News:

This is the TEST SITE - feel free to post but content will be deleted

Main Menu

AI Image Creation

Started by Slim, December 15, 2023, 11:50:34 AM

Previous topic - Next topic

Thenop

I found an old wedding photo from my parents. Late 60s, large format (a4) and I asked Gemini to colour it, true to the style and age. I was stunned.. truly magnificent. I could have asked to take out some background elements but decided it should stay as true to the original as possible.

PXL-20250917-073255863-MP.jpgHighslide Image Viewer

1758097681506.pngHighslide Image Viewer

Slim

That's brilliant, could almost be Kodachrome. Captures the colour film personality perfectly.

The Picnic Wasp

I've been tempted recently by an ad on Facebook to provide a short, coloured animated version of old still photos. Sorry the website name escapes me but it looks very impressive.

Nickslikk2112

If I got AI to colour my parent's wedding photos - from the 50s - they'd still come out grey. It's grim up North...

Slim

This is cropped from a school photo from September 1971.

jg11.pngHighslide Image Viewer

I asked Gemini to colourise it and resize it. It's a bit uncanny. I specified the colours and it got them all right, except that my eyes were too light brown. So I got it to redo it. Then I added a little bit of noise in the GIMP because it looked unnaturally smooth. A bit of film grain. I'm just gobsmacked by this.


schoolcrop.webpHighslide Image Viewer

Thenop

It really is unbelievable what this can do. I recall watching old episodes of say, CSI, in hich a tech blows up a video still and - low and behold - they discovered an object until then unnoticed in razor sharp detail. Impossible 25 years ago, but now.. here we are.

Thenop

1766672798266.pngHighslide Image Viewer

Slim

This is a photo I took at an airport in Spain in 1973:

airportbefor.avifHighslide Image Viewer

I got ChatGPT to remove the reflections on the glass, the specks of dust and the obstruction:

airportai.avifHighslide Image Viewer

Came out pretty well I think.

Slim

I sometimes knock up a few cycling stats charts at the end of the year in Google Sheets. This time I thought it might be quicker to get an AI to do it. I took a screengrab of the appropriate portion of my spreadsheet and asked ChatGPT for a doughnut chart.

distances.avifHighslide Image Viewer


OK - exactly what I asked for, but a bit crude. I asked it to make it a bit more stylish.

chatgptdo.avifHighslide Image Viewer

I was happy enough with this until I realised that it had hallucinated an extra 7.8% entry. It had another couple of goes but wasn't able to fix it. It explained:

chatgptexpl.pngHighslide Image Viewer

I took the first image to Gemini, to see if it could do any better. Just asked it to make it more stylish, prettier. No hallucination this time.

gem-distances.avifHighslide Image Viewer

So - in the end it wasn't actually any quicker than doing it myself in Google Sheets, though it might have been if I'd gone straight to Gemini - and I had to check the result more carefully. But the technology is improving fast.

Thenop

Did you specify which bike should have which colour?

Slim

Quote from: Thenop on December 31, 2025, 11:48:00 AMDid you specify which bike should have which colour?

No. I did think of that actually - green for the X, Red for the first Boardman and so on - but decided not to bother.

Thenop

Quote from: Slim on December 31, 2025, 12:15:01 PM
Quote from: Thenop on December 31, 2025, 11:48:00 AMDid you specify which bike should have which colour?

No. I did think of that actually - green for the X, Red for the first Boardman and so on - but decided not to bother.

So ChatGPT and Gemini 'decided' to use the exact same colour per bike. Curious, unless the engine is the same.

Slim

Quote from: Thenop on December 31, 2025, 01:29:40 PM
Quote from: Slim on December 31, 2025, 12:15:01 PM
Quote from: Thenop on December 31, 2025, 11:48:00 AMDid you specify which bike should have which colour?

No. I did think of that actually - green for the X, Red for the first Boardman and so on - but decided not to bother.

So ChatGPT and Gemini 'decided' to use the exact same colour per bike. Curious, unless the engine is the same.

That's because they were both working from the first ChatGPT image which already had the colours defined.

Slim

This pic of Neil at Advision in London came up on Instagram earlier.

neil1.avifHighslide Image Viewer

It's pretty rough - some sort of print pattern and yellowed, perhaps due to the film being out of date or overheated. I asked ChatGPT to fix both issues. One of those cases where the AI is a bit more real than the original image. Judging by the datestamp you can just see on the original, taken three weeks after I first saw them in Newcastle.

neil2.avifHighslide Image Viewer

Slim

Quote from: R6GYY on January 10, 2024, 01:04:11 PMI've had a few goes now just to experiment, but any text I use gets corrupted in the image, translated with extra or different letters. What gives?


Two years after this thread started things seem to be much improved on this point. When I was creating promos for Ravenstone Radio I was getting text like "RAVSSTONERRADI" but now generally the various image AIs get text right. I thought about this when I got an AI to generate this pic:

badtex.pngHighslide Image Viewer