This video inspired me to really push Opus 5.5 and test its limits. I followed the general workflow Donald described here and the results are great

Basically the same prompt he used with some minor modifications and pointing Opus to my pc-98 image gallery

there's a lot that needs to be fixed here, imo. But because it's a mixed flow and not pure code, it gives a lot more leeway in terms of expressivity. This video itself has a few bugs that can be ironed out, but I wanted to make this a true one pass experiment.

This was so peak when it was happening. Bing was fucking wild.

fucking chills oh my fucking god i don't believe what i'm seeing

Incredible

there's no way there's no way did the sprites come from your library or something? how were they constructed (eg. the shoggoth)? I'm in complete disbelief

for this workflow, both here and in Donald's original, the model is doing three things. First, it's generating assets with models and essentially "tracing" over generated references. It's using them to understand motion and sometimes it's just a generative output. Then it's compositing these into a scene and finally, it's overlaying the typography work and beat matching.

incredible. if you know, what image model generated the shoggoth? it's quite impressive

that’s pretty cool! i also gave it a try with a different style direction and the result (after a few rounds of tweaking) is impressive

ggwp

very fun thank you for sharing

If you want to see a "pure code" example you can look here

wtf

Hate to tell you this, but you missed the CDR reference.

blame that on Claude

I’d say this lacks the unsettlingly accelerated pace of the original, or the “what does Claude love” one (with the slightly-too-fast-to-read frames). The original QT’d x2 had a quality akin to tail-to-head overlapped speech that some influencers use, feels like perception hacking

The original is more impressive to me because it's pure code animation, and the step change from this line comes from a mixed workflow. it uses a sota video model for its more complex scenes. The point is to experiment with reproducibility, the novelty curve is already flat...

holy fuck this is amazing

so not claude

Donald’s workflow is an orchestration of SD2.5 outputs for motion reference and in some cases just straight up a filter pass on top of it. The pure code animations have a notably different look, which in here you can see in the typography work

dude you should have made this a nightcore remix this song goes hard af pitched up

I can't get over how good these videos are.

This is bonkers

this is so good. wtf

you did it. you created SOVL with ai. i can’t tell what is straight up generative and what is made with regular code. this is nuts

Much Better! I feel that with styles that are more stylized and limited it is much easier to maintain consistency. Rotoscoping probably would only work well with anime styles since it would be much easier to match distinct shadows (though you could have "animefication" step).

I think the rotoscoping point is moot if you’re just looking to stress test pure model spatial reasoning. I think the test here is more about it’s composition skills, beat matching the cuts and etc

Niiice

wtf did Anthropic unleash on us with 5.5? What a wild jump

Wtf (good)

It's so good. Did u upload it also on tiktok?

Holy peak

This is it. The Magnum Opus.

1) holy fuck holy fuck holy fuck. it’s like a raccoon got his hands on a flamethrower I am so happy you are using opus 5.5 2) it is really cute that this includes a bunch of the same bits (the box, ending with the performers taking a bow)

this is really well done, the animations are great - how did you get it to make this style?

@p1atdev_art This is so cool! Completely dopamine’d

banger

Is the main character of the video created by someone else, or created for this video?

oh my god

Ok, wow

Now we just need to remix it for OPNA

Hope fuel, the future can be so beautiful.