Real-time scene narration for blind & low-vision users. Point your phone ahead. It speaks up only when something matters (a person, a change, a hazard), and answers when you ask.
English · 简体中文
Most camera assistants either narrate every frame or wait for you to tap. Sightline runs on JD's JoyAI-VL-Interaction, a streaming vision model trained to decide when to speak, so it stays quiet until there's something worth saying.
⚠️ Experimental aid. Don't rely on it for safety.
Sightline needs JoyAI-VL-Interaction running as a local OpenAI-compatible adapter (default :8070). It's an 8B model, so a GPU is required. The weights are on Hugging Face (jdopensource/JoyAI-VL-Interaction-Preview, Apache-2.0), and JD's repo has a script that downloads and serves them. Once the adapter is up:
pip install -r requirements.txt
cp .env.example .env # point ADAPTER_URL at your adapter
python server.py # serves on :8092Open it on your phone over HTTPS (the camera needs a secure context, so use a Cloudflare or Tailscale tunnel), tap Start, point ahead, and ask questions out loud. focus = balanced / hazard / detail controls how chatty it is.
Built on JoyAI-VL-Interaction (JD, Apache-2.0) and faster-whisper (MIT). Licensed under Apache-2.0.