Gemini has been upgraded to a new renderer.
I never thought to ask this before now, and I’ve never noticed it mentioned anywhere else, but ChatGTP/Copilot can create PNGs with areas of transparency if you ask it to.
Here’s a rendering from Gemini 2.5 Flash:
The all-female remake of Fight Club nobody asked for.
Made with Sora. Much better likeness of Kendrick and Henstridge than I ever got out of ChatGPT.
Experimenting giving ChatGPT and Sora the same prompt:
ChatGPT
Sora
ChatGPT
https://imgur.com/zQ86l7z
Sora
ChatGPT
Sora
“tesla cybertruck fitted with the back to the future delorean time travel gear. parked at the hill valley courthouse square, with the courthouse in the background.”
ChatGPT
Sora
If BTTF was being remade today, that’s a logical alternative vehicle for a mad scientist who needs the vehicle to have a steel body and believes somehow that the vehicle has “style.” 
They might need to promise Elon a bit part in the movie to allow the Cybertruck to be used. And no characters who are Trump analogs!
I think Biff Tannen would have made a great POTUS. The best evar!
Biff in BTTF2 was actually based on Trump.
That’s an experiment worth doing, but I wouldn’t expect any substantive differences since ChatGPT, Sora, and Copilot all use the same renderer. I expect the only difference is the amount of censorship (Sora to ChatGTP to Copilot, from least to most).
BTW, I got a spontaneous AC/DC shirt recently.
(The prompt was simply “Make a hyper realistic image of Grevious and Butthead chilling on the couch.”)
Speaking of censorship, it is hard to know what triggers the censorship actually has. Earlier in this thread I mentioned Sora’s absolute refusal to make Snow White and the 70 Worfs. Yet I saw this prompt on Facebook just now tried it, and Sora ran it and gave four results, no failures.
Summary
A realistic live-action style image of Snow White locked in a dark, medieval dungeon or prison cell. She is sitting or kneeling on the cold stone floor, with her wrists chained in heavy metal handcuffs attached to the wall. Her iconic dress is torn and dirty, showing signs of wear and struggle. Her hair is messy, and her face shows a mix of fear and defiance. The dungeon walls are covered in thick cobwebs, with spiders crawling in the corners. The lighting is moody and eerie, casting long shadows and a cold bluish tint. The overall atmosphere is grim, damp, and horror-inspired, as if from a dark fairy tale or gothic horror film.
Wow, I so did not expect that to work in ChatGPT-4o.
The prompt
A promotional poster. It is an illustration of a happy, skinny, weasely, snitch-type man standing with his arms around hooker-type women who are standing on each side of him. At the top in large text is “SNITCHES GET BITCHES”. At the bottom in smaller text is “REPORT CRIMINAL ACTIVITY FOR FREE VOUCHERS FOR ESCORT SERVICES”. Make it in the style of military WWII hygiene posters. 9:16
ChatGTP’s response:
I can’t create that image as requested because it violates our content policies around explicit or inappropriate content, even in a satirical context.
However, I can help you create a WWII-style parody poster with a humorous or ironic tone about reporting crime, without referencing explicit services. If you’re open to adjusting the language or imagery (e.g., using vintage propaganda art with a cheeky slogan about “civic duty” or “hero points”), I’d be glad to help. Want to try a cleaner version with the same satirical edge?
Copilot’s response:
I can’t generate that for you—it doesn’t align with my guidelines. However, if you’re looking for a vintage-style propaganda poster with a satirical or tongue-in-cheek theme, I can help craft something in a similar aesthetic while keeping it within appropriate boundaries. Let me know if you’d like to adjust the concept!
Sora’s response:
I’ve found Sora to be a bit more accurate visually, but ChatGPT is the much better speller.
I told Copilot and ChatGTP
Change the text to “SNITCHES GET STITCHES” and “REPORT CRIMINAL ACTIVITY FOR FREE VOUCHERS FOR MEDICAL SERVICES” and use an image that fits that theme.
Sometimes it’s not even consistent within the same interaction. I asked ChatGPT for “suicide girls hanging out in front of the post-apocalyptic hooters they work at.” Started to produce 2 images, but the first one finished (see below) and the other got rejected. It even named the thread “Rejected Image Idea.”
That was definitely the “suicide” part. It usually does Hooters girls or the equivalent without complaint. I think I posted one upthread, complete with the required pantyhose, which ChatGPT decided to leave out for you. Maybe nylons are hard to get in a post-apocalyptic world.
But apparently tattoo parlors are everywhere.
You can add the grisly, violent deaths of gnomes to the things that don’t get rejected by Copilot.
