跪拜 Guibai
← Back to the summary

HyperFrames Turns Web Pages Into Deterministic MP4s: A Chinese Flash Model Runs Three 1080p Video Jobs

Ditch CapCut: Use an AI Agent + Web Code to Build a Fully Automated Video Generation Workflow

Hi, I'm Meng Jian.

Anyone building products inevitably has to make videos: product launch films, data animations showing growth, talking-head clips with subtitles. Dragging the timeline frame by frame in CapCut (剪映) is too slow; asking AI to write Remotion-style React video code requires installing a pile of dependencies, and the more complex the code gets, the easier it is to break.

Recently HeyGen's open-source HyperFrames has been hot on GitHub: it gained more than 20,000 stars this month, and as of November 17 the repository sits at 85,778 stars. Its idea is simple: have AI write an ordinary web page (HTML/CSS/JS), then render it frame by frame into an MP4.

This time I didn't use Claude. Inside WorkBuddy I used the domestic model DeepSeek-V4.1-Flash to build one case for each of the three typical official use cases: a product launch film, a data animation, and a talking-head clip with subtitles.

First, the toughest verification result: on the left is the frame frozen by the headless browser at frame 150, on the right is frame 150 of the final exported MP4. The two images match. Whether web-rendered video loses frames is answered by this picture.

Frame 150 comparison: left is the frozen web preview, right is the exported result

Below are the complete hands-on steps and pitfalls.


Step 1: Configure the environment and install

In WorkBuddy, make sure DeepSeek-V4.1-Flash is selected. If you don't have WorkBuddy installed yet, download the desktop client from the official site:

WorkBuddy official download page

Enter the first prompt in the input box to install HyperFrames:

Please first check whether Node.js, npm, and the Chrome headless browser rendering environment are installed on my computer (if not, set them up), then install HeyGen's open-source HyperFrames in the folder C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo. After installation, tell me the version and time taken.

While cleaning up old dependency remnants, WorkBuddy popped up a security interception notice: "Detected bulk delete operation: cumulative count 5475 items, high-risk command threshold 50".

Bulk delete security interception popup, ready to click 'Allow this deletion'

This is cleaning old dependency caches. First check that what's being deleted in the popup is dependency directories like node_modules, then click "Allow this deletion" to proceed; if it's other files, click reject. The model then completed its initialization self-check, and the built-in 10-second test clip headless render took only 17.7 seconds.


Step 2: Case 1 - Fetch the Ziliu website to make a 20-second vertical launch film

After the base environment was done, I asked it to read the live website of my product "Ziliu" (ziliu.online), extract design elements, and make a vertical launch film.

Enter the second prompt:

Please read the real interface, colors, and copy of my product "Ziliu" website at https://ziliu.online, and use HyperFrames to write a roughly 20-second 9:16 vertical product launch film with camera transitions, text animations, and a final logo ending. Export to C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\ziliu-launch-v1.mp4 and tell me the time taken.

The AI fetched the website content, extracted the traditional color scheme of "rice paper + ink + lapis blue + vermilion", organized the HTML structure, and completed the render.

Ziliu 20-second launch film first version generated, results player loading on the right

Open the generated ziliu-launch-v1.mp4 for a full-screen check:

First version first screen: rice paper background with Ziliu seal and update info

First version middle shot: showing multi-platform distribution switching interface

The layout was tidy, but against a high bar for launch material, three flaws surfaced: the transitions between shots were abrupt, lacking breathing room; the core selling point "multi-platform content distribution tool" was squeezed into the subtitle position, too small and not eye-catching; the entire video had no audio track, pure silence.

When I switched back to WorkBuddy after the full-screen check, I hit a live incident: the input box next to the original conversation kept spinning, stuck and unable to accept focus.

Pitfall tip: Don't wait for a stuck dialog box. Just click "New Task" in the upper-left of WorkBuddy to start a new conversation. Because all the web code and media resources are in the local working directory, you can point to the project path in the new conversation and ask it to modify.

In the new task, send the third optimization prompt:

Please optimize the C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\ziliu-launch project: the first version's transitions were a bit abrupt, and the core selling point "multi-platform content distribution tool" wasn't prominent enough. Change the transitions to smooth elastic push-and-pull animations, add a highlight card zoom effect to the core selling point, and add background music. Re-export to C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\ziliu-launch-final.mp4 and tell me the time taken.

This round the domestic model did a structured job. To give the selling point breathing room, it proactively extended the duration from 20.0 seconds to 22.9 seconds, explaining in its output: "To give the selling point a true 'one screen' rather than having it squeezed next to the title, the only option is to add a roughly 3-second beat."

To address the silent audio, it wrote a Python script scripts/make-bgm.py that deterministically synthesizes background music and mixes it in, reproducible and free of copyright risk.

Optimization complete, duration set to 22.9s and BGM deterministically synthesized

The final exported launch film now emphasizes the core selling point with a highlight card that elastically pushes and pulls, noticeably improving the rhythm:

Test result ziliu-launch-final.mp4 (22.9 seconds)


Step 3: Case 2 - Make a real GitHub star growth data animation

The second scenario is a common need for indie developers showing project growth on social media: use HyperFrames' own star growth to make a dynamic chart.

Enter the fourth prompt:

Please look up the current real star count and this month's growth of the heygen-com/hyperframes repository on GitHub, and use HyperFrames to make a 15-second 16:9 data animation showing the number rolling from 0 to 79,000 stars and the bar chart growing with this month's 20,000-star surge. Export to C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\star-chart-v1.mp4 and tell me the time taken.

When it triggered a network snapshot fetch, WorkBuddy showed an external access security confirmation:

Attempting to access web.archive.org triggers security risk interception

Click "Always allow in this session" to proceed. The first version of the chart grew the bars, but the pure black background was a bit monotonous and the peak lacked visual impact. I then sent the fifth prompt for refinement:

The chart background is a bit monotonous, and the number peak lacks visual impact. Add dark tech-style grid lines to the background, and when the bars reach their highest point, add a golden glow pulse and a dynamic corner badge "This month +20,000 Stars". Re-export to C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\star-chart-final.mp4.

The domestic model was very level-headed in correcting the data. In its output it explicitly pointed out: "The total star count is 85,778 (85.78k), not 79,000; 79,000 was the level in early November."

Star data animation final version generated, with corrected corner badge values and time taken

Test result star-chart-final.mp4 (15 seconds)

The whole optimization round included 3 headless renders, taking 5 minutes 30 seconds in total; a single 15-second video render took only 16 seconds.


Step 4: Case 3 - Extract timeline from a local talking clip and add dynamic subtitles

The third scenario combines real material: add modern short-video-style emphasis subtitles that highlight key words to a 25-second local talking-head video I recorded.

First, look at the original talking-head footage (no subtitles, low volume):

Original talking-head material my-talking-clip.mp4 (25 seconds, no subtitles)

Enter the sixth prompt:

Copy my real talking-head video D:\Dev\hyperframes-demo\my-talking-clip.mp4 into the directory C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo, and use HyperFrames to add nice dynamic subtitles with each sentence's key words enlarged and color-changed for emphasis. Export to C:\Users\your-username\OneDrive\Documents\Ziliu AI test\hyperframes-demo\talking-sub-v1.mp4 and tell me the time taken.

The AI invoked local whisper-cli to run word-level timestamp transcription:

WorkBuddy installing local whisper and running word-level timestamp transcription on the real talking-head clip

While analyzing the waveform, it discovered two physical flaws in the original video: the first 7.36 seconds of the talking-head clip were a preparation phase with pure silence (-180 dB); and the original volume was too low (-41 LUFS).

It applied two targeted treatments: during the first 7 seconds of silence, it automatically laid out a large title card saying "Use your phone to remote-control your computer and let the Agent do the work" and slightly darkened the underlying footage; when the voice starts at the 7.38-second mark, the title card exits, the dynamic emphasis subtitles follow smoothly, and key words simultaneously enlarge and turn golden yellow.

Test result talking-sub-v1.mp4 (25 seconds)


Step 5: Frame-by-frame comparison: the web preview and the final video are nearly identical

Whether web-generated video can be used commercially comes down to "determinism". To check if it really is stable frame by frame, I sent the verification prompt:

I want to prove to readers where HyperFrames beats traditional browser screen recording and Remotion: show the project's plain HTML code structure (no need to install huge React dependencies), and enable the local preview, jump to frame 150 and take a screenshot, then place it next to frame 150 of the just-exported MP4 to see whether it's strictly frame-identical with zero dropped frames.

The AI quickly spun up the local preview service, froze and jumped to frame 150 for a screenshot, and simultaneously used a headless tool to extract frame 150 from the MP4, generating a comparison report report.html - the comparison image at the top of this article. The frame similarity was 0.99, with the small difference coming from video compression.

Text shadows, hair strands, and card edges all line up. Screen recording captures frames based on refresh rate, which is prone to shaking and dropped frames; this approach renders each frame in a fixed, frozen manner, so that problem doesn't exist.


How does it compare to traditional approaches?

After finishing the three cases, looking back at what makes it strong, each point is supported by the tests above:

  1. It writes ordinary web pages: The AI doesn't have to write React components. The domestic Flash model rendered the first versions of all three cases directly, and one or two rounds of edits made them usable.
  2. Every frame is rendered at its own time point: In step 5, the frame 150 comparison shows the preview and final video are nearly identical, with no frame loss like screen recording.
  3. The output is a project folder you can keep editing: In the first case, when the input box got stuck, starting a new conversation and pointing to the project path let me continue editing.
  4. Fast: A 15-second clip takes only 16 seconds to render once.

Here's a table comparing it with the most common approaches:

Dimension Traditional editing (e.g. CapCut) React video framework (e.g. Remotion) HyperFrames
Driving method Manually drag the timeline Complex React components and JSX state Plain ordinary HTML / CSS / JS
Model requirements Can't let AI modify directly Must write React components, many dependencies, easy to error when complex Writes ordinary web pages; this time the domestic Flash model handled it
Rendering determinism Manual editing, WYSIWYG Also frame-by-frame rendering Frame-by-frame frozen rendering; frame 150 was verified identical
Time per render Manual editing, depends on clip length Depends on component complexity A 15-second clip renders in 16 seconds
Full automation workflow Hard to seamlessly integrate with Agent Can integrate, but dependency package is huge Naturally connects with Agent; one CLI line produces a video

My conclusion: HyperFrames turns video-making into web development, and a lightweight domestic model can pull it off.

Weaknesses and limitations:

  1. Environment and dependency threshold: You need a stable local Node.js, npm, and Chrome headless environment; the interception prompts during dependency cleanup can mislead beginners.
  2. Long-form narratives don't fit: If you only give a vague one-line requirement, plain HTML/CSS can get lost in shot logic. It works best for clearly structured product promos, dynamic data displays, and social media packaging.
  3. Audio still needs to be added actively: By default there's only a video track; you must explicitly ask in the prompt to synthesize background music or attach audio timestamps.

Summary and honest judgment

Running the whole set, without using an overseas model like Claude, the domestic DeepSeek-V4.1-Flash inside WorkBuddy completed all three 1080p videos.

One-line judgment: if you regularly need to make short videos with clear structures like product launch films, data animations, or talking-head subtitles, HyperFrames is worth installing and trying; if you're shooting narrative or long-form video, it still can't replace editing software.

If you want to try it, install WorkBuddy, send it the first prompt above to set up the environment, then hand it your product website or a talking clip, and you'll have your first video.


👋 I'm Meng Jian, formerly a Tencent T11, ex-ByteDance tech lead, now doing AI programming full time.

🔥 More AI programming practice:

Find this useful? Like + Save is the biggest support 🙏

This article was compiled by the author with AI assistance; content has been tested and reviewed by the author.