I actually think the first one is better.
The perspective is also off in both of them. They are the wrong size, too large, for where they are in relation to the background. Said another way, the camera POV is closer to them and farther from the background, but differently so. Which gives the “cardboard cutout in front of …” look.
It probably doesn’t help that the source photo was probably a Turneresque colorized version from the original black and white episode.
There are issues with both apps regarding this comp:
Depth of field: background is too soft; jars and cat are razor-sharp; Barney is mid-sharp. Mixed planes break a single focal plane.
Lighting direction and quality: scene light is soft, overhead/left. Cat and pickle/hand have harder, cooler light.
Shadows: no convincing contact shadow under cat paws or between jars and counter; hand/pickle cast none on the face.
Color grade / white balance: background has warm, colorized TV look; cat and pickle are modern, saturated, cooler.
Edge: hard cutout edges around ears, paws, and the hand; haloing on sleeves.
Reflections: highly reflective jar glass/lids show no reflection or color bleed from the cat.
Geometry: hand orientation to the pickle/mouth is off; bite angle doesn’t align with jaw; uniform folds show AI distortion.
These could be fixed in Photoshop by applying blur match, warm grade (a warm sepia LUT), contact shadows, scale adjust, edge feather + grain, and minimal reflections, but that would take time. I was hoping Sora and Gemini would combine the photos and render my prompt a bit better on the first generation, but they just did OK (at least they got the humor across: a riff on “The Pickle Story,” Season 2, Episode 11. Aired December 18, 1961; Barney called them “kerosene cucumbers”; supposedly Don Knotts’ favorite episode). Maybe their next models will do better. I get better better results with Firefly (its new “Boards” feature is great).
Here are the photos I started with:
Any cat would tell you that that’s a feature, not a bug.
Is that the only reference you gave it for your cat? If so, it’s impressive that it was able to completely change the pose. Though I suppose that cats being heavily represented in the training data would help a lot with that.
Yes, Benny is a Korat cat; they pretty much all look alike. Great cat, though! In contrast to my Dwarf/Highlander mix, Ollie. That cat’s tiny, but a disaster on four paws.
I tried a text-only prompt in Sora. Both outputs were black and white, and I told it to convert one of those to color:
Aunt Bea is feeding Barney Fife a large pickle from a jar in her kitchen. A gray cat is watching.
As you can see, these color versions look wrong, too. It seems like the training set for color photos of Barney Fife tell the AIs to make fake-looking inages (thanks to colorization).
A few tweaks to your image # 3:
I like to take drawings I run across and plug them into AI and tell it simply to convert it to a realistic photo without text. Here’s a Far Side in the new Gemini.
Which isn’t so bad. Here it is in Copilot
The cowboy is pretty good, but the buffalo are sheep, and the campfire is a burning chicken.
Here’s a couple more from Copilot
I sincerely consider this a victory. It also turned a rock into a bunny!
The Gemini one doesn’t look nearly as photorealistic as the other one.
The livestock in the Gemini image look more ovine than bovine, as well.
But there is, I think, something quintessentially Far Side in a burning chicken.
I also like turning interesting “text images” into images, telling the AI to create an image based on the text content. Here’s a recent one from Sora
My experience with Gemini thus far is that it’s super impressive for combining separate elements into one cohesive image and complete garbage for making compositional changes after the fact.
For instance, I put some characters on a beach and wanted one trying to wave seagulls away from his food. The overall addition of the characters was quite good but it only added one gull not very close to the guy. So I asked for more and it failed. Then I edited the image in paint and said “Put gulls here” and it sorta did it but not great. Later it had one bird sitting on the beach and facing the other way – it would not mirror the bird no matter what. All it needed to do was flip the bird (heh) but it would change the bird slightly without mirroring it or delete the bird or leave the bird the same but change other image elements…
So I rate the new Gemini the score of Zebra Stripes Gum: Amazing for the first ten seconds and progressively more disappointing every second after.
The livestock in the original cartoon look like sheep. I thought they were sheep until i read the comments, and went back to the cartoon and read the text.
I see your point. But … it depends on your focus.
Big difference whether you focus on the herd’s faces & horns which are very buffalo or their bodies which are indistinct, fluffy, & monochromatic white without buffalo’s characteristic long-haired furry hump. But still w a very large not-hairy shoulder hump by sheep standards.
I’d put the AI rendering them as sheep in the “wrong, but understandably wrong” category. If they were rendered as dogs or rabbits or … that’d be hopelessly stupidly wrong.
IMO YMMV …
Unrelated to the above …
Clearly being overrun by a buffalo herd in a narrow canyon would mean our cowboy would be pretty much instant people-burger. Or, more lime, buffalo toe-jam.
Would he fare much better w sheep? Or would it be the same death, just needing a few more tramplings to complete the burger- / jam-fication?
I wonder if it’s worth doing the setup in Gemini & then porting it over to another platform…
Possibly. My images so far were just to put it through the paces but if I actually really wanted a specific complex image then I might start in Gemini and then fix all the other elements using a local client for inpainting and whatnot.
They do kinda have a hump, but i think the faces and horns look like sheep, too.
At the time that I read the Gemini Nano Banana article I was working on a prompt idea for an action photo capturing the moment a soccer ball hit someone’s face, distorting the face and the ball. I had a version open in Gemini and told it to swap the girl with a creepy alien, then to swap the ball with Mars. It did a pretty good job of both.
Summary
An action shot of a tween girl playing soccer. The image captures the very moment that the soccer ball contacts the side of her face, flattening the soccer ball against her cheek. Iphone 15 photo with shallow dof. Other players are visible in the background. Her jersey says SPORTS TEAM
(BTW, Sora/Copilot was the only model tested that “got” what I was asking for.)
The “books you hated in high school” thread made me think of this.
