HyperFrames at 80k Stars: Three Videos Finished Overnight With a Free Model
An engineering-perspective hands-on test of an open-source video framework: HyperFrames (open-sourced by HeyGen, over 80k stars on GitHub) lets AI write a webpage and shoot it frame by frame into a video; animations and transitions are implemented the way you do webpages, without touching a timeline. The official positioning targets coding tools like Claude Code and Cursor. This article instead used WorkBuddy with the free-tier Hy4 flash, and ran three clips through in one night. Several engineering details are worth noting: first, it returns a storyboard before starting, writing out what each second shows, and proactively asks whether the "8 platforms" should be written as real names or fabricated; second, the self-review step is tough - it calls out white text on a light background appearing blurry, animations competing for the same line of text, text below 28 pixels becoming fuzzy after compression, and a scan-light effect with two tracks switching simultaneously causing frame flicker - fixing each one and re-checking. Third, encountering a serious issue like 66 elements with blurred backgrounds that would drag down the output, it stops before finishing the render, removes them, and starts over. Fourth, data animation is honestly labeled TREND instead of pretending to be precise. All three finished clips, complete prompts, time costs, and downsides are recorded as they happened.
I’m Xiaohu, with a bachelor’s and master’s from Zhejiang University’s computer science program, a “Breakthrough Actionist,” and author of “AI Super Individual – Complete Guide to a One-Person Company.” I focus on AI training, super-individual creation, and company efficiency coaching.
These days the HyperFrames open-source project on GitHub has jumped past 80,000 stars, gaining over 20,000 in a month. Behind it is HeyGen, which works on AI video.
Hi everyone, I’m Xiaohu.
This is a real prodigy for making videos. If you don’t want to pay for video generation models, HyperFrames is an especially good choice.
Famous as it is, when you actually want to make a video with it, the barrier hasn’t dropped at all - you need to know how to code, and you need to understand animation.
Let me explain the principle of HyperFrames code-to-video: AI writes a webpage, and it captures that webpage frame by frame, joining them into a film.
Animations and transitions are all handled the way you’d write a webpage; you never touch timeline software.
The official positioning is for AI coding tools. Claude Code, Cursor, and others can command it. This time I didn’t use Claude; a free domestic model ran the whole journey from start to finish.
I pasted prompts into WorkBuddy, let Hy4 flash write, revise, and render for me, and in one night produced three clips: an MG-animation B-Roll, a data animation of GitHub star counts, and dynamic subtitles for a real talking-head video.
Let’s look at the three finished pieces:
▲ Case 1 GIF: 15-second finished clip
▲ Case 2 GIF: number rolling from 0 to 86,333
▲ Case 3 GIF: key words highlighted in gold + enlarged
Case 1: A 15-second B-Roll, it returns a storyboard first
Open WorkBuddy, create a new task, switch the model to the free Hy4 flash, and paste the first prompt:
Please use hyperframes to create an MG animation, 3:4, 15 seconds long, as the B-Roll for a talking-head video. You help me conceive the complete script and transitions. Requirements: smooth and fluid, high aesthetic. Show the following content: one person is a company - morning AI writes a WeChat article, afternoon AI makes a talking-head short video, evening auto-posts to 8 platforms.
It didn’t rush to start; it first probed the machine, returning a storyboard table: what appears on screen at each second and what happens, written clearly.
After returning the table, it paused to ask me: should the “8 platforms” be written as real platform names or as virtual ones? I chose real platform names.
After writing the webpage file, it first ran a self-check and found a page of issues: the “afternoon” capsule had white text on a light background that looked blurry, two English words mixed into the body text, the first two animations fighting over the same line of text, and a blank first frame - it fixed each, then re-checked, and all turned green.
Then it started capturing. In about ten minutes, a 3:4, 15-second clip landed, just over 5MB. Overlaying it on the talking-head footage merged it seamlessly.
Case 2: GitHub star data animation, numbers are checked live from the official source
Second prompt:
Please look up the current real star count of the heygen-com/hyperframes repository on GitHub, use hyperframes to create a 15-second 16:9 data animation: the number rolls from 0 up to the latest star count, with a bar chart growing. Requirements: smooth and fluid, high aesthetic. Export the final video and tell me the time cost.
It added a clever touch in the design: the number rolls with strong deceleration, landing steadily without bouncing back - numbers aren’t balls, they don’t bounce. Seven bars grow from the bottom at staggered timings, and the last bar changes color, hitting the same beat as the number landing.
It also volunteered an honest note: the months on the bar chart are illustrative, not real data, so it only labeled the chart as TREND, to avoid pretending it’s precise data.
After writing it, the self-review again found two hidden bugs: the scan-light animation had two tracks switching states at the same time, which would cause a one-frame hard cut, so it split them into separate “position” and “opacity” tracks, staggered. And fifteen pieces of text were below 28 pixels, which would blur after video compression, so they were all raised above 28.
The most striking part came after the final render, when it looked again: a line of 220-pixel-tall “HYPERFRAMES” ghost text in the background was stealing attention and pressing on the numbers, so it lowered the opacity to only 3%, letting it recede as texture.
Final delivery: 15 seconds, 3.1MB, all checks green, 0 errors, contrast 80/80.
Case 3: Adding dynamic subtitles to a 3-minute+ talking-head video
The third one used real material - a talking-head video I recorded myself, 203 seconds:
Copy my real talking-head video D:/development/1007ArticleOnly/_hf2_1118/talking-raw.mp4 to your working directory, use hyperframes to add beautiful dynamic subtitles to it, with key words in each sentence enlarged and colored for emphasis. Export the final video and tell me the time cost.
▲ It first verified the material and set the subtitle technical route.
It first checked the source: 544x960 portrait, 25 fps, 3 minutes 24 seconds - it noted the portrait layout. The subtitle approach was decided as “keep the video still, just overlay a layer of subtitles,” with key words highlighted in gold and enlarged.
Before the official capture, another checkpoint: the check reported 66 elements with blurred backgrounds - a serious flaw that would drag down the final output in HyperFrames testing.
It didn’t wait for the video to finish; it directly stopped, removed the blurs, and started over.
Official capture: portrait 1080x1920, 5035 frames, finished in 1 minute 58 seconds, all 8 self-checks passed.
I sampled three frames: the subtitles run along the bottom edge without blocking the face; key phrases like “AI短剧算力,” “100万元,” “22万部” are highlighted in gold and enlarged, easy to catch at a glance.
For the same 12-second segment, first the original, then the subtitled version:
▲ Original talking-head: only visuals and sound
▲ With dynamic subtitles: key words gold-highlighted and enlarged
After three clips, a comparison table
Dimension | Traditional editing (Jianying) | React video framework (Remotion) | HyperFrames
Driving method | Manual timeline dragging | Writing React components | Plain ordinary webpage
Model requirement | AI can’t change it | Complex components easy to make mistakes | Free-tier model completed all three clips
Rendering | Depends on clip length, manual editing | Depends on component complexity | 15-second animation took about ten minutes; 203-second talking-head rendered in 1 minute 58 seconds
Automation integration | Hard | Possible | One-line command directly produces the video
Downsides also stated honestly
Default silent. All three clips are pure picture tracks; if you want sound, you have to specify it in the prompt. In Case 2, it even asked me whether I wanted sound effects.
Long videos scale with time. A 203-second render took nearly 2 minutes; a 10-minute video would have to wait around 6 minutes.
Final remarks
After one night and three clips, my only jobs were: paste prompts, click a few pop-ups, and sample frames to check the result.
Writing the webpage, finding flaws, stopping and re-rendering - all done by the AI. The framework is open-source and free; the model was the free tier.
My assessment: for short videos with clear structures like product promos, data animations, and talking-head subtitles, it’s worth installing and trying. For shooting narrative or editing long videos, it still can’t replace editing software.
Which one do you most want it to do for you? Talk in the comments.
The above is the complete hands-on record, and the three prompts can be reused directly. By distilling the workflow of “storyboard first → self-review and fix each item → stop render and restart on serious flaws → sample frames for final verification,” you can quickly understand the temperament of any code-based video framework.