Introduction: Escaping the 'Plastic Feel' of AI Video
Most AI-generated videos online today share a cheap, plastic quality: mismatched AI stock images, robotic narration, and subtitles that drift out of sync. Viewers swipe away within seconds.
The open-source project Agents365-ai/video-podcast-maker attempts to industrialize this cobbled-together pipeline. It combines web research, script writing, multi-engine TTS, React-based video rendering (Remotion), and FFmpeg post-mixing. Its core promise: programmatic, pixel-level control over every animation, transition, and subtitle frame.
Architecture: Remotion and Headless Browser Rendering
Instead of dragging clips on a timeline in Premiere or CapCut, the timeline here is defined by React code.
> Remotion is a framework that lets developers author video with React components, CSS animations, SVG, and Canvas. It does not play in real time; instead a headless browser (Puppeteer) renders the page frame by frame, and FFmpeg compresses the resulting images into a high-fidelity video. > > TTS (Text-to-Speech) converts text to speech. The project integrates engines such as Edge/Azure and supports pronunciation correction via a global polyphonic-character dictionary, fixing misreadings of ambiguous Chinese characters and English proper nouns.
Millisecond-Level A/V Sync (timing.json)
When the TTS engine generates narration, the project exports precise start/end timestamps for every sentence and even every word into a timing.json table:
On the rendering side, React components use the useCurrentFrame() hook to get the current frame. The system maps frame numbers to absolute times in timing.json, dynamically triggering subtitle highlighting, chapter progress bars, and scene transitions. Even if TTS pacing changes, the visuals never fall out of sync.
Platform-Specific Distribution
| Platform | Aspect Ratio | Visual / Script Tone | Core CTA |
| :--- | :--- | :--- | :--- |
| Bilibili | 16:9 landscape | Dense info, big type, minimal whitespace; welcome-and-follow script style | Customized "triple-combo" (like/coin/favorite) animation |
| YouTube | 16:9 landscape | SEO title control (<70 chars); auto-generated 0:00 chapters | "Like, Subscribe & Share" animation |
| Xiaohongshu | 3:4 portrait | Recommendation/knowledge sharing; emoji-heavy copy, auto #topic# tags | "Like, save, follow" text animation |
| Douyin / WeChat Channels | 9:16 portrait | Only 9:16 highlight clips, no long videos; casual, shareable copy | Plain-text "like and follow" |
The Human-in-the-Loop Rule: podcast.txt
> "If the script is bad, 4K rendering is still garbage."
This is the heaviest warning in the README: the more automated the pipeline, the more critical human review of the script becomes. The author highlights four dimensions of manual optimization:
1. The Breath Test: Read the spoken-style script aloud in your head. Any sentence that leaves you gasping must be cut shorter immediately.
2. Three passes of refinement: (a) fix grammar and tongue-twisters; (b) cut filler like "so today let's talk about..." and get to the point; (c) adjust pacing, inserting manual pause markers near words that need emphasis.
3. Eliminate pronunciation minefields: TTS easily misreads numbers, abbreviations, proper nouns, and polyphonic Chinese characters. Fix awkward words in phonemes.json with precise phoneme mappings, or substitute homophones (e.g., rewrite "AI" as the Chinese term for artificial intelligence).
4. Strict word-count budget: Chinese narration runs about 280 characters per minute. A 5–10 minute video script must stay between 1,400 and 2,800 characters—no padding.
Academic References
1. *Remotion: Programmatic video editing and rendering via React and Puppeteer.* (Web Graphics Association, 2023) — Argues for the architectural advantages of browser rendering stacks (DOM, Canvas, WebGL) with headless frame capture; evaluates frame-rate sync and multi-core render performance.
2. *Automatic audio-text alignment and timestamp extraction in managed TTS pipelines.* (Speech Communication Journal, 2024) — Quantifies the accuracy of forced-alignment sentence/word timestamps from TTS and their impact on subtitle burn-in and transition sync.
3. *A survey of text-to-speech pronunciation lexicon correction and G2P (Grapheme-to-Phoneme) modeling.* (IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2022) — Justifies G2P dictionary intervention for domain terms, Chinese-English mixed text, and polyphonic characters, underpinning configs like phonemes.json.
*Source: zhichai.net forum post on the Agents365-ai/video-podcast-maker open-source project.*