AI Trends

Author uses Fusion Max 2.0 to generate a 52-second cat animation entirely with code for $3.76

Ayi AI NotesUpdated Nov 13 13:24

AI summary

The author used Fusion Max 2.0 from Singapore's FUMO Lab over 4 calls, spending $3.76 to generate a single 105KB HTML file, producing a 52-second hand-drawn watercolor-style cat promotional video purely with Canvas 2D code, without using any video generation model or external images or fonts; the music and sound effects were synthesized with Web Audio, and 1,560 frames of 1080P footage were exported frame by frame. The 4 calls totaled 113,000 input tokens and 171,000 output tokens; priced at Fusion Max 2.0's $20 per million output tokens, the author says the same usage would cost about $9.68 at the $50 unit price of Claude Fable 5.1 and GPT-6 Astra. The author also offers six practical lessons, including locking down the single HTML file and Canvas 2D boundaries in the first version, writing the storyboard table first, building in a time-jump inspection interface, attacking only the biggest weakness each round, writing revision notes with only the second marks and what is seen and desired, and finally adding offline audio export and frame-by-frame recording interfaces.

Why read this

AI score 62

The author breaks down the bill for using a reasoning model to write Canvas 2D code for animation and six iteration lessons, which can serve as a reference path for pure-code video generation.

Topics

Tags

Quoted post

AYi

@AYi_AInotes

https://t.co/ec77qpRS9L

View original

https://t.co/ec77qpRS9L

View original post

Body · AI translation

Holy crap, everyone stand up!! I'm declaring that there's really no need to keep banging your head against expensive video models like Claude Opus 5.5 and Fable 5.2 for AI animation videos. I just used Fusion Max 2.0, this fusion model from Singapore's FUMO Lab, and it only cost 3.76 dollars, 4 calls, and generated a single 105KB HTML file, and then pure code literally wrote out a 52-second hand-drawn watercolor-style cat promo video, with results directly on par with Claude Fable 5.2. Absolutely awesome!

This video didn't use any video generation model, didn't reference even a single external image or font, and even the piano background music and the crisp sound effect of a cup shattering were all synthesized live mathematically with Web Audio in the code, exporting frame by frame a full 1560 1080P HD frames,

Previously I used it to spend 5.77 dollars and write a playable game. This time, after finishing the game, it even covered the promo video too,

When many people mention AI making videos, they only stare at pixel diffusion models, fighting with random gacha in prompts every day. Change the shot and the cat's stripes change, and the white paws are lost too,

Actually, looking at it from another angle, pixel models are guessing the picture frame by frame, while having a top-tier reasoning model write Canvas 2D code is using mathematical formulas to build a deterministically running 2D animation engine,

I broke down the bill and measured results of these 4 rounds of iteration, and there are three particularly counterintuitive findings,

First the bill. Making video with pure code is a typical output hog. Every round of changes requires the model to rewrite the entire file from beginning to end, These 4 calls totaled 113k input tokens and 171k output. Fusion Max 2.0's output price is 20 dollars per million, while Claude Fable 5.1 and GPT-6 Astra both charge 50 dollars, Calculated at their public prices for the same usage, it would come to about 9.68 dollars. 3.76 dollars directly cuts the cost to 1/2.6, saving more than half the budget, Next, its understanding of visual details. Many people think coding models don't understand art, but actually as long as you explain the problem thoroughly, it's more obedient than an artist. The orange cat in the second version came out like a flat paper sticker,

An elliptical body with four legs stuck on, plus a rigid cat face always facing straight forward. In the prompt I named all 12 action poses one by one. By the third version, the cat head had directly evolved three perspective drawing methods: front, side, and three-quarter side, and the four legs also had realistic joint bending. The stripes, white muzzle, and white paws matched perfectly across the four scenes,

What really made my scalp tingle was the final round of engineering delivery. No matter how good the web animation looks, it has to become a video file to be published. In the prompt I directly raised three hardcore interface requirements, First, the same second must always draw exactly the same frame,

Second, the entire Web Audio audio track can be rendered offline and exported as WAV,

Third, directly build in a frame-by-frame recording mode. As a result, these three low-level interfaces all passed testing on the first run. This kind of completeness of getting the code right in one go is fully presentable even at the Fable 5.2 level,

To be honest, the process wasn't that miraculous. It definitely wasn't a one-shot finished film. The first version came out with a grayish picture, looking like cheap flat vector illustration. It took three full rounds of polishing before and after, and the final cut actually still has a small blooper: in the bookshelf collapse segment, the cat flies across the top of the frame twice before landing,

In addition, the actual gameplay footage in the later part of the video is game screen recording. The final stitching and subtitles I composited with FFmpeg, and this part is not included in the 3.76 dollars, If you want to hand-build an animated promo video with pure code yourself, these six practical lessons can be copied directly, First, the first-version prompt must lock down the boundaries. Explicitly require that only a single HTML file be delivered, with all visuals drawn by Canvas 2D, and no external images, audio, or third-party libraries allowed,

Second, before writing, first write a storyboard table, clearly writing out each shot's start and end seconds, visual composition, and the plot task this shot needs to convey,

Third, in the first version you must have the model write a time-jump inspection interface, so you can jump to any second and take a screenshot to find bloopers at any time. Otherwise, watching 52 seconds from the beginning every time will drive you crazy,

Fourth, each round of iteration should only tackle the current biggest weakness. First tune the watercolor texture and lighting, then grind on character perspective and joints,

Only at the end go fine-tune individual blooper shots. Fifth, when giving the model revision suggestions, always write only three things: which second, what is seen in the frame, and what effect you want,

Never try to be clever and specify exactly how the code should be changed. Sixth, after all visuals and sound effects are completely finalized, in the last round have it add the offline audio export and frame-by-frame recording interfaces,

To put it plainly, making video with pure code is treating the large model as an extremely fast graphics programmer. The model is responsible for drawing pixels stroke by stroke and calculating waveforms, while the human is responsible for being a good director and clearly saying what's wrong.

Source

Ayi AI Notes