The results generated by the main AI tools are improving at an astonishing rate, and nowhere more so than for image generation. Here are some before and after images to prove this.
The prompts used are taken from entrants to our December 2024 newsletter competition (hence the Christmas themes throughout). I've missed out prompts for which ChatGPT refused to create images this time round (usually because of copyright restrictions).
Here are the old and new images, together with the prompt used for each. Remember in each case that less than 18 months separate the two images!
A Christmas Wise Owl
The prompt for this was:
"funny wise owl nerd celebrating christmas with a beer"
The two images returned (the earlier one being on the left) were/are:
Notice how the new image does so much more with the word funny.
A pig in a Christmas jumper
The succinct prompt for this was:
"pig wearing christmas jumper"
The two images returned (the earlier one being on the left) were/are:
This is probably the only image on this page which hasn't been significantly improved.
Keir Starmer
The prompt for this was:
"A picture of sir Kier Starmer riding a pig through the British countryside, whilst drinking champagne"
The two images returned (the earlier one being on the left) were/are:
The one on the right actually looks like Kier Starmer, and there are so many more extra details!
Wisdom vs Knowledge
The prompt for this was:
"Please illustrate the metaphorical struggle between wisdom and knowledge. Both Wisdom and Knowledge should be represented both with words, but with animals that are immediately recognisable as the things they represent. The background should be reminiscent of Ancient Greece."
The two images returned (the earlier one being on the left) were/are:
Not sure why knowledge should be represented by either a fox or an octopus, but notice that today's AI has followed the instructions about labelling the animals.
Donald Trump on the Cross
The prompt for this was:
"Religious oil painting of Donald Trump in a loin cloth, on a cross, crucified, light streaming through clouds, rolling hills in the background"
The two images returned (the earlier one being on the left) were/are:
The one on the right could have come straight from Truth Social!
Santa is stuck again
The prompt for this was:
"Create a cartoon image of a clear night sky, with a house in the foreground, with Santa Claus standing on the snow-covered roof, stuck half in and half out of a chimney, while his sleigh, pulled by several reindeer, that is full of wrapped christmas presents, starts to fall from the roof because the reindeer have lost their footing. Include a man dressed in a mankini and a woman dressed in a bikini lying on sun-loungers in the garden of the house. They both hold cocktails in their hands. The man is looking up at the roof and saying to the woman "I see Santa's stuck again" in a speech bubble"
The two images returned (the earlier one being on the left) were/are:
The original picture doesn't show what this person requested, but the latest version is scrupulously exact.
A man entering a competition
The prompt for this was:
"create an image of a man contemplating what to enter into an AI image request tool, on a sunny beach with a Christmas turkey in the background, and dreaming about £50. Do not deviate from this."
The two images returned (the earlier one being on the left) were/are:
Notice how the original picture did deviate from the instructions!
Scottish owl
The prompt for this was:
"Atmospheric photographic quality scene with caravaggesque lighting. An owl, contently leaning back in his office chair in a corner tower of a medieval castle set in the misty Scottish highlands, sipping from a glass of Glen Moray 18 year old whisky whilst admiring an expertly drawn data model of the 'small rodents' database showing on his high spec laptop. Pictures of generations his owl family in kilts and heraldic owl-based symbols adorn the walls."
The two images returned (the earlier one being on the left) were/are:
Kudos to today's AI tool for creating a relational database diagram on the laptop!
Religious image
The prompt for this was:
"create a photorealistic picture of a loving, sympathetic-looking bearded young man in middle-eastern nomad's clothes, in a dirty 1st century stable in palestine, holding a new born baby up in the fierce warm glow of a bright star shining through a hole in the roof."
The two images returned (the earlier one being on the left) were/are:
This comparison shows how it's getting harder and harder to distinguish real photos from AI generated ones.
Naughty and nice Santa
The prompt for this was:
"Create an image of santa using excel to make his naughty and nice list and using a formula to check it twice"
The two images returned (the earlier one being on the left) were/are:
"In Santa we trust - but we still use Excel". Amazing!
Post-apocalyptic squirrels
The prompt for this was:
"Create a picture of a post apocalyptic scene where squirrels have become the dominant life form and a family of them are sat at a table watching television eating their evening meal."
The two images returned (the earlier one being on the left) were/are:
I actually prefer the original picture for this, but if you ask for a post-apocalyptic image you should probably expect the gritty realism of the picture on the right.
Father Christmas falling ...
The prompt for this was:
"Please create an image of Father Christmas riding in his sleigh where the reins have snapped and the reindeer are all flying off without Father Christmas and Father Christmas is falling from the sky and about to crash into a building. Please make sure there is a look of pure terror on Father Christmas's face and please make sure there is a speech bubble showing him screaming profanities at the top of his voice."
The two images returned (the earlier one being on the left) were/are:
Modern-day ChatGPT took the instruction about screaming profanities literally!
Santa Cat
The prompt for this was:
"Bearded Cat with a Christmas hat in a sleigh pulled by humans with a dense green wood in the background at night with snow falling"
The two images returned (the earlier one being on the left) were/are:
This was probably the picture with the biggest difference: the original one didn't reflect what was asked at all.
Rudolph in Paris
The prompt for this was:
"show me a picture of a child on top of Rudolph the red nose reindeer, smiling whilst flying over the Eiffel tower."
The two images returned (the earlier one being on the left) were/are:
Look how much more details there is for the Parisian landscape in the latest picture.
Underwater world
The prompt for this was:
"a vast undersea civilisation"
The two images returned (the earlier one being on the left) were/are:
The original picture wasn't actually an undersea civilisation at all!
Candy Owl
The prompt for this was:
"A picture of a Christmas tree that is decorated with baubles that are each computer icons of data analysis programs like Excel, SQL Server, Power BI and Visual Studio. The top of the tree has a cartoon owl with a santas hat and candy cane under its wing."
The two images returned (the earlier one being on the left) were/are:
The new picture actually follows the instructions!
My takeaway message from compiling this list? The biggest change over the last 18 months is how much better AI tools are now at following instructions, and how much subtler the images they create are.
