Piano Caption Loom
What if the caption an image is dreamed from were not written in one shot, but refined across visible rounds — a draft proposed from Karel's real piano, then a critic pushes back, the agents revise, and you watch the caption sharpen before the image regenerates?
Multi-agent propose → critique → revise loom. Emotion-aligned via valence × arousal (Russell 1980 circumplex). Cycle 2 of The Latent Piano Room.
Drop a Welcome Home track (or any audio file)
mp3 · wav · ogg · m4a · flac
No file? Begin plays Karel's recording (when logged in) or a warm resolving synthesized piano — both drive real musical analysis and the caption loom.
Read the design notesDesign Notes
The Loom: At each phrase boundary or onset cluster (min 7s apart), four specialist agents (Scene, Palette, Motion, Style) propose independent clauses from the MusicalFrame. A Critic checks each clause against the valence × arousal emotion target and emits concrete fixes. Revisers apply the fixes, producing a sharper caption each round. The final round also appends the numeric emotion anchor for maximum image alignment.
Emotion alignment: Valence (major + consonant → +1, chromatic/dissonant → -1) and Arousal (RMS dynamics + onset density) follow Russell's (1980) circumplex model of affect, as operationalized in arXiv 2512.23320 for music-to-image generation.
No-key fallback: When no FAL_KEY is configured, the route returns 501 and the UI switches to synthesized mode — the plasma / particle field renders continuously, driven by the same musical features and valence-arousal readout. The loom still runs every cycle; you watch captions improve even when no image arrives.
References: arXiv 2507.20536 (T2I-Copilot) · arXiv 2511.11483 (ImAgent) · arXiv 2512.23320 (Semantic Emotion Aligned Music→Image) · Russell 1980 circumplex · Bello et al. 2005 onset detection · Refik Anadol (Machine Hallucinations) · Memo Akten (Learning to See)