Maestro v2.0 is out! Its got stuff and things!

Maestro v2.0 is out! & it’s loaded with new features!

Maestro is a free, open, local, unrestricted AI media studio that’s loaded w/ the latest SOTA models & can autonomously gen films & music videos.

v2.0.0 (2026-09-03)

A complete create-to-edit workflow

• Added the new full-screen Editor Mode with multi-track video, audio, and title editing; 21:9 canvases; canvas transforms and snap guides; transitions, speed, opacity, volume, fonts, undo/redo, project history, and hardware-aware export.
• Director productions can be opened as editable timelines with separate shot clips and the complete soundtrack, while any timeline clip can make a round trip through Maestro AI and return as a new take.
• Reorganized Studio into clear Video, Image, and Audio workflows. Video editing tools now live beside generation, Image adds dedicated New/Edit/Upscale/Outpaint modes, Audio includes Revoice, and Finish adds reusable film grain.
• Added unified long-form planning up to 60 minutes with duration presets, exact timecodes, direct window counts, media/story-aware Auto duration, and Faithful or Creative AI window prompts. Frequently used Studio modes, models, planning choices, and H3 optimizations now survive restarts.

MiniMax H3, references, and local intelligence

• Added H3's native 768p tier, 21:9 canvases, and Regenerate 2K workflow, plus Alibaba PAI FL2VA and Ref2VA acceleration presets with native PDD support.
• Added optional experimental H3 Fused 4-Step Frames and References models. Both share one pinned INT8 ConvRot checkpoint, default to the published four-evaluation recipe, expose 4-8 Total Steps, and use SLA sparse attention with a safe dense fallback.
• Added a reusable Omni character library for named image + voice or video references, automatic reference-duration budgeting, and stronger reference isolation so identity media is not mistaken for a start frame. Exact Subject/Speaker bindings keep each character's face, voice, and dialogue together and reject phantom subjects.
• Rebuilt H3 prompt planning around causal story continuity, exact dialogue preservation, official Context-IR guidance, model-aware token fitting, and safer multi-window continuation. Faithful Studio planning now keeps the user's event/dialogue schedule authoritative while the LLM concentrates on cinematography.
• Added Qwen3.8 27B Uncensored with creative thinking controls, prompt-enhancement telemetry, and non-thinking structured-output paths.

Remote workflow and release polish

• Added completion alerts, optional chimes, encrypted closed-app Web Push, an installable Maestro web app, and optional private HTTPS access through each user's own Tailscale account. Windows restores opted-in Tailscale access after Maestro restarts without repeated approval prompts.
• Added per-clip, multi-window, and full Director completion estimates, including cache-aware calibration for First Block Cache and private local timing history for more accurate future estimates.
• Expanded gallery details and search across model, resolution, LoRAs, H3 optimizations, prompts, window counts, and generation timing, while making the viewed or playing clip the reliable active Studio target.
• Updated Maestro's orange app icon, unified the responsive Director / Studio / Editor header and version display, added Director first-frame thumbnails and full-rate iOS Editor preview playback, simplified the Pinokio menu, and preserved the v1.9.1 llama.cpp nightly-download hotfix.