And A Merry Christmas To One And All!
Notes on AI Used In This Christmas Message
Oh, yes, there was some...
First, I took a photo of Beryl, flat out and fast asleep after Christmas Lunch. I cropped and edited it in Paint.net, after which it looked like this:
Then I uploaded it to NightCafe, where I used it as a starting image for Flux Schnell. I gave it the prompt "A sleeping black and white dog wearing a santa hat, surrounded by piles of opened presents under a tastefully decorated Christmas Tree", set the run time to "Long" and asked for four images. The only usable one was this:
I used several models, trying for something better but they were uniformly terrible. Honestly, it was like the last three years of AI development never happened. I haven't tried generating from an image for a while. I was expecting a lot more. I even used some of my freebies to access the Pro models but they were even worse than the free ones, which ties in with Suno's supposedly best model actually being much worse than its predecessor.
I think some of these companies are just trying too hard now. Either they're getting desperate or they're so far into the forest they can't see the trees any more.
Anyway, I did manage to get one I was fairly happy with so I uploaded it to ClipChamp. Then I went to Suno and generated some "music" to go with it. That was a low-effort move on my part for which I'd apologize if I didn't also think it fits pretty well with the cheesy effect I was after.
All I did to get started was type in a half-assed prompt: "Vaporwave Christmas melody, ethereal, glitchy, otherworldy, tuneful, soothing, seasonal, bells, jingling, faint caroling". That gave me something all but unlistenable so I tried to fuzz it up by covering it under the prompt "As heard through the faulty speaker of a malfunctioning radio, broadcast by a station that fades in and out, quite faint, hard to hear". No luck. The cover sounded pretty much identical to the original. Apparently Suno can't do effects like that.
MeloBytes can, though, so I exported it as a .wav file and uploaded it over there, where I used a filter to make it sound like an old-time radio broadcast. I downloaded the result and uploaded that to ClipChamp. I put the image and the sound file into a timeline, added a couple of effects (Vaporwave and VHS) to get the scratchy, glitchy effect, then I exported the final version to my PC.
I was going to get Blogger to just embed the video directly but the file was too big so I uploaded it to my YouTube channel instead. Once it had processed, I embedded it from there and that was that.
It took me longer to write up this account of what I did than it took me to do it. I probably should have got an AI to do the summary.
How much of all of that was really AI, though? If I'd been a bit less lazy and made up my own tune and uploaded that, instead of letting Suno do it for me, most of what the "AI" did would have been nothing more than post-processing. Calling it AI just makes it sound more impressive, or so the AI companies would like to believe.
It passed a Christmas afternoon while Mrs Bhagpuss and Bery sleep off their lunch, anyway! Happy Christmas everyone!
Posts
Beryl pics are the best!
ReplyDeleteReal Beryl is a million times cuter than AI Beryl!
DeleteI of course had to try:
ReplyDeletehttps://dragonchasers.com/wp-content/uploads/2025/12/CyV4OMbMwN6y3DzHfkxm-0-3w5jf.webp
Beryl has a santa hat. I guess I should have specified it should be on her head and size appropriate!
It's also weird how the presents (which are NOT torn open) appear to be below the rug she is laying on.
I used a Flux 2 Dev freebie to generate that one (on Night Cafe)
Prompt: "Give the dog in this picture a santa cap (red cap,white puffball) and add some torn open presents around the periphery."
That's very... umm... interesting...
DeleteBeryl herself is pretty much lifted straight from the photo. It looks as though the AI barely changed anything at all. It's particularly impressive that it kept all the curls and markings virtually identical, whereas all the images I got, from all the models i tried, smoothed everything out and made it look plasticky.
The problem with that, though, is that if I'd just wanted a photo of beryl with a Santa hat, I could have photo-shopped one. No-one's needed an AI for that kind of thing in a couple of decades. Or I could have got Beryl to wear one and taken a picture of that!
Everything else about the image you got, though, is just wrong. The presents are far too uniform and most of them aren't "torn open". They are "around the periphery" but only in the most literal sense. The depth perception only makes sense if Beryl is lying on some kind of table. The bizarre placement of the hat is the piece de resistance, though. Magnificently misunderstanding the concept and the context!
I don't think the problem is so much the prompting, either. It's that not all the models on NightCafe are all that good. They need a lot of close management to get precise results. Here's a link to a video I watched earlier today that shows just how much an AI can get right with minimal prompting. Well, Google's NanoBanana can, anyway. ChatGPT, not so much...
That video is so meta. An AI avatar talking about using AI to create something but "it" isn't allowed to influence... it's cousin, presumably?
DeleteI watch one channel, and the host is pretty 'out there', but the host got really sick (long covid or something) and couldn't sit in front of a camera to do videos, so she made an AI Clone of herself. It announces itself as a clone, so she's not trying to pull a fast one or anything. But what I find is the clone is so ordinary looking, aside from the off-brand Star Trek uniform: https://www.youtube.com/@JuliaMcCoy
Again, not condoning her content which is pretty fringe, to put it mildly. But the avatar and her (most of the shorts are the real her) are pretty darned close. The avatar does certain gestures over and over and that's all that really gives it away, to me.