17
gibberish. Yeah, totally. And the illusion just
0:19
shatters completely. It shatters very quickly.
0:21
Yeah. We see the messy reality right behind the
0:25
algorithms. Welcome to the deep dive. I am Incredibly
0:29
glad you are here with us. Today we have a very
0:32
specific mission. Google just released NanoBanana
0:36
2 .1 into the wild. Finally. Right. And it promises
0:40
vastly better image quality. It claims highly
0:42
accurate text rendering. It boasts significantly
0:46
improved prompt following. And it is four times
0:49
cheaper. Wow. Cheaper than NanoBanana Pro, that
0:51
is. I mean, those are some incredibly massive
0:54
claims, but we really need to see hard evidence.
0:56
We cannot just trust the marketing brochure.
0:58
Exactly. We're testing four major competitors
1:01
today. Nano Banana 2 is our baseline model. Nano
1:05
Banana Pro brings the high quality visual details.
1:08
OpenAI's GPT Image 2 .5 is the heavy hitting
1:11
heavyweight. And Nano Banana 2 .1 is our new
1:15
challenger. It's quite the lineup. It is. We
1:18
put these four models through five rigorous tests.
1:21
Real -world visual tests. We want to separate
1:24
the marketing hype from actual performance. Okay,
1:27
let's unpack this. We have to start with the
1:30
foundational elements, I think. Right. Before
1:33
we test complex physics, we start with reality,
1:36
the basic foundations. We need to see perfectly
1:38
natural lighting. We need to see believable glass
1:42
reflections. We absolutely need readable text
1:45
in a static environment. So for the first test,
1:48
we wanted an everyday scenario. Picture a busy
1:51
neighborhood coffee shop. It is a slightly cloudy
1:54
afternoon. Very specific mood. Yeah, exactly.
1:57
Two friends in their late 20s are chatting. We
2:00
requested a 35 millimeter lens right at eye level.
2:03
We wanted realistic facial expressions. We wanted
2:06
relaxed body language. Right. We even threw in
2:08
a subtle curveball. We asked for light condensation
2:11
on the windows. We gave the models very specific
2:14
text challenges, too. The main storefront sign
2:16
must say, good coffee, good people. The window
2:19
lettering needs to read, freshly brewed daily.
2:22
The takeaway coffee cups say, morning club. And
2:26
the poster inside. Yes, and there is a promotional
2:28
poster inside. It says, buy one, get one, 50
2:32
% off. The comparison results here were incredibly
2:34
revealing. GPT Image 2 .5 won this round overall.
2:38
It really did. It gave us a highly accurate storefront.
2:42
The human characters looked completely natural.
2:44
The lighting perfectly suited that cloudy afternoon
2:47
vibe. The smaller lettering on the cups wasn't
2:50
entirely perfect, but the overall image looked
2:52
highly convincing. Nano Banana 2 .1 had very
2:57
mixed results here. Yeah, it was weird. It handled
2:59
the large storefront text fairly well, but it
3:02
completely failed on small coffee cup labels.
3:05
The overall result just looked slightly less
3:07
polished. Nano Banana 2 is actually much worse.
3:10
It repeated freshly brewed daily in multiple
3:13
places. Just everywhere. Yeah, it just pasted
3:16
the text everywhere. Some smaller signs contained
3:18
completely distorted and complete text. And Nano
3:22
Banana Pro made a classic mistake. It reversed
3:24
the storefront name in a reflection. I'm looking
3:28
at these results and I just don't get it. It's
3:30
baffling, I know. We have a multi -billion dollar
3:32
system that can generate perfect photorealistic
3:34
condensation on a window, but it spells morning
3:37
as M -O -R -N -N -G. How is it so smart and so
3:41
incredibly dumb at the same time? Because it's
3:43
not thinking. Right. Generating AI text is like
3:46
giving a child a massive box of Lego bricks.
3:48
You ask them to build a word they do not know
3:50
how to spell. Oh, that's a good way to look at
3:52
it. They know what the shape of an S looks like,
3:55
but they might build it backward. They might
3:57
use the wrong colors entirely. They just do not
3:59
know what the letter actually means. That is
4:02
a brilliant way to picture it. People constantly
4:04
forget what an AI model actually is. A math system
4:08
trained to recognize and recreate complex visual
4:11
patterns. Exactly. It is fundamentally just math.
4:14
It does not actually understand our world. It
4:16
does not think. Why do these incredibly advanced
4:19
systems still struggle with simple spelling on
4:21
a tiny coffee cup? Because they do not really
4:24
read text. They process text as complex visual
4:28
shapes. They do not understand linguistic characters
4:30
at all. So they're just drawing? Yeah, the system
4:33
just tries to draw detailed squiggles. Squiggles
4:36
that statistically look like human letters. It
4:39
is mimicking the visual appearance of writing.
4:42
It is not actually writing. So the AI draws text
4:46
like shapes, not as actual readable letters?
4:49
Yeah. That shape matching works OK for simple
4:51
textures, but it really struggles with strict
4:54
underlying logic. So it cannot render text because
4:57
it is just mimicking two -dimensional shapes.
4:59
But what happens when we force the AI to think
5:02
in three -dimensional space? That's where it
5:04
gets messy. Right. Text is flat. Real life has
5:07
deep depth left, right, and dynamic perspective.
5:11
That brings us directly to our second test. Let's
5:14
move from static text to spatial awareness. We
5:17
set up a rainy gas station evening. We were extremely
5:21
specific about the camera placement. Super specific.
5:23
Yeah. The camera is exactly 30 centimeters above
5:26
wet pavement. It is approximately four meters
5:29
away from the human subject. And it is tilted
5:32
up exactly 20 degrees. This creates a highly
5:35
dramatic low angle perspective. We gave the models
5:38
very strict left and right rules. A woman is
5:41
standing near the center. she holds an open red
5:44
umbrella in her left hand. And a white coffee
5:47
cup in her right hand. Yes. Both hands must be
5:49
anatomically correct. Which is always a gamble.
5:52
Always. And there is a white four -door sedan
5:54
parked behind her. The driver's side is facing
5:57
the camera. And there is a gas pump on the far
5:59
left. Everyone struggled with this complex prompt.
6:03
Every single model failed somewhere. GPT image
6:06
2 .51 for overall cinematic perspective. It did
6:10
look beautiful. It also got the hand placement
6:12
perfectly right. But it faced the white sedan
6:14
completely backward. Which ruins the whole shot.
6:18
Nano Banana 2 .1 got the low angle correctly,
6:21
but it put the umbrella in the wrong hand. And
6:24
it put the car backward too. Yeah. The woman
6:27
appeared much too far away as well. Nano Banana
6:30
Pro gave us great cinematic lighting, but the
6:34
umbrella was in the wrong hand again. And the
6:37
camera angle simply wasn't low enough. I have
6:40
to admit something here. I still wrestle with
6:42
prompt drift myself. Oh, me too. Absolutely.
6:44
You spend 20 minutes typing 50 careful details.
6:48
The machine just casually forgets five of them.
6:50
It is incredibly frustrating. What's fascinating
6:53
here is how the failure happens. These systems
6:56
can render ultra photorealistic rain. They can
6:59
make wet pavement reflect neon light beautifully.
7:02
Which is visually stunning. It is. But that requires
7:05
intense mathematical lighting calculations. Yet
7:09
they cannot grasp basic spatial geometry. Why
7:12
does a system that can generate photorealism
7:14
completely fail at knowing left from right? Because
7:18
spatial relationships are fundamentally different
7:20
from lighting. Geometric logic is processed differently
7:24
than surface texture. How so? Texture is just
7:27
localized pixel prediction. Spatial awareness
7:30
requires understanding deep three -dimensional
7:32
space. These models lack a true structural grounding.
7:36
They just guess. Yeah, they do not build a mental
7:39
map of the scene. They just predict adjacent
7:41
pixels. Right. Spatial logic is totally different
7:43
from just recognizing static pixel patterns.
7:47
It gets even more complicated from here. Wait
7:48
until things start physically moving around.
7:51
We have seen AI fail at static left and right
7:53
instructions. Let's escalate this further. What
7:57
happens when we introduce gravity, momentum,
7:59
and dynamic physical movement? Our third test
8:02
is a typical backyard trampoline. A 25 -year
8:05
-old man is jumping very high. He is exactly
8:07
80 centimeters in the air. And we used a professional
8:11
sports photography style for this. Right. We
8:13
asked for a 70 -millimeter lens and a blazing
8:17
-fast 1 2 ,000th second shutter speed. The physical
8:20
details really matter here. That fast shutter
8:23
speed freezes motion instantly. His medium -length
8:26
hair needs to be lifting up. His loose gray t
8:30
-shirt should show realistic folds. A plastic
8:33
water bottle is bouncing nearby. The safety net
8:36
must be completely intact. And he must be wearing
8:40
sneakers. And Nano Banana 2 .1 finally takes
8:44
a clear win. It absolutely nailed the dynamic
8:48
physical motion. It really impressed me. The
8:50
bouncing water bottle was physically believable.
8:53
The sneakers looked completely correct in midair.
8:55
The entire jump felt incredibly realistic and
8:58
grounded. GPT image 2 .5 really struggled here.
9:02
It made him jump way too high. The water bottle
9:05
just floated awkwardly in space. Like it had
9:08
no gravity. Exactly. The hair movement wasn't
9:10
entirely convincing either. Nano Banana 2 was
9:13
even weirder. It made the guy completely barefoot.
9:16
Which we explicitly said not to do. Right. His
9:19
jumping pose was terribly unbalanced and awkward.
9:22
The water bottle's movement felt highly unnatural.
9:25
Nano Banana Pro left the bottle standing perfectly
9:28
upright. The safety net also looks very broken
9:30
and incomplete. Generating all of this from scratch
9:34
is just wild. It really is. Whoa. Imagine scaling
9:38
to a billion queries. It is actually staggering
9:42
when you think about it. It really boggles the
9:43
mind. Every time you hit enter, this model is
9:46
hallucinating reality. It is effectively hallucinating
9:49
the physics of a bouncing bottle. It builds it
9:52
from scratch based solely on latent space patterns.
9:56
And millions of people are requesting this simultaneously.
9:58
Exactly. How does the AI understand the physics
10:01
of a bouncing water bottle without actually having
10:04
a physics engine? It relies entirely on massive
10:07
data sets of frozen motion. It has seen millions
10:10
of high -speed sports images. It essentially
10:13
hallucinates the correct physical trajectory.
10:16
So it's just guessing where things should be.
10:18
Yes. It predicts where pixels usually go during
10:21
motion blur. It does not calculate mass or real
10:24
gravity. It mimics visual patterns of falling
10:27
objects instead of calculating real physics.
10:29
Spot on. Mid -roll sponsor read placement. Welcome
10:33
back. So we have conquered dynamic physical movement.
10:37
Let's pivot from the physical world to abstract
10:40
data. Can an AI organize complex scientific information?
10:44
This is a tough one. We wanted to see a vertical
10:47
4x5 layout. The topic was decompression sickness.
10:51
It is also known as the bends. We wanted an educational
10:54
infographic aimed primarily at beginners. The
10:57
requirements were very strict. We needed accurate
11:00
science on human nitrogen absorption. We needed
11:03
the facts on clear bubble formation. And it had
11:06
to be readable. Right. We wanted simple accessible
11:09
English and absolutely no invented statistics
11:11
or fake medical claims. GPT Image 2 .5 wins this
11:15
round, clearly. The infographic was highly detailed
11:18
and structured. The text rendering was incredibly
11:21
strong. The scientific content was actually surprisingly
11:23
solid, too. It was. Nano Banana 2 .1 had a very
11:27
clean visual layout. It was extremely easy to
11:30
read, but it had a massive glaring problem. Oh,
11:33
yeah. It was bad. It gave overly specific, hallucinated
11:36
medical advice. It gave context -free ascent
11:39
rates for scuba divers. That is a terrifying
11:42
safety issue. Nano Banana Pro was very messy
11:45
overall. It duplicated a whole section entirely,
11:48
and it had blatantly inaccurate nitrogen absorption
11:51
facts. And Nano Banana 2 just distorted several
11:56
key scientific terms completely. Here's where
11:58
it gets really interesting. Think about the last
12:00
time you searched for a quick explainer online.
12:02
Oh, all the time. The prettier the design is,
12:05
the more easily we believe the data. We trust
12:07
a beautifully clean layout instinctively. We
12:10
just drop our natural guard. We do. We totally
12:12
do. We assume the math is correct because it
12:14
looks professional. Which is a very dangerous
12:17
psychological trap. It makes the hallucinations
12:19
much harder to spot. We lower our critical thinking
12:22
for good graphic design. Based on these hallucinations,
12:25
is it safe to trust AI for educational graphics
12:27
right now? No. It is generally not safe at all.
12:30
Visual aesthetics often mask deep factual inaccuracies
12:34
entirely. AI outputs often look highly professional
12:37
and polished. They look like a real textbook.
12:39
They do. And that professionalism bypasses our
12:42
natural skepticism completely. You have to verify
12:44
every single medical claim manually. You cannot
12:47
outsource your scientific accuracy to a diffusion
12:50
model. Always fact check, because it prioritizes
12:53
pretty layouts over actual scientific accuracy.
12:55
That brings us to our final most personal challenge.
12:59
Right. For our final test, we moved from general
13:01
data to personal identity. What happens when
13:04
you put yourself into the AI's hands? This was
13:06
fun. We used a standard 16 by 9 YouTube thumbnail
13:10
format. We uploaded your actual personal portrait
13:13
as a reference. Yes, you did. We wanted to preserve
13:15
your exact facial identity and structure. The
13:18
prompt was quite ridiculous. I am wearing a bright
13:21
yellow shirt. I am reacting with total exaggerated
13:25
surprise. An anthropomorphic boxing banana is
13:28
punching me directly in the cheek. With shiny
13:30
red boxing gloves. The banana wears shiny red
13:32
boxing gloves, yes. And the bold text must say,
13:35
Nano Banana 2 .1. It needs vibrant colors and
13:38
highly dramatic YouTube lighting. Nano Banana
13:41
2 .1 takes the crown here. The colors are incredibly
13:44
vibrant and punchy. My face is highly recognizable
13:47
and accurate. It really looked like you. The
13:49
dynamic punch is visually clear and hilarious.
13:52
And the text is perfectly readable and well -placed.
13:55
GPT Image 2 .5 had a truly great punch effect,
13:59
but it badly distorted your human face. It also
14:02
overcrowded the text completely. It was way too
14:04
cluttered. The text was just far too massive
14:06
for the frame. Nano Banana Pro made the banana
14:09
look much too friendly. The punch lacked any
14:11
real visual impact or kinetic energy. And Nano
14:14
Banana 2 made the background far too dark. Your
14:17
surprise expression felt completely disconnected
14:19
from the actual punch. It just looked weird.
14:22
This raises an important question. Why did the
14:24
newer model excel at simplicity here? Why was
14:27
the layout so much cleaner? Why did the older
14:30
GPT 2 .5 overcomplicate the thumbnail compared
14:32
to the newer Nano Banana 2 .1? Because newer
14:36
models are tuned very differently today. They
14:38
are better tuned for modern visual hierarchy.
14:41
They understand the deep importance of negative
14:44
space and design. Less is more. Exactly. Older
14:47
models just throw massive pixels at the canvas.
14:50
They desperately try to fill every single empty
14:52
gap. Nano Banana knows exactly when to pull back.
14:56
GPT overcomplicates the effects, while Nano Banana
14:59
just understands basic visual hierarchy. If we
15:02
connect this to the bigger picture, being the
15:04
newest model doesn't automatically mean being
15:06
the best at everything. We saw wildly different
15:09
strengths today across the board. Google's newest
15:11
release did not dominate completely. GPT Image
15:14
2 .5 remains the absolute champion for pure photorealism.
15:19
It is best for complex camera instructions and
15:21
strict framing. and it completely dominates with
15:24
detailed text -heavy infographics. But Nano Banana
15:27
2 .1 is the new go -to for dynamic action. It
15:30
handles realistic physical motion beautifully,
15:33
and it is absolutely perfect for highly stylized
15:36
YouTube thumbnails. You absolutely need a multi
15:39
-tool approach. You cannot rely on just one single
15:41
generator for everything. You have to pick the
15:44
right tool for this specific job. You really
15:45
can't. The technology is too fractured right
15:48
now. So what does this all mean? The funniest
15:51
part of this entire deep dive, even the absolute
15:54
best, most expensive models still make bizarre
15:56
mistakes. They really do. They fail in completely
15:59
unpredictable, hilarious ways. Like backwards
16:02
sedans. Or entirely mutated human hands. Exactly.
16:06
Mistakes that a quick glance might completely
16:08
miss. Getting a 90 % perfect image is incredibly
16:12
easy now. Getting a 100 % perfect image is still
16:15
incredibly hard. Very hard. The illusion is incredibly
16:19
fragile. The magic breaks the second you look
16:21
closely. Now, we're curious. Based on these rigorous
16:25
tests, which AI image model would you trust with
16:28
your next big project? Thanks for joining us.
16:31
Right. See you next time.