🫡🪴

Corridor

eye

System prompts often hint at some pretty funny bugs that the prompt tries to paper over: "Search results aren't from the human - do not thank the user for results"

A new prompt to add to the fofr-benchmark: > a man counts out loud from 1 to 10, using his fingers and holding them up as he goes > a man counts out loud from 1 to 10, "1, 2, 3, 4, 6, 7, 8, 9, 10", he counts using his fingers and holds them up as he goes

I made some Veo 3 videos, they've been popular on X, you might have seen them already. > a video with dialogue of two muffins while baking in an over, the first muffin says "I can't believe this Veo 3 thing can do dialogue now!", the second muffin says "AAAAH, a talking muffin!"

gm

PixVerse v4 is some sort of perfect AI slop generating machine replicate.com/pixverse/pix...

Sink hole. Experiments with the new “exp” param in MJ v7

If I add a watermark does it make it more believable?

swipe right

And a distressing number of people are okay with slop

I wonder how many tokens my brain processes daily

Imagine what you could do with an open source fine-tuneable gpt-image-2.

Watching o3 guess a photo’s location is surreal, dystopian and wildly entertaining simonwillison.net/2025/Apr/26/...

That blink

You can just do things

Makes my skin crawl

I was working away from home the last couple of days and managed to make my hotel more "shiny" with Pika's new Pikaswaps and Pikaaddition features. Incredible what you can do with a mobile phone and a couple of minutes with Gen AI technology!

Current mood

the glass is half dull

Imagen 3 is so good at anime pixel art styles

Impressed with this Imagen 3 output, mostly in how the dragon is cropped out of the composition

There you are! How did you get up here?

I’m seeing how I can experiment with Imagen 3 and fractal prompts

Trying out Minimax's video-01-director model, I like how this generation finds the subject in the crowd then follows them replicate.com/minimax/vide... (Prompt in alt)

Pearly gates

Google's Imagen 3 now on Replicate 🚀 replicate.com/google/image...

AI video really excels in the horror genre These are all from Kling v1.6: > a very strange black and white silent arthouse movie scene, a strange creature is twitching in a dark room, it's unnerving, scary, old and weird

You can now use a character reference image when using the Minimax "video-01" (Hailuo) model on Replicate. And, dang it's sooo good. replicate.com/minimax/vide...

he did what!?

I'm finding a couple of things with flux pro finetuning: - you need fewer steps to get a likeness or style than dev - the training generalises differently and can overfit more easily Here's two flux ultra outputs from my Mona Lisa training (300 steps, character, quality)

Same. “Oh, did I already have my coffee” 😧

One day I might get Veo access. Sigh.

I don't think people know that making plausible fake videos is now something that doesn't require a lot of skill or time I made this Veo video of a man trying out a flying skateboard at CES from a text prompt, 2nd try. Notice the shadows, green light, his body shifts for balance

I'm enjoying playing with flux ultra again

I've found that the "ceefax" keyword looks great on flux ultra. (CEEFAX was an early text service that broadcast on BBC channels in the empty parts of a TV signal)

Space suite

I've been trying out Hunyuan fine-tunes on this Homer meme. Attempting to massively overfit and then tweak, I'm not getting the results I want, but I did get a nice output that works well as a follow on clip.

I tried to see how Kling v1.6 would handle the trolley problem. But it just backed away slowly.

soul traversal

(Left) Top #aiDoctorMoreau's 2024 creation on Reddit: White Humpback Dachshund (Right) Top 2024 on X: Flatfish Dachshund

together forever