xhs-video-workflow

Feed the Markdown note into the pipeline: it analyzes the content into an 8-13 slide storyboard, builds GSAP-animated HTML slides at 1080x1920, generates ChatTTS voiceover per slide, and renders an MP4 with HyperFrames. Total runtime is roughly 35-45 minutes.

Use ChatTTS with source set to huggingface and a fixed seed of 42 for consistent output across runs. Write numbers as Chinese words (e.g., 百分之八十 instead of 80%) so they are pronounced correctly, and export each segment as a 24kHz WAV file.

ChatTTS is the primary engine, offering offline generation and reproducible results via fixed seeds. Edge TTS is the fallback when ChatTTS is unavailable or a Microsoft voice like zh-CN-XiaoxiaoNeural is preferred, but its output must be manually split into per-slide segments.

Blank slides are caused by a missing CSS rule: you must add .slide.clip{opacity:1!important} or HyperFrames renders nothing visible. Also verify the GSAP script tag appears only once, since duplicate inclusion breaks the animation timeline.

Extract each WAV duration with soundfile into a timing.json file, then set each audio element's data-start to its slide's start time plus 0.3 seconds. This offset lets the GSAP fade-in finish before the narration begins.

The pipeline does not include auto-publishing to Xiaohongshu and requires local dependencies including Python 3.8+, Node 18+, ffmpeg, ChatTTS, and Playwright. Dynamic APIs like Date.now() or Math.random() are forbidden in slides because they break deterministic rendering.