MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent
MIRA revises text-to-music prompts so generated audio better matches explicit and implied user intent.
The paper frames text-to-music intent alignment as satisfying a per-request rubric of explicit requirements and implied musical intent, covering instrumentation, structure, rhythm, and mood. MuRA-Bench is built from real platform requests curated by music experts. MIRA, a test-time agent, grounds a request into rubrics and searches prompt revisions for a black-box generator, iteratively generating audio and verifying it with a trajectory-aware tree search under a bounded budget. Across open-source and commercial backends, MIRA improves alignment enough for an open-source generator to approach systems such as Suno and Mureka.
- Intent alignment is scored with per-request rubrics instead of one relevance score.
- MuRA-Bench uses expert-curated real platform music requests.
- MIRA searches prompt revisions with generate-and-verify tree search.
- An open-source generator approaches commercial systems such as Suno and Mureka.
From the original article
1 min read at huggingface.co
Text-to-music systems produce increasingly convincing audio, yet evaluation reveals little about whether the result matches user intent. A global text-audio relevance score can overlook the implicit intent in underspecified prompts and mask failures in specific requirements, such as instrumentation, structure, rhythm, or mood progression.
By Zekai Liu, Zhilin Wang, Xuzheng He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.