Skip to main content

Embodied Space Studio [UE Plugin]

Embodied Space Studio is an Unreal Editor workflow panel for authoring embodied motion from text, spatial constraints, saved motion data, and batch prompt collections. It connects the editor to the Kimodo bridge/runtime service, lets you stage prompt timing and body constraints in the Timeline, previews generated motion on either the embedded Preview viewport or the main Level viewport, and turns the selected motion into UE Animation Sequence assets for SOMA, Manny, Quinn, MetaHuman, or a custom retarget target.

The motions were generated using Kimodo, a controllable human motion generation system described in: Davis Rempe et al., "Kimodo: Scaling Controllable Human Motion Generation," arXiv:2603.15546, 2026. https://arxiv.org/abs/2603.15546

Quick Start​

Recommended: AI-assisted setup

For the fastest setup, use an AI coding agent such as Codex, GitHub Copilot, or Claude Code. Give the agent access to the KimodoBridge repository and ask it to configure and verify the local Docker runtime using the Docker Compose commands below. Review all generated commands and confirm local paths, Hugging Face credentials, model files, output paths, and GPU settings before running them.

Tutorial videos​

Download the SOMA Source Kit​

The Fab plugin does not bundle the Kimodo/SOMA source meshes or their third-party license files. Download the SOMA Source Kit separately, extract it to a user-chosen folder, and keep both glTF files together.

Download Embodied Space Studio SOMA Source Kit

The ZIP contains SK_KimodoSOMA.gltf, SK_KimodoSOMA_Rooted.gltf, Kimodo_LICENSE.txt, Kimodo_NOTICE.md, setup instructions, and integrity hashes.

ZIP SHA-256: E15C9CE9B920AAB471039D03F048CB49F00DBCC72D6F4CEECCE210B072BE18CC

Configure the local Docker runtime​

  1. Install and start Docker Desktop with the WSL 2 backend and NVIDIA GPU support where available.

  2. Clone or download the KimodoBridge repository and treat its local root folder as <BridgeRoot>:

    richieBao/embodied-space-studio-kimodo-bridge

  3. Sign in to Hugging Face in a web browser, request access to the gated Meta Llama model, and manually download the following four model packages from their Files and versions pages:

    Model packageManual download pageCopy the downloaded files to
    Kimodo motion checkpoint: nvidia/Kimodo-SOMA-RP-v1.1https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1.1/tree/main<BridgeRoot>/checkpoints/Kimodo-SOMA-RP-v1.1/
    Meta Llama base model: meta-llama/Meta-Llama-3-8B-Instructhttps://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/tree/main<BridgeRoot>/text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/
    LLM2Vec MNTP adapter: McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntphttps://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/tree/main<BridgeRoot>/text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-adapter/
    LLM2Vec supervised adapter: McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervisedhttps://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/tree/main<BridgeRoot>/text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/

    Download the repository files themselves, not only the README page. Preserve the filenames and folder contents. At minimum, verify that the Kimodo folder contains model.safetensors and config.yaml, the Meta Llama folder contains model.safetensors.index.json and all model shard files, and both adapter folders contain adapter_model.safetensors and adapter_config.json.

  4. Copy .env.example to .env. The default TEXT_ENCODER_BASE_MODEL_HOST, TEXT_ENCODER_MNTP_ADAPTER_HOST, and TEXT_ENCODER_ADAPTER_HOST values already match the folders above. Configure the Hugging Face cache/token path, TEXT_ENCODER_DEVICE, UE_BRIDGE_HOST_PORT, and KIMODO_OUTPUT_HOST for the local machine.

  5. Open Docker Desktop, open its built-in Terminal, change directory to the local KimodoBridge root, and build the Compose project:

    cd "<BridgeRoot>"
    docker compose build
    docker compose create text-encoder ue-bridge

  6. Start the required services either from the same Docker Desktop Terminal:

    docker compose up -d text-encoder ue-bridge

    Or, in Docker Desktop, open Containers > kimodo-main and use the existing UI controls:

    • Start/Run to start a created service container.
    • Stop to stop it.
    • Restart to restart it.

    Start text-encoder first, wait until it is healthy, and then start ue-bridge.

  7. Confirm that both containers are running:

    docker compose ps
    docker compose logs --tail=100 ue-bridge

  8. Verify the local bridge endpoints:

    http://127.0.0.1:18027/bridge/v1/health
    http://127.0.0.1:18027/bridge/v1/models

Use the plugin in Unreal Editor​

Required Unreal Engine project setting

Set Edit > Project Settings > Platforms > Windows > Default RHI to DirectX 12, then restart Unreal Editor. Embodied Space Studio requires Default RHI: DirectX 12.

Unreal Engine plugins enabled by ESS​

The Embodied Space Studio plugin descriptor declares and enables these Unreal Engine plugin dependencies:

  • IK Rig (IKRig)
  • Interchange Framework (Interchange)
  • Interchange Editor (InterchangeEditor)
  • Control Rig (ControlRig)
  • Spatially Aware Retarget Ops (RelativeIKOp)
  • MetaHuman Creator (MetaHumanCharacter)
  • MetaHuman Core Tech (MetaHumanCoreTech)

When Embodied Space Studio is enabled, Unreal Editor may request a restart while enabling these dependencies.

  1. Install and enable Embodied Space Studio in Unreal Engine 5.8 and allow its required plugin dependencies to be enabled.
  2. Set Default RHI to DirectX 12 using the project setting above, then restart Unreal Editor.
  3. From the Unreal Editor main menu, select Window > Embodied Space Studio. This opens the dockable Embodied Space Studio editor window.
  4. Select the Generate tab in the right-side tab bar.
  5. Click Connect Docker Bridge, then click Refresh Health. Connection and health messages appear in Bridge Output and Runtime Output.
  6. Click Refresh Models, choose the required Training dataset, Skeleton, and Version, then click Load model.
  7. Download and extract the SOMA Source Kit above. In Generate > Kimodo Mesh / Skeleton, click Choose & Import Kimodo SOMA Mesh / Skeleton, then select SK_KimodoSOMA.gltf. Keep SK_KimodoSOMA_Rooted.gltf beside it; the plugin imports the Mesh/Skeleton into the current Unreal project and remembers the Kit path.
  8. Create or select a prompt section in the Timeline, enter a natural-language motion prompt, and add spatial constraints if required.
  9. Return to Generate and click Generate. The button becomes available after the bridge is healthy, a model is loaded, and valid prompt content exists.
  10. Use Preview Target for embedded preview. For the main Level viewport, click Level Target, then click Spawn SOMA. The spawned SOMA actor can be used for Level preview and for editing constraint poses directly in the Level viewport.
  11. Open Export to bake the selected motion into an Unreal Engine Animation Sequence.
  12. If an action is unavailable, review the status summaries and copyable Bridge Output and Runtime Output logs.
Main workflow
  1. Open Generate, import SK_KimodoSOMA.gltf from the separately downloaded SOMA Source Kit, use Connect Docker Bridge and Refresh Health to connect to the local service, refresh the model catalog, choose dataset/skeleton/version, load a model, and generate motion from Timeline prompts and constraints.
  2. Use Preview Target for fast isolated inspection, or choose Level Target and Spawn SOMA to preview and edit constraint poses directly in the main Level viewport. Use Selected, Focus, and Unlink manage an existing Level Target.
  3. Use the Timeline to create prompt sections, scrub frames, play generated samples, import bulk prompt text, and place Full-Body, 2D Root, hand, and foot constraints.
  4. Use Constraints when a motion needs exact body poses, hand/foot placement, root path control, saved skeleton edit states, or external constraints.json input.
  5. Use Load/Save to reopen saved ue_preview_manifest.json outputs, save reusable example metadata, and open generated output folders.
  6. Use Visualize to control playback speed, mesh/skeleton/foot-contact overlays, floor/environment display, camera FOV, Timeline visibility, and constraint label display.
  7. Use Export to bake the active motion as a SOMA Animation Sequence, build or refresh rooted SOMA retarget assets, select target character assets, tune collision/foot anchoring options, and export a target Animation Sequence.
  8. Use Batch to process many prompts or already-generated SOMA jobs, export multiple target families, preview generated naming, and resume or stop a batch run.
  9. Use Misc for production utilities: placing batch animations into a level, creating a MetaHuman player character, merging MetaHuman body/garment meshes, packaging a Fab product, offline collision/grounding cleanup, MetaHuman Level Sequences, and prompt subtitles.

When a button is disabled, check the status summaries above the active section. Most actions require a healthy bridge service, a loaded model, a valid preview clip, a matching source/target mesh, a selected Level Target, or a valid /Game output path. Runtime Output and Bridge Output logs are designed for troubleshooting and can be copied from the UI.

ess

Kimodo Docker Runtime Setup​

AI-assisted setup

You can use AI coding agents such as Codex, GitHub Copilot, or Claude Code to help configure the Kimodo Docker runtime automatically by following the instructions in this section. Give the agent access to the bridge repository and ask it to complete and verify the documented setup steps. Before running the runtime, review any generated commands and confirm that local paths, credentials, model files, and GPU settings are correct for your machine.

Embodied Space Studio uses a local Docker Compose runtime to serve Kimodo motion generation to Unreal Engine. The Unreal Engine plugin is distributed through Fab, while the Docker bridge runtime is distributed separately through GitHub:

https://github.com/richieBao/embodied-space-studio-kimodo-bridge

In this bridge repository, kimodo-main is the Docker Compose project name declared in docker-compose.yaml. The always-on services are text-encoder and ue-bridge; demo is optional for the standalone Kimodo web demo.

The Docker bridge is self-contained. It can be placed in any local folder and does not need to live inside or beside an Unreal project. Any Unreal project with the Embodied Space Studio / KimodoMotionAuthoring plugin can connect to the same local bridge endpoint:

http://127.0.0.1:18027/bridge/v1

ess

Install order​

  1. Buy/download and install the Embodied Space Studio plugin from Fab into any Unreal project that will use Kimodo generation.
  2. Download and extract the SOMA Source Kit from this page. Keep SK_KimodoSOMA.gltf and SK_KimodoSOMA_Rooted.gltf in the same folder.
  3. Install and start Docker Desktop.
  4. Download or clone the GitHub bridge repository to a user-chosen folder.
  5. Sign in to Hugging Face in a browser and request access to the gated meta-llama/Meta-Llama-3-8B-Instruct repository.
  6. Manually download the four model packages listed in Download model files and copy their files into the exact checkpoints/ and text-encoders/ folders under <BridgeRoot>.
  7. Copy .env.example to .env, then edit paths for your machine.
  8. Build/start the Docker bridge and connect to http://127.0.0.1:18027/bridge/v1 from the plugin's Generate tab.
  9. In Generate > Kimodo Mesh / Skeleton, choose and import SK_KimodoSOMA.gltf before previewing or exporting SOMA motion.

Prerequisite: install Docker Desktop​

Docker is required before you can start the Kimodo runtime. Embodied Space Studio does not install Docker automatically.

  1. Install WSL for the Docker Desktop WSL 2 backend. Open PowerShell as Administrator and run:

    wsl --install --no-distribution
    wsl --update

    The --no-distribution option installs WSL without installing an additional Linux distribution such as Ubuntu. Docker Desktop manages the WSL distributions it requires. If the installation prompts you to restart, restart your computer.

  2. Verify the installed WSL version:

    wsl --version

    Docker Desktop requires WSL version 2.1.5 or later; using the latest available WSL version is recommended.

  3. Install Docker Desktop for Windows from the official Docker website.

  4. Start Docker Desktop and wait until the Docker Engine is running. In Settings > General, enable Use WSL 2 based engine. On supported systems, this option may already be enabled by default and may not be visible.

  5. For GPU generation, install a compatible NVIDIA driver and make sure Docker Desktop can access the GPU.

  6. Open PowerShell and confirm Docker is available:

docker --version
docker compose version
note

Docker is required for Generate, Refresh Models, Connect Docker Bridge, and other Kimodo runtime features. Loading already generated previews or using some offline Unreal-only utilities may not require the Docker runtime to be running.

Python environment note​

The official Kimodo local virtual-environment setup uses Python 3.10:

conda create -n kimodo python=3.10
conda activate kimodo

Embodied Space Studio normally runs Kimodo through Docker, so users do not need to create this local Python environment unless they want to run Kimodo scripts outside the containers.

Download the bridge repository​

Choose any local folder for the bridge. The commands below use <BridgeRoot> as the folder where you cloned or extracted the GitHub repository.

git clone https://github.com/richieBao/embodied-space-studio-kimodo-bridge.git
cd embodied-space-studio-kimodo-bridge

Downloading the repository ZIP from GitHub and extracting it to a folder such as D:/EmbodiedSpaceStudio/KimodoBridge is also fine.

The bridge folder should contain:

<BridgeRoot>/docker-compose.yaml
<BridgeRoot>/Dockerfile
<BridgeRoot>/ue_bridge_service
<BridgeRoot>/scripts
<BridgeRoot>/checkpoints
<BridgeRoot>/text-encoders

The ue-bridge container runs the bridge API service packaged in <BridgeRoot>/ue_bridge_service; it no longer mounts a specific Unreal project or runs a server.py from the Fab plugin folder.

Runtime layout​

ServiceContainer namePurposeHost access
text-encodertext-encoderLoads the LLM2Vec text encoder and provides embeddings to Kimodo services.http://127.0.0.1:9550
ue-bridgeue-bridgeRuns the bridge API used by the plugin Generate / Models / Health actions.http://127.0.0.1:18027/bridge/v1
demodemoOptional Kimodo interactive demo service.http://127.0.0.1:7860 by default

The services are built from Dockerfile into the kimodo:1.0 image. The image is based on NVIDIA PyTorch, installs the Kimodo Python runtime, includes the local kimodo-viser fork, and runs through the container entrypoint that maps files back to the host user.

Download model files​

Model weights are intentionally not included in the GitHub repository. The user must sign in to Hugging Face, accept or request any required model access, manually download the files from each Files and versions page, and copy them into the exact <BridgeRoot> folders below. Review the license shown on each model page before downloading or using its files.

1. Kimodo motion checkpoint​

  • Model: nvidia/Kimodo-SOMA-RP-v1.1

  • Manual download: https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1.1/tree/main

  • Copy all downloaded repository files to:

    <BridgeRoot>/checkpoints/Kimodo-SOMA-RP-v1.1/

  • Required files include:

    <BridgeRoot>/checkpoints/Kimodo-SOMA-RP-v1.1/model.safetensors
    <BridgeRoot>/checkpoints/Kimodo-SOMA-RP-v1.1/config.yaml
    <BridgeRoot>/checkpoints/Kimodo-SOMA-RP-v1.1/stats/

2. Meta Llama 3 base model​

  • Model: meta-llama/Meta-Llama-3-8B-Instruct

  • Manual download: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/tree/main

  • This repository is gated. Request access and wait for approval before downloading.

  • Copy all downloaded repository files to:

    <BridgeRoot>/text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/

  • Despite the local folder name ending in -mntp, this folder must contain the full raw Meta Llama base model. Verify that it includes model.safetensors.index.json, every referenced model shard, tokenizer files, and configuration files.

3. LLM2Vec MNTP adapter​

  • Model: McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp

  • Manual download: https://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/tree/main

  • Copy all downloaded repository files to:

    <BridgeRoot>/text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-adapter/

  • Verify that this folder includes adapter_model.safetensors and adapter_config.json.

  • There is no separate Hugging Face repository named ...-mntp-adapter; -adapter is only the required local folder name.

4. LLM2Vec supervised adapter​

The completed manual folder layout should resemble:

<BridgeRoot>/
├─ checkpoints/
│  └─ Kimodo-SOMA-RP-v1.1/
│     ├─ model.safetensors
│     ├─ config.yaml
│     └─ stats/
└─ text-encoders/
   └─ McGill-NLP/
      ├─ LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/
      │  ├─ model.safetensors.index.json
      │  ├─ model-*.safetensors
      │  └─ tokenizer/configuration files
      ├─ LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-adapter/
      │  ├─ adapter_model.safetensors
      │  └─ adapter_config.json
      └─ LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/
         ├─ adapter_model.safetensors
         └─ adapter_config.json

Runtime loading order:

  1. Load raw meta-llama/Meta-Llama-3-8B-Instruct files from the local LLM2Vec-Meta-Llama-3-8B-Instruct-mntp folder as the base model.
  2. Merge the MNTP adapter from TEXT_ENCODER_MNTP_ADAPTER_HOST; by default this is the local LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-adapter folder downloaded from McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp.
  3. Apply McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised as the final supervised adapter exposed by the text-encoder service.

Configure .env​

From <BridgeRoot>, copy the example file:

Copy-Item .env.example .env

Then edit .env for your machine. Use forward slashes in Windows paths for Docker volume mapping.

HF_HOME_HOST=./.cache/huggingface
HOST_USER=<your-windows-user-name>
SERVER_PORT=7860
TEXT_ENCODER_DEVICE=auto
TEXT_ENCODER_BASE_MODEL_HOST=./text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp
TEXT_ENCODER_MNTP_ADAPTER_HOST=./text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-adapter
TEXT_ENCODER_ADAPTER_HOST=./text-encoders/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised
UE_BRIDGE_HOST_PORT=18027
KIMODO_OUTPUT_HOST=D:/EmbodiedSpaceStudio/KimodoBridge/outputs

HF_HOME_HOST sets the cache directory mounted into the containers. With the documented manual-download workflow, the required model weights are loaded from the local checkpoints/ and text-encoders/ folders, so a Hugging Face CLI download is not required. If another workflow needs authenticated Hugging Face access at runtime, place the approved account token in the selected cache directory.

KIMODO_OUTPUT_HOST is the host folder where generated motion.npz, ue_preview_manifest.json, logs, diagnostics, and prompt metadata are written. The packaged start script resolves KIMODO_OUTPUT_HOST=./outputs to an absolute bridge-local path automatically. If you run raw docker compose commands, set KIMODO_OUTPUT_HOST to an absolute Windows path so Unreal can open generated results.

TEXT_ENCODER_DEVICE is a Docker startup-level setting for the text-encoder service:

TEXT_ENCODER_DEVICE=auto
TEXT_ENCODER_DEVICE=cuda
TEXT_ENCODER_DEVICE=cpu

Use auto for CUDA when available with CPU fallback, cuda to force GPU, or cpu to force CPU text encoding. Use cuda, not gpu; Kimodo passes this value to PyTorch device selection.

Start the Unreal bridge runtime​

Open Docker Desktop, open its built-in Terminal, and change directory to the local KimodoBridge repository:

cd "<BridgeRoot>"

Build the image and create the Compose-configured containers:

docker compose build
docker compose create text-encoder ue-bridge

Start the required services from the Docker Desktop Terminal:

docker compose up -d text-encoder ue-bridge

Alternatively, after docker compose create has created the services, use the existing Docker Desktop UI controls:

  • Open Containers > kimodo-main.
  • Use Start/Run to start an already-created service container.
  • Use Stop to stop it.
  • Use Restart to restart it.
  • Start text-encoder first, wait until its status is healthy, and then start ue-bridge.

Connect from Unreal Engine​

Open any Unreal project that has the Embodied Space Studio / KimodoMotionAuthoring plugin installed. In the plugin's Docker bridge settings or Generate tab, connect to:

http://127.0.0.1:18027/bridge/v1

The Docker bridge is independent from the UE project folder. Multiple compatible UE projects can use the same local bridge by connecting to the same endpoint.

Apply .env changes and recreate containers​

Docker reads environment variables when a container is created. If you edit .env, especially TEXT_ENCODER_DEVICE, model paths, ports, or KIMODO_OUTPUT_HOST, simply clicking Stop/Start or Restart in Docker Desktop can keep the old container configuration and the old value.

After changing .env, run:

cd "<BridgeRoot>"
docker compose up -d --force-recreate --no-build text-encoder ue-bridge

Use --build only when the Docker image itself needs to be rebuilt, such as after changing Dockerfile, docker_requirements.txt, ue_bridge_service, or dependency files:

cd "<BridgeRoot>"
docker compose up -d --build --force-recreate text-encoder ue-bridge

Verify the active device setting inside the recreated container:

docker exec text-encoder python -c "import os; print(os.environ.get('TEXT_ENCODER_DEVICE'))"

Verify the runtime​

Run the packaged verification script:

cd "<BridgeRoot>"
.\scripts\Verify-EmbodiedSpaceStudio-Bridge.ps1

A healthy bridge should respond at:

http://127.0.0.1:18027/bridge/v1/health
http://127.0.0.1:18027/bridge/v1/models

The expected model key for this packaged setup is:

kimodo-soma-rp-v1.1

You can also inspect the containers directly:

docker ps --filter "name=text-encoder"
docker ps --filter "name=ue-bridge"
cd "<BridgeRoot>"
docker compose logs -f ue-bridge

Stop or reset the runtime​

To stop the services while keeping containers available for a faster restart:

cd "<BridgeRoot>"
.\scripts\Stop-EmbodiedSpaceStudio-Bridge.ps1

To remove the compose containers completely:

.\scripts\Stop-EmbodiedSpaceStudio-Bridge.ps1 -RemoveContainers
tip

If Unreal Engine cannot connect from the Generate tab, first check http://127.0.0.1:18027/bridge/v1/health, then confirm that Docker Desktop is running and that text-encoder and ue-bridge are active. The Unreal plugin's Connect Docker Bridge button is intended to connect to the same local bridge endpoint expected by these scripts.

User Manual​

Purpose​

Connect Unreal Editor to the Kimodo runtime, choose the active generation model and source character, configure sampling/postprocess settings, and submit the current Timeline prompts plus native or external constraints for motion generation.

Where to find these controls​

From the Unreal Editor main menu, choose Window > Embodied Space Studio. In the dockable Embodied Space Studio window, use the right-side control panel. Its horizontal tab row is directly below Bridge Output; select Generate. The Generate action buttons are at the top of the scrollable tab content, below the Source Character summary. Scroll down in the same right-side panel to reach Runtime Output, Model Selection, Kimodo Mesh / Skeleton, and Generate Settings.

Bridge controls versus motion generation

In the Generate tab, Connect Docker Bridge starts/connects the local bridge service and Refresh Health checks its status. The green Generate button does not start the bridge; it submits motion generation only after the bridge is healthy, a model is loaded, and the Timeline contains a valid prompt.

ess

Action Buttons​

  1. Generate / Generating submits the current request. The button is enabled only when the bridge is healthy, a model is loaded, there is valid prompt content, and the service is not busy.
  2. Stop Compute asks the bridge/runtime to stop the active compute job without resetting all editor state.
  3. Connect Docker Bridge and Disconnect start or stop the local bridge service. Refresh Health rechecks service health, and Reset clears editor authoring state and preview state.

Runtime Output​

  1. Output Name Prefix is used when naming generated job outputs and follow-up assets.
  2. Text Encoder Device is configured in Docker, not per Generate request. Edit TEXT_ENCODER_DEVICE in KimodoBridge/.env, then recreate text-encoder and ue-bridge with docker compose up -d --force-recreate --no-build text-encoder ue-bridge.
  3. Runtime Log records user-facing Generate/Export/Batch actions. Copy Log sends it to the clipboard; Clear removes the local display text.

Model Selection​

  1. Refresh Models queries the bridge model catalog.
  2. Training dataset, Skeleton, and Version filter the model list. Load model makes that model active in the current bridge session.
  3. The selected model summary explains whether the model is loaded, unavailable, or waiting on service health.

Kimodo Mesh / Skeleton​

  1. Choose & Import Kimodo SOMA Mesh / Skeleton opens a file picker. Select SK_KimodoSOMA.gltf from the separately downloaded SOMA Source Kit. The plugin validates that the source contains Hips and does not contain the synthetic root bone, imports the Mesh/Skeleton into /Game/EmbodiedSpaceStudio/Kimodo/SOMA, and remembers the external Kit path for Rooted SOMA generation.
  2. The SOMA Source Kit status line shows the currently selected external glTF path. Keep SK_KimodoSOMA_Rooted.gltf in the same folder as the selected original file.
  3. Kimodo Source Preview Mesh overrides the source mesh used for preview binding and request metadata.
  4. Source Skeleton Override can provide a skeleton when a mesh is not enough or when a specific binding is required.
  5. Reset To SOMA Default clears custom source selections and returns to the already imported Kimodo/SOMA source asset. If the asset is not installed, it prompts you to use the external Source Kit import button; it no longer imports a bundled fallback.

Generate Settings​

  1. Num Samples controls how many candidates the runtime returns for the same authored prompt/constraint request.
  2. Seed becomes active when edited and is used for reproducible generation.
  3. Denoising Steps controls diffusion refinement cost and quality.
  4. Reduce Foot Sliding enables runtime foot-contact stabilization. Level 1 is lighter; Level 2 adds stable curves and XY foot lock.
  5. Classifier-Free Guidance balances Text Weight and Constraint Weight when CFG is enabled.
  6. Transition Frames controls blending between adjacent prompt segments.
  7. Post Processing toggles runtime postprocess, and Root Margin adjusts the root-trajectory margin used by that postprocess.

First author prompts in the Timeline, then add constraints if needed, confirm the bridge health and loaded model, generate one sample set, and inspect it in Preview or Level Target before exporting or batching.

Example Prompts​

Download MetaHuman Living NPC Pack Vol.1 Example Prompts

View the Generated Animation Pack on Fab

MetaHuman Lexi MHPKG File Download​

Download Lexi