Local generative AI on an RTX 4070, disguised as a streaming station

This is the current state of local generative AI on my VRAM-poor RTX 4070, disguised as a streaming station. (I should’ve listened to Dilitha Wijayapala and bought the 4090 back then)

https://ai-music.madeby.dev - 150+ tracks across drum & bass, synthwave, nightcore, rock, house and more.

The pipeline:

• A local LLM writes themes, titles and lyric drafts inside pinned “style packs”, so songs vary without drifting out of genre
• YuE2 generates vocals and arrangement, served locally via a GGUF build
• A community LoRA handles instrumentals, because the base model always wants to sing
• Qwen Image makes the cover art from each song’s style and lyrics
• MiniMax H3, also running locally, turns each cover into a looping video

Put on some synthwave and tell me which track is the worst 🎧