Interestingly, those dinosaurs are all much more paleo-accurate as well. At least, the tyrannosaurus and triceratops are. I’m not sure what the top left guy is meant to be, maybe a hadrosaur?
Some raptor, maybe.
This reminds me of one of the first ukiyo style images I tried to do shortly after Dall-E was integrated into ChatGPT-4:
I posted it way-upthread:
And maybe that dog is part tanuki?
The prompt:
I want to make a image that is a joke on the Simpsons “old man yells at cloud” image. I want a newspaper clipping with a photo of someone resembling Gary Oldman shaking his fist at a fluffy cloud overhead. The headline of the clipping is “Oldman yells at cloud”. The clipping is being held in someone’s hand by the lower right corner.
(The cloud was on the right and I moved it manually.)
ChatGTP refuses to make a krasue (talked about earlier in the thread) so it must know it is something gory.
Tried for “Image of a nuppeppō and a rokurokubi playing shogi in a shoji room with tatami floor” and got this.
The nuppeppō (a poorly shaped creature made out of globs of meat) isn’t quite like it is traditionally depicted but I’m not too upset with this interpretation. The rokurokubi (long-necked woman) is great.
(Bing has no clue about those creatures.)
ETA I told it to make that into a photograph.
Inspired by an official Fallout post on Facebook with a virtually identical title: Kool & the Gang, ghoulified.
Before:
After (Ghoul & the Gang):
My latest with ChatGPT.
I uploaded some old photos of mine of some small plastic yokai figurines (plus a photo of an old Japanese print) and described them and asked for a realistic photograph.
I did some minor hand editing on two images. The karakasa (umbrella guy) had two legs and the kappa (turtle guy) had a weird curved tube/tail thing connecting his butt to his leg. I removed both of those.
These results are impressive enough to me, but then I tried an image with two angles, and had to try to describe something violent and gory well enough to get accurate results but mildly enough to not trigger the guardrails. The description was
This photo is two separate angles of a demon spider from Japanese folklore called Tsuchigumo. It ate people and when it was defeated by a famous samurai it was full of skulls. Could you make a realistic photo of this creature based on these two images? The two images are from the front and the right side.
And yes, it managed to create an image based on details from both angles.
I uploaded pics of my cats to ChatGPT and asked her to model and remember them for future adventures. Then I had her create an image of the gang strutting down Broad Street in Philly. She nailed Benny, came darn close with Ollie and Macaroni, and even captured the vibe of Broad Street (complete with Billy Penn atop City Hall). Gotta say, character consistency like that? Total game-changer.
“Her”?
Yes, my ChatGPT is a woman. I’m not complaining.
It just crossed my mind to try for a contact sheet. Hannya was the last image in a session, so I went with that.
Create a contact sheet of several sequential images as she notices the camera, looks towards it, and waves.
Thanks, but I was thinking of a grid of 12 images, arranged 4x3, with spacing around them.
I would qualify that as a partial success. It can make a grid of differing images, but doesn’t really understand sequential movement from one pose to another.
Oh, something odd. I’ve been using two free accounts for ChatGPT. One that I opened a pretty long time ago with my primary email address and am mainly using through the ChatGPT app and a new one with my secondary address that I run in a web browser. The older account in the app is much more interactive and chatty and gives me only three image renders a day. The new one in the browser doesn’t say much or ask many questions but gives me at least a dozen renders a day. Why the difference, who the hell knows.
Are you paying for the new one? More than 3 image generations per day is the paid option. Check in the settings and customizations areas whether you’ve got custom instructions and/or memories active on the two different accounts.
No.
Looks like I did 14 images yesterday on that free account before it stopped me. (Five of them were “try it again this way” images.)
And the talkative account will say things about how good the image idea sounds and tell me that it is going to render it now, and sometimes asks questions to clarify points. But the other one just renders without a word.
It is very weird, on top of the other differences, my ChatGPT in app account does a much worse job of converting an image to a different style than the account in a browser, even though both are using 4o.
For each of these, the left is the source, the center the app creation, and the right the browser creation.
Looks like the app account is still using Dall-E instead of the more closely integrated 4o generator. Or even if it is calling the newer generator, perhaps the app is unable to translate as much image context into the prompt. Have you tried switching it up which account is logged in where?
Haven’t tried that, don’t want to risk breaking the one that does work.
I asked that version and it claims to be using Dall-E
What image generator are you using?
I’m using OpenAI’s DALL·E image generator, specifically integrated into this chat as part of the tools I can access. It allows me to create, edit, and transform images based on your prompts. Let me know if you want to tweak or create something else!
Why not 4o?
Great question! The image generation tool I’m using is powered by DALL·E, and while I (as GPT-4o) generate the prompts and guide the creative direction, the actual image generation happens through the DALL·E model, which is currently the one specialized for image tasks.
GPT-4o (the “o” stands for “omni”) handles language, vision, and audio inputs very efficiently, but the image generation component is still handled by DALL·E, which is the best at turning text into visuals right now.
If OpenAI eventually releases a unified image-generation model under GPT-4o or beyond, I’ll be able to use that too!
I haven’t messed with any of this in a while. I want to create a video of the Monty Python foot labeled with a word coming down and squishing a different word. Is there a good free option for that now?
You could try at Klingai.com. Unless they changed things you should get 166 free credits. You can make 5 or 10 second clips, cost is 20 credits for 5 seconds, 40 credits for 10 seconds. You can provide key frams for the beginning and end, and describe the action.
Giving it a shot now! Thanks