cycle-5 · 2026-08-08

Two tools in one page.

Part A is a static AMT landscape — what's NEW since the cycle-3 sweep (MuScriptor, MIROS, HeartMuLa, mason369, Magenta RT2 carried forward) plus four cycle-4 entries: ACE-Step v1.5 (Apache-2.0 music gen on 8 GB), Libretto (UC Berkeley's LLM agent for symbolic music), Rubato (one-pass audio→timestamped notation), and Stable Audio 3.0 (commercial-safe open music gen). Part B is the Long Horizon Prompt generator — type a concept, get production-ready prompts in nine formats: the original six (horizontal / vertical / square video, audio, music production, social post) plus three new publishing outputs for the personal-publishing leg: Lead sheet, Engraving spec, and MEI prep checklist. Vanilla JavaScript, no server, copy buttons.

No server No API calls Vanilla JS Copy-to-clipboard Dark + light 9 output formats cycle-5 sweep + publishing

Part A

AMT landscape — cycle-4 sweep

Agent brain · symbolic

Libretto

UC Berkeley · Jun 21 2026

paper only

"Giving LLM agents a sense of musical structure." LLM-native grammar with explicit onset slots, voices, and bar-level organization. Corpus-calibrated statistical music cloud for retrieval, generation, diagnosis, and revision through musician-readable feedback. Closes the cycle-2c "no good AMT agent interface" open question.

End-to-end · single-pass

Rubato

U. Washington · May 22 2026

paper only

Prompt-conditioned encoder-decoder that outputs timestamped notation in one autoregressive pass — no cascaded MIDI→engraving. New "InterMo" text representation. Beats cascades on notational accuracy. Closes the cycle-2c "cascaded hack" gap. License: research-only at paper stage.

Commercial-safe · open weights

Stable Audio 3.0

Stability AI · May 20 2026

commercial ⚠

Open-weights Small/Medium (459M / 1.4B). 6:20 music on consumer GPUs. Trained on licensed AudioSparx + ~500k Freesound CC — the only 2026 commercial-safe open music-gen. Sister: Stable Audio Open Small (on-device, 11 s clips in <8 s on Arm CPU, free for commercial use).

Max length
6:20
Medium
1.4B
Small SFX
459M
License
Comm. ⚠

First-of-kind · video-only

V2N (Video → Notes)

KAIST · ISMIR 2026

research-only

The first complete video-only piano transcription (VPT) system. From a top-view video alone, with no audio, predicts onset, offset, key hold, and velocity for every note. Shared temporal backbone with task-specific heads, per-frame supervision. Pairs with the cycle-3 KAIST stack (D3RM + D3PIA + PianoVAM).

Docs

Read the full cycle-4 sweep

RESOURCES-cycle-4.md

Full categorized GitHub + HuggingFace list, the cycle-4 recommended stack (with ACE-Step replacing YuE for the budget path), the license summary delta, and the next-action watchlist. Cycle-3 (MuScriptor hot-take + MIROS + HeartMuLa) carried forward — unchanged.

Carried forward from cycle-3 (still the right call): MuScriptor (multi-instrument SOTA, CC BY-NC 4.0 weights) · MIROS (2025 AMT Challenge winner, F1 0.5998) · HeartMuLa (Apache-2.0 lyrics / generation family) · mason369/music-to-midi (turnkey pipeline harness) · Magenta RealTime 2 (real-time foundation model). Full cycle-3 catalog in RESOURCES-cycle-3.md.

Part B

Long Horizon Prompt — full production pipeline

One idea in. Nine production-ready prompts out. Six for social / music / audio / video (horizontal, vertical, square, audio, music production, social post) plus three for the personal-publishing leg: lead sheet (LilyPond / Sheet Sage), engraving spec (page layout, font, hairpins), and MEI prep checklist (MusicXML / MEI validation, metadata, naming). Each output has a copy button. No server, no API, no rate limits.

Optional: add keywords / brand terms (comma-separated)

Horizontal video 16:9 · 1920×1080


              

YouTube, Vimeo, broadcast, embedded web hero, TV — anything that fills a widescreen frame.

Vertical video 9:16 · 1080×1920


              

TikTok, Reels, Shorts — full-screen portrait. Keep text in the upper third, action centred.

Square video / image 1:1 · 1080×1080


              

Instagram feed, LinkedIn carousel, Facebook — symmetric, composition-friendly.

Audio Podcast / Spotify / master


              

Use for Suno / Udio / Lyria / MusicGen — vocal directions, mix, mastering targets.

Music production DAW / notation


              

For a producer: tempo, key, structure, instrumentation, FX, mix targets.

Social post Twitter / LinkedIn / IG caption


              

Hook + body + CTA + 3-5 hashtags. Tight, conversational, platform-aware.

Lead sheet LilyPond / Sheet Sage / Verovio


              

Melody line, chord symbols, lyrics block, key + time + tempo. Drop into Sheet Sage or a fresh .ly file.

Engraving spec page layout · font · hairpins


              

Page size, margins, font, stem direction, system spacing, rehearsal marks, hairpin style. The "personal publishing" feel.

MEI prep checklist MusicXML / MEI · metadata


              

Work title, composer, lyricist, encoding date, source, rights, file naming, MEI schema version, validation tool.

What this is: a static prompt-templating engine. It takes your idea and re-shapes it into nine production-ready prompts, each tuned for a specific platform or output medium. Six for social / music / audio / video (horizontal, vertical, square, audio, music production, social post) plus three for the personal-publishing leg (lead sheet, engraving spec, MEI prep). No network calls, no API keys, no rate limits. Outputs are plain text — paste them into Suno, Udio, Midjourney, Veo, Sora, Runway, Photoshop, your DAW, LilyPond, Verovio, or music21.

What this is not: a content generator. The text is a hand-tuned prompt, not the finished video / track / image / score. The point is to short-circuit the blank-page problem — give the user a structured first draft that they can edit, not a finished product.