365 ·GitHub
What is this?
A control layer that lets an AI virtual streamer run her own livestream: it filters and queues viewer chat (danmaku), manages the character's wardrobe and small gestures, composites a live overlay, and hands the feed to OBS or a built-in stream, while Vidu S2 generates the actual avatar video, voice and replies.
A control layer that lets an AI virtual streamer run her own livestream: it filters and queues viewer chat (danmaku), manages the character's wardrobe and small gestures, composites a live overlay, and hands the feed to OBS or a built-in stream, while Vidu S2 generates the actual avatar video, voice and replies.Is this a full avatar-generation app or just a control layer?The avatar's video, voice and spoken replies are all generated by Vidu's S2 real-time digital human service; this project is just the middle layer - a local Python server plus workbench that filters chat, triggers outfit/action changes, and composites the overlay for OBS.
The avatar's video, voice and spoken replies are all generated by Vidu's S2 real-time digital human service; this project is just the middle layer - a local Python server plus workbench that filters chat, triggers outfit/action changes, and composites the overlay for OBS.How do I get it running?Run the included script to fetch and verify Aliyun's RTC web SDK, then start server.py - the service only listens on localhost and the same address opens the local workbench UI where you build characters, outfits and manage the stream.
Run the included script to fetch and verify Aliyun's RTC web SDK, then start server.py - the service only listens on localhost and the same address opens the local workbench UI where you build characters, outfits and manage the stream.What do I need before trying it?macOS or Linux, Python 3.11+ and ffmpeg built with libx264/aac, plus a Vidu Open Platform API key (starting with 'vda') loaded with credits - real-time avatar generation costs about 169 RMB per hour at the official rate. Evolink (for images/outfits/voice cloning), OBS 30+, and a YouTube stream key are optional.
macOS or Linux, Python 3.11+ and ffmpeg built with libx264/aac, plus a Vidu Open Platform API key (starting with 'vda') loaded with credits - real-time avatar generation costs about 169 RMB per hour at the official rate. Evolink (for images/outfits/voice cloning), OBS 30+, and a YouTube stream key are optional.What can the avatar actually do, and what can't she do yet?She can chat, change into a pre-made outfit (about 6 seconds per swap), and do small gestures like heart-hands or waving, plus idle small talk when no one chats. She can't sing, stand up, show her full body, dance much, remember viewers across sessions, or push product sales, and changing background/props is disabled by default.
She can chat, change into a pre-made outfit (about 6 seconds per swap), and do small gestures like heart-hands or waving, plus idle small talk when no one chats. She can't sing, stand up, show her full body, dance much, remember viewers across sessions, or push product sales, and changing background/props is disabled by default.Why is background/prop changing turned off by default?Outfit changes reopen a whole new reference photo so clothes stay intact, but background/prop changes instead get Vidu to redraw the live frame around a reference image that doesn't include her outfit - in one public test, swapping backgrounds twice visibly distorted her clothes, so it's disabled unless you set safety.livevisualedits in config.json for private testing only.
Outfit changes reopen a whole new reference photo so clothes stay intact, but background/prop changes instead get Vidu to redraw the live frame around a reference image that doesn't include her outfit - in one public test, swapping backgrounds twice visibly distorted her clothes, so it's disabled unless you set safety.livevisualedits in config.json for private testing only.Can I try it without spending money on Vidu?Yes - the live room page has a dry-run mode that skips Vidu entirely and just demonstrates chat filtering, queueing and overlay, and tests/devmockstack.py runs the whole control flow against a local mock Vidu.
Yes - the live room page has a dry-run mode that skips Vidu entirely and just demonstrates chat filtering, queueing and overlay, and tests/devmockstack.py runs the whole control flow against a local mock Vidu.Are the streamer's face, voice and persona included in the repo?No - the author's avatar image, persona, voice and overlay-card assets used in his own tests are excluded via .gitignore, so you need to supply your own character image, voice and persona.
No - the author's avatar image, persona, voice and overlay-card assets used in his own tests are excluded via .gitignore, so you need to supply your own character image, voice and persona.Can it recover automatically if Vidu drops the connection?Yes - turning on autorecover (via API call or config.json) retries every 3 seconds after a disconnect, up to 6 times per hour, but won't retry if the session ended due to low credits or content moderation.
Yes - turning on autorecover (via API call or config.json) retries every 3 seconds after a disconnect, up to 6 times per hour, but won't retry if the session ended due to low credits or content moderation.How does it route viewer chat without calling an LLM?Danmaku first passes a moderation.json filter list, then live/router.py matches it by rule into chat, outfit, prop, background, action or reset categories - no model call, so it's instant and free.
Danmaku first passes a moderation.json filter list, then live/router.py matches it by rule into chat, outfit, prop, background, action or reset categories - no model call, so it's instant and free.What are similar projects?dsh-voice-ai-girlfriend, openspline, TalkLikeYou-Streaming-RealTime-Demo, Duix-Avatar, mesh-avatar-studio, hongguo-player, srs, EverydayWechat
dsh-voice-ai-girlfriend, openspline, TalkLikeYou-Streaming-RealTime-Demo, Duix-Avatar, mesh-avatar-studio, hongguo-player, srs, EverydayWechatWhich topics is it in?obs-studio, avatar, digital-human, danmaku, video-streaming, voice-cloning, chat, real-time, chatbot, streaming, china, local-first
obs-studio, avatar, digital-human, danmaku, video-streaming, voice-cloning, chat, real-time, chatbot, streaming, china, local-first