Valkyrie Studio

Valkyrie Studio is a Rust desktop app for loading a local .model3.json, moving and voicing the avatar, and recording video. It uses the MIT-licensed Purism Core .moc3 runtime; the official Live2D Cubism SDK is not built or shipped. Bring your own model and FFmpeg. No model assets or API keys are in this repository or the release packages.

Watch the full-screen Vaelari walkthrough. The 81-second video shows a scripted agent driving the model and ElevenLabs speech. It was recorded with an earlier renderer, so it demonstrates the workflow rather than the exact appearance of this Rust build.

Get started

1. Download the package for your platform from Releases. Windows and macOS each have their own package; Ubuntu, Fedora, and Arch have separate Linux builds. These are portable alpha archives, not installers or signed apps.
2. Extract it. Install FFmpeg with utvideo and the encoder for your desired output: libx264 for MP4, libvpx-vp9 for transparent WebM, or prores_ks for transparent MOV. Put ffmpeg on PATH or set FFMPEG to its full path.
3. Start valkyrie-studio (valkyrie-studio.exe on Windows). In Stage, select Browse models… and choose your local .model3.json. The app remembers recent models and reopens the last available one next time. Keep its referenced .moc3 and textures in place.
4. Choose Vertical 9:16 or Horizontal 16:9 in Stage or Capture, then frame with Zoom and Pan. Press F11 for fullscreen. The controls follow desktop display scaling; UI size adds a 75–250% adjustment. For vertical shots, pick a social platform to preview its safe area and representative chat, action, caption, and navigation controls.
5. In Voice, enter an OpenAI or ElevenLabs key, select a voice, choose an expression (strength 0–5), type text, and select Speak and animate. Eleven v3 can receive a matching emotion tag for its vocal delivery. On Windows you can save the key encrypted for your user account; on macOS/Linux keys remain session-only or can come from environment variables.
6. In Capture, choose a new .mp4, .webm, or .mov output path and matching codec, then select Start recording and Stop recording. Wait for saved. Export runs in the background and may need substantial temporary disk space.

The user guide covers the controls and current alpha limits. The build guide covers source builds. Validation separates measured Windows behavior from CI compilation on the other platforms.

Connect an agent

Keep Studio running with a model loaded. In the extracted package, run valkyrie-agent --connect (.exe on Windows). It should report READY. The CLI reads the session token from the app's per-user data folder and connects only to 127.0.0.1:4141. An agent can send one JSON command at a time, for example:

valkyrie-agent '{"op":"emotion","name":"joy","duration":3}'
valkyrie-agent '{"op":"puppet","energy":1.5}'
valkyrie-agent '{"op":"gesture","name":"nod","duration":0.8}'

In agent mode, the model now keeps moving between commands with gentle head and body sway, eye movement, and blinking. The Inputs tab has a Motion energy slider and gesture preview buttons; puppet accepts 0–2 (1 is normal, 0 disables automatic movement). An agent can still send precise tracking at 20–30 Hz, and gesture adds a short nod, shake, tilt, or lean without replacing that tracking. Speech continues to drive the mouth.

For a long voiced performance, copy the JSON example from the Voice tab or package, then run valkyrie-perform --script performance.json --output take.mp4. Each line's emotion now controls an Eleven v3 voice tag, facial expression, and matching gesture; expression_intensity runs from 0–5, and optional motion cues turn the head during that line's measured audio. The runner prepares voice lines before capture so generation does not pause the recording, aligns mouth movement and 20 Hz head cues, records bounded chapters, and joins them. The default pause between lines is 0.18 seconds. See agent connection, performance scripts, and the API.

Current alpha scope

The Rust app provides the Purism renderer, A.R.I.A-derived particle physics with motion styles and per-group tuning, agent/UDP tracking with editable listening addresses, typed OpenAI and ElevenLabs speech, a voice picker, portrait safe-area guides, scalable UI, and H.264/H.265 MP4, VP9 WebM, or ProRes MOV recording. Model rendering and audio playback were tested on Windows with a local Vaelari model. CI compiles and tests the six build targets, but macOS and Linux GUI, camera, FFmpeg, and package startup still need physical target-machine tests. Phone tracking uses a UDP sender; the optional Python webcam helper is started separately. The older Ruby interface's attachment/background editing and direct VTube Studio API connection have not been ported. See alpha status.

The app is MIT licensed; Purism Core's MIT notice is included. Model rights, FFmpeg build terms, and other dependency notices are separate. See third-party components. The cleared walkthrough is a video permission, not permission to redistribute the model files.