Maestro：本地 AI 视频创作与剪辑

An all-in-one, 100% local AI video, image, and music studio. Director mode plans full music videos and short films from a single prompt. …

Maestro A one-click local AI creative studio, director, and video editor for creators. Maestro combines a modern interface with powerful image, video, and audio generation, an LLM-directed production workflow, and a non-destructive multi-track editor. Optimized for LTX-2.5, MiniMax H3, and the latest local creative models and LoRAs. Features · Install · Updates · Guides · Full update history What it does Create with Studio, plan a production with Director, and finish it in the Editor. Local models run on your own PC; external AI services and private remote access are optional. Create, direct and edit Director — music videos and short films. Turn a song or story into a shot plan, prepare character references and keyframes, generate the clips, and assemble the result. Use Auto mode for a complete production or Manual mode to review each stage. The project dashboard keeps plans, prompts and takes together so you can revise individual shots. Director guide. Studio — video, images and audio. Generate from text or references; animate, extend, blend and edit video; create and edit images; or make music, speech, cloned voices and sound effects. Queue jobs, enhance prompts and adjust model-specific controls. Models include MiniMax H3, LTX, Flux, Qwen, MOSS and Fish Audio. Studio guide. Editor — finish on a timeline. Arrange video, audio and titles with trimming, transitions, speed, volume and canvas controls. Bring in a Director production or gallery clips, send a shot back through AI generation, and export the finished edit. Keep control of your creative workflow Reusable characters and references. Save and share characters with appearance references and optional voices. Combine image, video and audio references in supported workflows to guide identity, action and sound. Characters and RefMods. Built-in AI writing. Local language and vision models help develop prompts and plan productions, with models downloaded on first use. Optional external providers include OpenAI, Anthropic and compatible local endpoints such as LM Studio. Local AI guide. Models and LoRAs in one place. Browse compatible CivitAI LoRAs and community checkpoints, manage downloads and updates, and use prompting guides and recommended strengths. H3 checkpoint imports. Performance matched to your PC. Auto-Tune uses your GPU, VRAM and RAM to choose starting settings. Quantization, memory offloading and duration recommendations help smaller GPUs, while advanced controls remain available. Performance guide. An organized library, wherever you work. Separate projects into workspaces, search across folders, save favorites and reload generation settings. Use the responsive interface on a phone or tablet, with optional completion alerts and private Tailscale access to your running PC. Remote access and notifications. Available inputs, editing tools and acceleration depend on the selected model and hardware. See the guides and model details for workflows, setup and model-specific licenses. Updates View the complete update history · Detailed changelog The version you are running is shown next to the Maestro title in the UI. To update, use the launcher's Update button in Pinokio. Runtime upgrades are explicit menu actions. A missing performance-runtime marker does not automatically run Update when you open the launcher. If offered, Start with Compatibility Runtime keeps the existing runtime available. v3.2.5 (2026-10-24) Faster Fish speech: whole-frame CUDA Graph decoding when the full model fits, cached experts and partial acceleration for smaller GPUs, and compatible RTX 30-series attention. Add Fish Drama 3 Small with voice references and streaming synthesis assembled into one gallery audio file. Long-form speech writing: saved style/personality instructions, draft targets based on the selected duration, bounded section writing and recovery, richer Fish performance cues, and retained incomplete drafts with clear reasons. Fish Auto length lets the script finish naturally within safety limits. H3 checkpoints and memory: support Eros Max beta5/beta6 mixed layouts, share compatible imports across Frames and References, and correct Auto-owned H3 placement on Windows 16GB RTX 50-series / 64GB RAM systems. Director recovery: repair repeated source-action allocations and missing physical-continuity metadata without rewriting accepted shot content. Fix concurrent Windows soundtrack-cache publication. Available on Main and Dev. Update the idle installation in Pinokio, restart Maestro and refresh the browser. Short Fish tests completed on 24GB, 12GB and 8GB GPUs; speed and delivery depend on hardware and the script. The reported RTX 5080 regression still needs a matched hardware comparison. See the v3.2.5 release notes. v3.2.0 (2026-10-22) Startup and mobile: prioritize model controls, recover interrupted startup, keep mobile controls visible without a refresh, and show workspace loading continuously while the saved folder and gallery are restored. Director and H3 references: improve source-action extraction, outline compatibility and style handoff; retain rejected writer drafts with reasons, preserve accepted multi-window prompts during repairs, and add explicit video soundtrack roles in Studio and Director. Repair Animate trim bounds and H3 VDN, Music3 and local-writer downloads. Audio and drafts. 21 October fixes: multi-window Face Refiner blending, visible warnings when optional finishing fails, native DLSS availability checks, large-input WGPKernels preparation, preserved manual INT8 choices during Auto-tune, and external-LLM Director music-arc response compatibility. The Qwen processor download repair is also verified against the current public asset pin. FastH3 V2: experimental eight-step text-to-video with generated audio, trained VSA-H3 attention and its own sampling recipe. FastH3 OmniRef: experimental eight-step video/audio generation from ordered image, video and audio references. Official BF16, INT8 and FP8 options, plus a separate Bla