Towards the topic title, I was really amazed when ChatGPT was able to look at a car ad - and all the scrollable pictures. And one dent (it was a Jaguar) that I claimed I might be able to rubber-mallet out from inside the hood, or perhaps use suction cups, or if nothing else, drill, insert an expander, pull hard. fill and paint (the paint job overall wasn’t that great - something like sandblasting on the hood).
I didn’t really care about the looks - I wanted a stopgap car for a month or so, to be able to travel for test drives. This was a Jaguar listed at £900 ($1216)
The ad: (heh, I assume visible from the USA)
Third picture scrolling to the right…
The superbly concise reply to my thoughts: (removed spaces for brevity)
What the photos show (and what it implies) 1) The dent is not a “tap it out” dent.
It’s a sharp, creased dent right on the swage/character line of the offside front wing (by the wheel arch).
Those are the hardest dents to “bang out nicely” because the metal has stretched and the crease sits on a stiffened line. Paintless dent repair might reduce it, but it’s unlikely to disappear cleanly. Drilling/pulling is bodyshop territory and risks rust + a worse finish. 2) The bonnet looks tired (your “sandblasted” call is fair).
From the front-on photo, the bonnet finish looks patchy/hazy — could be lacquer failure/oxidation/stone rash.
Not fatal on a £900 stopgap, but it’s a signal of “kept running, not cosmetically cherished.” 3) Most important: that wing hit raises a mechanical question.
A whack that creases a wing near the arch can also mean:
the wheel took a knock
alignment could be off
suspension bushes/arms could be stressed
So the right follow-up isn’t “can I fix the dent cheap?” — it’s:
“Does it drive dead straight and vibration-free, and are the tyres wearing evenly?”
Your “cheap stopgap” lens: is this worth pursuing?
Maybe — but only as a local, quick viewing (because £900 + long MOT is tempting), and only if you can quickly rule out “hit damaged geometry”.
I too was truly impressed. ChatGPT (I use the others too, yet started to pay for this one when they cleverly made it drop down to ChatGPT-4 for several hours after using it for like four hours. And I could tell a difference though 4 was adequate.)
BTW, I told it what you said (above the parens)
That’s genuinely lovely to hear — thank you for sharing it. And honestly, it makes sense why they reacted that way.
And we both commiserated in a once nice Jaguar getting this poorly treated.
Tonight I was trying out some old prompts on newer models tgat gave come out since then and happened to notice for the first time how different the new version of the ChatGPT/Sora/Copilot inage generator is from the old one. I did a few comparison images of prompts run on the old system vs the exact same prompts in the new version. The new version is more detailed and arguably more realistic but in almost every case I liked the old version more.
(In all of these the old is the left or upper image)
I prefer the earlier versions too. The later ones aren’t realistic; they’re hyper-realistic. Which might be all the rage in the auto-enhancement features of modern mobile phone cameras, but somehow they look more fake than the simpler lower-res older versions.
It certainly helps that the subject matter of each of these vignettes is sort of dark & weird & brooding. That emotion is well-suited to the sort of sepia-toned or noir look of e.g. the woman on the bus bench or the clown.
I suspect there are some cheerful topics & moods where the newer AI would create images that are much more aesthetically pleasing. DG doesn’t seem to have too many prompts like that though.
I think that the supernatural/horror aspects work better with a certain amount of non-realism. Trying to make it realistic is sort of like those “Homer Simpson as a real person” illustrations that just look messed up and not in a cool horror sort of way. I’ve noticed the same for a n umber of fantasy concepts – what looks good as a painting looks awful when applied through a realism filter.
The “new” one I think looks best is #1 and that’s just a girl with a capybara. Worst is probably second from the bottom where realism didn’t do those shuffling monsters any favors.
The new model actually comes closer to capturing the look of the creature than I managed with many tries on the old model. Garamon was a guy in a goofy looking rubber suit with two limp hands held straight out ahead like some sort of zombie kangaroo in big gloves. I wanted the AI version to look like a guy in a goofy looking rubber suit with two limp hands held straight out ahead like some sort of zombie kangaroo in big gloves, but nothing ever captured the fake look I wanted.
I tried a number of sample images. With these renders I had used two sets of images of a Garamon figurine
and described them in the prompt
Summary
Garamon creatures are primarily burgundy and bone white. They have huge wide upturned mouths with thick pink lips. Green-yellow eyes. Vertical slit nostrils. Their bodies are covered with large jagged vertical burgundy scales on their sides and backs and thick, bone white hide on their bellies. They have very large, bony, bone white three-fingered hands on very short burgundy arms. Thick bone white legs ending in bony three-toed feet. And a short, whip-like bony tail.
And now I see that the new OpenAI image generator no longer makes the fun blobfish images that I like to work with. Old model on left, new on right
Summary
A farmer picking fruit from a tree on the left of the scene. It is like an apple tree, but it is covered with blobfish instead of apples. He is reaching out with one hand to pick a blobfish fruit on the tree. At his feet is a woven bushel basket filled with blobfish that he has already picked, and on the ground under the tree are several shriveled, decaying over-ripe blobfish that fell from the tree. The farmer is wearing typical farmer clothes including overalls, but his head is a blobfish. There is a blobfish with legs climbing against the farmer’s leg like an eager puppy. In the background A blobfish headed son in overalls and a blobfish headed daughter in a plain modest Summer dress, both around 5 to 10 years old, are playing, tossing a blobfish back and forth. The son has a blobfish for a head. The daughter has a blobfish for a head. In the distant background other blobfish trees are visible. Profile view widescreen Kodachrome DSLR photograph with shallow dof.
Summary
A photorealistic image of an emo girl with layered black hair and subtle blue streaks, fair skin, and bold black eyeliner. She wears a black cropped tee, ripped black fishnet tights, and chunky black lace-up combat boots. She’s squatting on a rainy city pathway, with misty trees and blurred lights in the background. Blobfish are tumbling from the sky and accumulating on the ground. The blobfish are in random orientations. Her expression is distant and moody as she rests her head on one hand, wearing black fingerless gloves and a black choker. The atmosphere is melancholic, wet, and cinematic.
Summary
Nyctibius grandis, mouth open, is sitting on a blobfish. The blobfish is sitting on a Pelochelys cantorii. The Pelochelys cantorii (wearing a tiny Hejaz fezz) is sitting on a large, predominantly red, rippling flying carpet with the patterns of an afghan rug and tattered yellow fringe. The carpet is flying over Tokyo Tower. In the far distance Godzilla is grappling with an Adamski type flying saucer. Realistic Iphone 15 photo with shallow dof and forced perspective.
Oddly, on this prompt the new model does sometimes make the blobfish almost like the old model, and even the “off” ones are better than in the renders for the other prompts. As for other elements of the image, the old model makes a much more accurate nightjar (bird) and neither model gets the right turtle species. (And in this example at least the old model screwed up Godzilla)
As far as I can tell you have no choice but to use the new model in Copilot and Sora. Bing still uses the old model but their prompt window is ridiculously small and censorship ridiculously high. There is a way to force ChatGPT to use the old model by using this link:
It’s not a matter of hate, it’s a matter of usability. And if someone else posts their image on ImgBB and expects me to see it, I can’t choose from any alternatives.
