You need to try YuE2 NOW! It’s Suno v5 level, open weights, low VRAM, FREE, FAST music gen & can even be trained in Maestro v2.2.0!
Also, MAJOR v2.2.0 AI prompt improvements! GPT-6 ASTRA trained Maestro’s local agents & made them awesome!
Full Maestro v2.2.0 Release Notes:
Unified AI Prompt Enhancement, YuE2 music, and more control over Director productions
•One Enhance workflow: develops short ideas and adapts detailed scripts to the selected model. Enhance now, or let a queued job enhance immediately before generation. An optional Use by default setting remembers Enhance on generation; saved source prompts and drafts remain available for review and retries.
•Stronger H3 prompt writing: improved silent-action parsing, dialogue ownership, choreography, camera direction, first-frame continuity and multi-window timing. Director shares the relevant writing guidance. Prompt review remains available for drafts that need attention.
•YuE2 3B music: the new default (and best) music model produces 48 kHz stereo songs, with Direct generation selected initially, optional composition planning, ABC scores and source-song covers.
•My music (Experimental): review recordings and lyrics, train and resume personal style adapters, compare checkpoint auditions, reconstruct source audio, and import/export styles.
•Director music controls: adjustable clip maximum and GPU limit, working Cut Speed, full-song timing, and musical/vocal cues for camera planning. Longer H3 clips can approach 14.4s; instrument cutaways retain the assigned singer's voice off screen. Shot edits now save reliably.
•All folders gallery: browse and search across output folders while keeping each item's source folder and the generation destination clear. Refresh, pagination and folder-qualified actions are fixed.
•TaoMate H3: optional experimental three-step Frames acceleration, with pinned downloads and compatibility checks. Includes internal token-refiner weights; it is not a separate video-refinement pass.
•Image, queue and runtime fixes: 21:9 image aspect, original/enhanced prompt details, multiline-image and Z-Image decoding fixes, corrected CivitAI model reloads, collapsible completed-job history, repeated-upload deduplication, and improved Linux local-LLM setup and diagnostics.