Blizaine/Maestro: a local AI video, image and music studio driven by Director Mode
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. Built on the WanGP pipeline. Install via Pinokio.
At a glance
- What is it?
- Maestro bundles image, video and audio generation behind one interface and adds an LLM-directed pipeline that plans music videos and short films from a single prompt. It installs through Pinokio, and the README is explicit about where the model licences stop being permissive.
- Who is it for?
- Maestro fits creators with a capable local GPU who want shot planning, generation and editing in one place instead of stitching three tools together, and who are comfortable with the Pinokio install path. It is the wrong choice if you need a hosted service, if your GPU is small, or if you intend to sell output built on YuE2 or the real-audio tokenizer, whose weights carry their own noncommercial terms.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Maestro actually solves for a solo creator
Generating a music video locally is not one task. It is shot planning tied to the beat, keyframe generation with a consistent character, per-model prompt formatting, clip generation, and then assembly. Most local setups leave the creator to move files between a text model, an image model, a video model and a timeline editor by hand. Maestro's claim is that it collapses that chain into one application: the README describes it as a "one-click local AI creative studio, director, and video editor", and Director Mode is the part that does the directing.
The audience is narrow on purpose. Everything runs locally, so the ceiling is your GPU, not a subscription tier. The README targets LTX-2.5, MiniMax H3 and what it calls the latest local creative models and LoRAs. If you already run Wan or Hunyuan checkpoints and want a planning layer on top, this is aimed at you. If you want a browser tab that generates video without a GPU, it is not.
Director Mode: how a prompt becomes a finished cut
Director Mode has two skills. Music Video takes an audio track and plans shots against it: the README states the LLM analyzes BPM, sections such as verse, chorus and bridge, and energy, then writes shots that land on the downbeats. Speaker transcription and diarization let you name and target individual voices or singers. Clip length can be set to Auto or to a model-aligned maximum, and the README notes shorter clips are available for smaller GPUs while preserving the full song. Short Film instead works from a screenplay, with named characters, dialogue and continuity across cuts, and a pacing-bias slider that controls cut frequency.
The architecture matters more than the feature list. Director v2 separates screenplay writing, structured shot planning and model-specific formatting. Native H3 plans go straight through an H3 compiler; other models get a per-model polish step. The README says Director shares Studio's adaptive writing and continuity guidance, protects exact dialogue and source-song vocals, and retains full action and camera descriptions through final compilation. That separation is the reason a plan can survive a model swap at all, and it is the most defensible design decision described in the repository.
Auto Mode runs the whole chain end to end: analyze, plan, generate images, generate clips, combine. Manual mode stops at each step for review. The trade-off is that Auto Mode is only as good as the planner, and the README does not describe what happens when a planned shot is rejected by the video model at generation time.
Installing Maestro through Pinokio and running a first job
The homepage listed for the project is https://pinokio.computer, and the repository root contains pinokio.js, install.js, start.js and update.js, so Pinokio is the supported install path. There is no pip or npm install described in the README. The scripts at the repository root are the entry points Pinokio calls; you do not run them by hand.
Install Pinokio first, then add Maestro as an application inside it. The install script handles dependency setup, and the separate torch.js, sol_torch.js and sam_install.js entries suggest the installer branches on your hardware rather than assuming one CUDA build.
git clone https://github.com/Blizaine/Maestro.gitCloning gives you the working tree, but the README does not present a manual launch command. The launcher entries are start.js, start_classic.js and start_sol.js, with launcher_profile.js selecting between them. Treat the clone as a way to read docs/ and CHANGELOG.md before installing, not as the install itself.
On first launch, Performance Auto-Tune detects GPU, VRAM and RAM and picks a profile, quantization, VAE tiling and a VRAM safety coefficient. The README frames this as removing "Profile 1 vs 2 vs 4.5" guesswork, and manual control remains under Show advanced settings.
For a first real job, open Director Mode, choose Music Video, drop in an audio track, and leave Clip length on Auto. Auto Mode then runs analyze, plan, generate images, generate clips and combine without stopping. Expect a download on first use: the README notes the transcription model is roughly 300 MB and that the UI shows live download status rather than a spinner.
VRAM, out-of-memory runs and where the pipeline breaks
The honest limitation is memory. Maestro runs image, video and audio models on your own card, and the README's own answer to that is the Auto-Tune layer plus an OOM recovery banner that suggests lowering VRAM headroom when a generation runs out, with one-click apply. A recovery banner is a useful feature, but it is also an admission that out-of-memory runs are expected rather than exceptional.
The optional models raise the floor further. The README states H3 VDN requires Triton and additional VRAM. For YuE2 music training, it says the exercised training GPU has 24 GB VRAM, and that matching a specific singer's voice is not guaranteed. Those two sentences together tell you what the training path costs and how reliable it is.
There is also a gap in the documentation around failure. The README says interrupted jobs are held after an app restart, and that enhancement failures stop for attention instead of silently generating a fallback. What it does not document is rollback for a completed but bad generation, or how to resume a Director plan that failed halfway through clip generation. If your workflow depends on deterministic recovery, that is unverified ground.
Studio Mode versus a dedicated node-based tool
Studio Mode is the manual counterweight: direct access to every model and knob. Video covers create, extend, blend, retake, edit, outpaint, repaint, recast and upscale across MiniMax H3, LTX-2.5/2.3, SCAIL-2, Wan and Hunyuan. Image covers Flux 2 Klein 9B, Krea 2 RAW/Turbo and Identity Edit, Qwen Image Edit. Audio covers MiniMax-Music3, ACE-Step and YuE2 for music, H3 Voice Audio, Kugelaudio and Qwen3 TTS for speech and cloned voices, and H3 or MMAudio for sound effects.
The natural alternative is ComfyUI. The difference in approach is real: ComfyUI exposes a node graph where you wire the sampling, conditioning and post-processing steps yourself, which gives you control over the exact graph and makes sharing workflows a matter of a JSON file. Maestro hides that graph behind a fixed composition workspace, with reference cards above a prompt editor, model selection beside Generate, and Advanced groups for Performance, Finishing, LoRAs and Generation. You trade graph-level control for a planner that keeps character and dialogue continuity across shots.
If you need a custom sampling chain, ComfyUI is the better fit. If you need a hundred shots that hold the same character and hit the same beat, the graph is not the hard part, and that is the case Maestro is built for.
Maintenance, updates and the licence question
The repository is not archived and the last push was on 2026-09-17, so the codebase is current as of that date. There are no releases retrieved, which means updates arrive through the repository rather than tagged versions. The root contains update.js and VERSION, and CHANGELOG.md is tracked, so version state is inspectable before you pull. Pinokio's update path is the intended upgrade route; the README does not describe a migration process for saved projects across versions.
The licence field reports NOASSERTION, which means GitHub could not classify the LICENSE file automatically. Read LICENSE and THIRD_PARTY_NOTICES.md yourself rather than assuming a permissive default. The README is unusually direct about model terms: it states that YuE2 and the real-audio tokenizer weights have their own noncommercial model terms. That is a constraint on the music path specifically, and it sits alongside whatever the application licence turns out to be. Check both before building anything you intend to sell.
Editorial conclusion
Maestro fits creators with a capable local GPU who want shot planning, generation and editing in one place instead of stitching three tools together, and who are comfortable with the Pinokio install path. It is the wrong choice if you need a hosted service, if your GPU is small, or if you intend to sell output built on YuE2 or the real-audio tokenizer, whose weights carry their own noncommercial terms. Before committing, verify your exact GPU and VRAM against the Performance Auto-Tune profiles in docs/Performance-auto-tune.md, and check whether the models you plan to use carry separate terms in THIRD_PARTY_NOTICES.md.
Frequently asked questions
How do I install Maestro on Windows?
The README points to Pinokio, and the repository root contains pinokio.js, install.js and start.js, so Pinokio is the supported install path rather than a manual Python setup. Install Pinokio, add Maestro through it, and let the install script handle dependencies. No pip or npm install command is documented.
How do I use Maestro in Pinokio?
Once Maestro is installed as a Pinokio application, the launch scripts start.js, start_classic.js and start_sol.js handle startup, with launcher_profile.js selecting between them. On first launch, Performance Auto-Tune detects your GPU, VRAM and RAM and picks a profile, quantization, VAE tiling and a VRAM safety coefficient.
How do I use Maestro AI to make a music video?
Open Director Mode and choose the Music Video skill, then drop in an audio track. The README states the local LLM analyzes BPM, sections such as verse, chorus and bridge, and energy, then writes shots that hit the downbeats. Auto Mode runs analyze, plan, generate images, generate clips and combine end to end.
How do I use Maestro Studio?
Studio Mode gives direct access to every model and knob, with a composition workspace that places compact reference cards above a large prompt editor. Characters stay on the left, while Recipes, Resolution, Aspect, Duration and Advanced group on the right, and the model selector sits beside Generate and Add to Queue.
Community notes