PSD2Live builds the Live2D rig from your PSD layer names

The mesh is decided by layer names, not by a drawing pass

PSD2Live begins before any drawing. It reads the layer names in the imported file, matches them against Chinese, English and Japanese naming, and assigns each layer a part type and a side. Paired parts such as left and right limbs are split apart on their own, and an adaptive mesh is fitted over what was recognised. On top of that it builds a head and body deformer chain, parameters for eye and mouth opening, gaze and eyebrows, idle, blink, nod and shake motions, and physics for front hair, back hair and jelly eyes.

The Start screen that opens right after import sets the model preset through Minimal, Default and Full choices, where Default carries loose clothing simulation. Two checkboxes carry the most weight: whether the file holds several independent parts such as separate left and right legs, and which layers should be split by mesh.

Quality tracks the PSD layering more than any other input. The parts called out are eye whites, pupils and upper eyelashes separated, with the pupil keeping the portion hidden behind the eyelid; an open mouth asset, or upper teeth, lower teeth and tongue on their own layers; front hair and back hair separated with margin around the occluded area; a body standing roughly upright; and layer effects and text rasterised beforehand.

Anything it cannot name is kept rather than dropped, and the type can be assigned by hand in the interface.

Four canvas modes sitting over six workspace presets

Editing runs in four modes: select, deform, edit and paint. The deform brush, mesh subdivision and cutting, front and back layering, creation of Warp and Rotation deformers, Glue, and deform paths marked experimental all live on that canvas.

Workspaces are stored as six presets covering edit, mesh, rig, animation, preview and physics. Moving to preview lets you check motions and physics before committing, then returning to edit, rig, animation or physics to change them. Keybinding presets exist for Photoshop, Blender and Cubism, with both a light and a dark theme available. A single project file opens in multiple tabs and keeps a branch-style history.

Motion work runs on a timeline with keyframes and curve editing under live preview. When a skeleton is present, preset motions such as crouching, waving and cheering can be appended on top of it. Assets sit in the same place: transparent images can be imported and placed, layers can be drawn and trimmed, and texture upscaling by 2x or 4x is optional.

Posing a skeleton that leaves as Cubism native shapes

The skeleton pass infers bones for limbs, tails and wings, and you pose them with FK or IK by dragging the end of a chain. Those bones are a working aid rather than something the shipped model depends on: at export everything is baked into Cubism native deformers, parameters and corrective key shapes, so the delivered model does not need this program running.

Differential handling belongs to the rig rather than the artwork pass. Layers can be marked as an on and off difference or as a pick one of many difference, which is how a mouth drawn in several variants stays attached to its parameter.

Physics editing pulls the bobs directly. Response curves update while you drag, and chained physics let one segment follow another. Evaluation is aligned with the Cubism Native Framework, so a swing set up here is meant to match what the official runtime produces. Swing generation for left and right and for up and down is automatic and comes with its own bobs, which you can then adjust visually.

Export writes four kinds of file and reports what it dropped

You can target Cubism 3.0 through 5.0, with 5.0 as the default. Features the chosen version cannot express are downgraded, and the downgrade shows up in the diagnostics rather than failing quietly.

Ctrl+G opens the export settings. Four outputs matter. The .psd2live file is the project itself, holding the original PSD, assets, settings, every edit and the branch-style history, and it is the one to keep if you intend to come back. The .cmo3 is a Cubism Editor model project for checking and refining in the official editor. The runtime family is .model3.json plus .moc3 plus textures, physics and motions, which have to be delivered together and loaded from .model3.json. The .psd2live.json is an export diagnostic report and cannot stand in for the project.

Export succeeding does not mean every runtime renders the result the same way, so check the model in the target editor and the target environment before delivery.

Packaging covers Windows x64 and Linux x86_64, and nothing else

Windows 10 and 11 on x64 get portable ZIP, EXE and MSI builds carrying the Java runtime, so unpacking or installing is enough to run. Linux x86_64 gets a Deb that bundles the runtime and the Cubism native preview, and it needs X11 with GLX, with XWayland acceptable. Anything else, meaning other architectures and other distributions, runs from source on JDK 21.

The native preview is optional and the built-in renderer needs no official SDK at all, so work is possible without obtaining one. Where the native preview cannot run, the application falls back to the built-in renderer on its own. It cannot run on pure Wayland with no XWayland, on aarch64, or on musl based distributions such as Alpine.

Three releases landed on 2026-10-02: v2.0.0, v2.0.1 and v2.0.2, and the last push to the master branch is dated 2026-10-01. The code is Kotlin under GPL-3.0, with third-party components recorded in a separate notices file. The project states that it is independent of Live2D Inc. and ships no proprietary Cubism SDK components.

Twenty five MCP tools, and no image generation among them

The application hosts a local authenticated MCP service exposing 25 public tools across observation, shape, assets, parameters, skeleton, motion, physics, simulation, export and history. Open it from the MCP menu entry for connection and installation, then copy the configuration matching your host. Hosts that speak Streamable HTTP connect directly; hosts limited to Stdio use mcp_proxy.py from the repository root.

The intended order is to let the agent call inspect to read the project first and only then make edits. Every write lands in the same history the interface uses, so it can be undone from the canvas.

What it will not do is generate images. MCP itself produces no imagery, so adding a new asset depends on the host having image generation of its own. Being able to call a tool is not evidence that a complicated modelling job succeeds, and the project keeps a status document recording real tasks alongside their successes and failures.

For agents writing code against this project, contributions are asked to record the host, the model, the number of rework rounds and the cost.

The command line can generate a model without opening the canvas

A JDK 21 install and the bundled Gradle wrapper are all the source path needs. The wrapper starts the GUI, runs the test suite, packages an installer for the current platform, and can also drive a conversion straight from a file:

./gradlew run                      # 启动 GUI（Windows：.\gradlew.bat run 或 run-gui.bat）
./gradlew run --args="--input examples/tml/psd-input/tml.psd --output build/example-output"   # 命令行直接生成模型
./gradlew test                     # 运行测试
./gradlew packageDistributionForCurrentOS   # 打当前平台安装包

A source build uses the built-in renderer by default, and wiring up the official Cubism SDK means obtaining it yourself and building the bridge library separately. Sample inputs live under examples in a tml and a ds directory. Saving a project uses Ctrl+S, and a first run is expected to open the in-application tutorial on F1, which comes in an 18 lesson route for beginners and a 13 lesson route for people who already know Cubism.

Editorial conclusion

PSD2Live suits a pipeline where a layered PSD already exists and rigging is the bottleneck, because the mesh, parameters, motions and physics arrive before any manual canvas work. It suits you less if your source art is flat, if you run aarch64 or musl and need the native preview specifically, or if you must target a Cubism version below 3.0. Before committing, check your layer naming against the project PSD layer specification, generate one model at the Cubism 5.0 default and open it in the runtime you actually ship to, then read the export diagnostics instead of trusting a clean result.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. README
4. Releases
5. tsunehimatoi/psd2live on GitHub