xhs-video-pipeline

Create a 1080x1440 animated HTML with GSAP slides summarizing the article, capture frames at 10fps with Playwright, then combine frames with BGM and optional TTS using FFmpeg. The pipeline outputs an H.264 MP4 in about 10-15 minutes.

Write one narration line per slide, convert numbers to Chinese full forms, and run a single continuous chat.infer() call with InferCodeParams and manual_seed for consistent voice. Mix the WAV with BGM at 12% volume using FFmpeg amix.

ChatTTS runs locally and produces natural prosody but requires the newer load(source='huggingface') API and InferCodeParams object. edge-tts is an online alternative using voices like zh-CN-YunxiNeural with adjustable rate.

ChatTTS pronounces digits and percent signs as English by default. Write numbers as Chinese full forms in the script, such as 百分之八十 instead of 80%, to force correct Chinese pronunciation.

Windows console encoding mangles Chinese filenames in FFmpeg output display, though the file itself is fine. Copy the output to an ASCII English alias after composition to avoid confusion.

Videos are limited to 6-10 slides at 3.5-4.5 seconds each, and dense content like feature grids or bullet lists is unreadable at video speed. Each slide should carry only one core message with large fonts.