Sound / Vision: open-source self-hosted app turns an idea into songs, album art, and AI music videos
LawrenceOfTheLabia · reddit · 2026-10-11
A Reddit dev open-sourced Sound / Vision, a self-hosted web app that takes you from an idea to a finished song, album art, and a lip-synced music video on one local Linux GPU box.
What it does
- Songs via YuE2, ACE-Step 1.5, or MiniMax Music 3 in local ComfyUI; multiple takes, covers, or your own tracks
- Analysis: stems, beats, sections, word-level lyric timing
- Album art from Krea 2 / Qwen Image 2.1 in one house style
- Videos: animated art, code-written motion graphics timed to the song, and lip-synced MVs via MiniMax H3
- Auto-cut TikTok/Reels/Shorts clips; Influence Engine trains a 'sound' to steer YuE2
Architecture & limits
- FastAPI systemd service + Dockerized ComfyUI; LAN access without keys, token pairing for outsiders
- Use a local LLM via Ollama or your own Claude/ChatGPT/Grok subscription (no API keys); MCP servers for hosted models or ElevenLabs Music
- Needs a 24–32 GB NVIDIA GPU (built on a 5090), 200 GB disk, 1h first install
- YuE2 is non-commercial; MiniMax H3 excludes US/EU/UK/Korea. MIT licensed.
More from coding & agent
- Cache mismanagement is silently burning agent budgets: how to fix it — brandon_galang · 2026-10-12
- How to Review Agent-Written PRs: Break the Code on Purpose and See If Tests Fail — ITower__Education · 2026-10-12
- Matt Pocock's Opus chief-of-staff skill runs 'bonkers' latency tests against OBS — mattpocockuk · 2026-10-12
- User cancels $200 ChatGPT Pro over Codex rate limits and slow 35 tok/s speed — karmay007 · 2026-10-12
- ThursdAI weekly: OpenAI drops 722 math manuscripts, 100k agents simulate Toronto, Haiku 5.5 hits 10 cents — altryne · 2026-10-12
- 100,000 agents simulate Toronto without crashing: every light is a person you can talk to — altryne · 2026-10-12