Docs

Technical Documentation · MotionSmith AI v1.0.0

From Prompt to In-Engine Animation Pipeline

Complete reference guide for setup, model configuration, motion generation, multi-segment authoring, animation continuation, editing, retargeting and Unreal Engine 5.8 output.

1. Quick Start Guide 2. System & Hardware Requirements 3. Installation & Setup 4. Model & Weights Configuration 5. Backend & Pre-Flight Diagnostics 6. Studio Interface Overview 7. Generation Settings 8. Recipes System 9. Prompting Guide & Syntax 10. Multi-Prompt Timeline & Chaining 11. Import & Continue Workflow 12. In-Studio Motion Editing 13. Props & Spatial Context 14. Skeleton Retargeting 15. Animation Output 16. Troubleshooting & FAQ 17. Third-Party Licenses & Credits 18. Technical Support

1. Quick Start Guide

MotionSmith AI provides a local motion-generation and animation-authoring workflow inside Unreal Engine 5.8. Describe a motion, generate and review the result, extend it with additional segments, refine the performance, retarget it to a supported character workflow and create a standard Unreal animation asset.

Prompt → Generate → Review → Continue → Retarget → AnimationSequence

MotionSmith AI Studio — generation, preview, timeline authoring and retargeting inside Unreal Engine.

First-Time Setup

C:\Program Files\Epic Games\UE_5.8\Engine\Plugins\Marketplace

1. Install and enable MotionSmith AI Plugin for Unreal Engine 5.8 (C:\Program Files\Epic Games\UE_5.8\Engine\Plugins\Marketplace)
2. Download the separate MotionSmith Runtime package.
3. Extract the Runtime contents to %LOCALAPPDATA%\MSR\.
4. Obtain Meta Llama 3 8B Instruct separately from the official Hugging Face repository.
5. Open Editor Preferences → Plugins → MotionSmith.
6. Select your local Meta Llama 3 8B Instruct model folder.
7. Click Rescan and confirm that the required MotionSmith components report READY.
8. Run Diagnostics / Pre-Flight and confirm all required checks pass.

Create Your First Motion

1. Open MotionSmith AI Studio.
2. Enter a short, focused motion prompt.
3. Generate and review the motion in the viewport.
4. Use Add Segment to continue the performance without rebuilding already approved motion.
5. Retarget to your target character when required.
6. Create the final Unreal AnimationSequence.

Existing animations can also be imported into the MotionSmith workflow, adjusted, placed on the timeline and continued with newly generated motion.

Installation & First-Time Setup

MotionSmith AI Animation Studio · Unreal Engine 5.8

Video Tutorial

Install MotionSmith, configure the Runtime and Meta Llama model, run Pre-Flight diagnostics, and generate your first animation.

Watch on YouTube ↗

2. System & Hardware Requirements

MotionSmith AI runs its generation workflow locally. Actual memory usage depends on the current Unreal project, viewport workload, runtime configuration and model state, so MotionSmith 1.0.0 does not publish a fixed minimum VRAM figure.

Component	Requirement
Engine	Unreal Engine 5.8
Operating System	Windows 64-bit
GPU	CUDA-capable NVIDIA GPU
Model	Meta Llama 3 8B Instruct stored locally
Storage	Sufficient local SSD space for the MotionSmith Runtime, Meta Llama model and project assets

No Manual Python Environment Required

The separate MotionSmith Runtime package includes the Python environment, PyTorch/CUDA runtime, NVIDIA Kimodo-SOMA-RP-v1.1 model and the required LLM2Vec adapters. You do not need to create a Conda environment or manually install Python, PyTorch, Kimodo or LLM2Vec.

Meta Llama 3 8B Instruct is not included with MotionSmith or the MotionSmith Runtime and must be obtained separately by the user.

If GPU memory becomes constrained, reduce unnecessary GPU-heavy workloads in Unreal while generating and close unrelated GPU-intensive applications.

3. Installation & Setup

Install the MotionSmith Plugin

1. Install MotionSmith AI for Unreal Engine 5.8 through Fab.
2. Open the target Unreal project.
3. Enable the plugin if Unreal asks you to do so.
4. Restart Unreal Editor when required.

Install MotionSmith Runtime

MotionSmith Runtime is distributed separately from the Fab plugin. MotionSmith does not automatically download, install or extract the Runtime.

Download MotionSmith Runtime 1.0.0

1. Download the MotionSmith Runtime ZIP.
2. Open %LOCALAPPDATA% in Windows Explorer.
3. Create an MSR folder if it does not already exist.
4. Extract the contents of the Runtime package into %LOCALAPPDATA%\MSR\.

The Runtime package already includes the required Python environment, PyTorch/CUDA runtime, NVIDIA Kimodo-SOMA-RP-v1.1 model and LLM2Vec adapters.

Configure Meta Llama 3 8B Instruct

1. Obtain access to the official Meta Llama 3 8B Instruct repository on Hugging Face.
2. Download the model to any local folder you prefer.
3. Open Editor Preferences → Plugins → MotionSmith.
4. Set Meta Llama Folder to the downloaded model folder.
5. Click Rescan.

Verify the Setup

The MotionSmith Setup section should report READY for:

• MotionSmith Runtime
• Kimodo-SOMA-RP-v1.1
• LLM2Vec MNTP Adapter
• LLM2Vec Supervised Adapter
• Meta Llama 3 8B Instruct

When all required components are detected, the Backend Setup Summary should report READY.

Run Diagnostics / Pre-Flight and confirm that the release gate reports READY before generating motion.

4. Model & Weights Configuration

MotionSmith Runtime already includes the NVIDIA Kimodo-SOMA-RP-v1.1 model and the required LLM2Vec adapters.

The only model that must be obtained separately by the user is Meta Llama 3 8B Instruct. MotionSmith does not redistribute the Meta Llama model weights.

Download Meta Llama 3 8B Instruct

Obtain access and download the model from the official Hugging Face repository in accordance with Meta's applicable license terms:

Meta Llama 3 8B Instruct on Hugging Face

You may store the downloaded Meta Llama model anywhere on your computer.

Select the Local Model Folder

Open Editor Preferences → Plugins → MotionSmith and set Meta Llama Folder to the downloaded model directory.

A valid model directory contains files such as:

Meta-Llama-3-8B-Instruct/
├── config.json
├── tokenizer_config.json
├── model.safetensors.index.json
├── special_tokens_map.json
├── tokenizer.json
├── model-00001-of-00004.safetensors
├── mode2-00001-of-00004.safetensors
├── mode3-00001-of-00004.safetensors
├── mode4-00001-of-00004.safetensors
└── supporting model configuration files

Click Rescan. When the model is valid, Meta Llama 3 8B Instruct will report READY.

Alternative Download Method — Hugging Face CLI

As an alternative to downloading the model files manually, you can use the official Hugging Face CLI to download the required Meta Llama 3 8B Instruct files directly into a local folder.

First, make sure your Hugging Face account has access to Meta Llama 3 8B Instruct. Then install the Hugging Face Hub CLI:

py -m pip install -U huggingface_hub

Sign in to Hugging Face:

hf auth login

Download the required model files to a folder of your choice:

hf download meta-llama/Meta-Llama-3-8B-Instruct config.json generation_config.json model-00001-of-00004.safetensors model-00002-of-00004.safetensors model-00003-of-00004.safetensors model-00004-of-00004.safetensors model.safetensors.index.json special_tokens_map.json tokenizer.json tokenizer_config.json --local-dir "C:\MotionSmith\Models\Meta-Llama-3-8B-Instruct"

When the download is complete, set Meta Llama Folder in Editor Preferences → Plugins → MotionSmith to the folder you selected, then click Rescan.

5. Backend & Pre-Flight Diagnostics

MotionSmith runs a local backend for motion generation. Setup and validation are available from Editor Preferences → Plugins → MotionSmith.

Setup Areas

• MotionSmith Setup: verifies the installed Runtime, bundled Kimodo model, LLM2Vec adapters and selected Meta Llama folder.
• Environment: validates the local Python, GPU and CUDA environment.
• Lifecycle: starts, stops or restarts the local backend when required.
• Diagnostics / Pre-Flight: performs the final environment and configuration checks before generation.

Run Diagnostics / Pre-Flight after first-time setup and after changing important runtime or model configuration. Required checks should pass before starting production generation.

If configuration has changed while the backend is already running, restart the backend so the new settings are loaded.

6. Studio Interface Overview

MotionSmith AI Studio brings generation, preview, timeline authoring, motion adjustments and retargeting into one Unreal Editor workflow.

• Authoring Viewport: preview the current character motion and spatial relationship to optional props.
• Prompt & Generate Controls: describe the next action and start generation.
• Timeline: review and manage generated or imported motion segments as a continuous performance.
• Details Panel: access generation, motion editing and retargeting controls relevant to the current workflow.
• Retargeting: transfer the working motion to a supported target character workflow without leaving MotionSmith.

7. Generation Settings

MotionSmith exposes the generation controls required for the current build directly in the Studio / Details workflow. Use the visible controls rather than modifying Kimodo or Python configuration manually.

Recommended Workflow

1. Write a focused motion prompt.
2. Select an appropriate Recipe when the current task benefits from one.
3. Review the current generation defaults in the Details panel.
4. Generate the motion.
5. Judge the result in the viewport before changing multiple settings at once.

For predictable iteration, change one meaningful setting at a time. If the result does not match the intended action, simplify or clarify the prompt before making large parameter changes.

8. Recipes System

Recipes provide convenient starting configurations for different motion-generation tasks. They are workflow presets, not replacements for a clear prompt.

Using Recipes

1. Select the recipe that best matches the type of motion you want to create.
2. Write a concise physical-action prompt.
3. Generate and review the result.
4. If the motion intent is wrong, revise the prompt first.
5. If necessary, test another recipe while keeping the prompt focused.

Recipe availability and exact labels are determined by the MotionSmith 1.0.0 build shown in the Studio interface.

9. Prompting Guide & Best Practices

Core Prompting Rules

• Describe motion, not story context: write what the body should physically do.
• Keep prompts focused: one clear action or beat usually produces easier-to-control results.
• Use direction and pace when useful: forward, backward, left, right, slowly, quickly, cautiously, etc.
• Use the timeline for choreography: split longer performances into multiple segments.
• English is recommended: concise English motion descriptions generally provide the most predictable text intent.

Less Useful	Better Motion Prompt
"The soldier is terrified and searches for an escape."	walk forward cautiously while looking from side to side
"Do an awesome fight move."	throw a right high kick and recover to a fighting stance
"Walk around happily."	walk forward with a light energetic step

Build Complex Actions as Segments

idle → walk forward → stop → turn left → step back

10. Multi-Prompt Timeline & Chaining

MotionSmith treats a longer performance as a sequence of editable motion segments. Each prompt can become its own clip on the timeline.

SEGMENT 01  idle and scan the room
SEGMENT 02  walk forward cautiously
SEGMENT 03  stop and turn toward the sound

Add Segment

Add Segment continues from the existing motion tail instead of regenerating the entire approved performance. MotionSmith preserves the earlier animation and uses the previous segment's transition context to generate the next motion.

This makes it practical to approve a section, continue from it, and revise the next beat without discarding earlier work.

11. Import & Continue Workflow

Existing Unreal animation can become part of the same MotionSmith working timeline rather than serving only as an external reference.

Imported animation and generated motion can be combined in the same MotionSmith timeline.

Imported Animation
↓
Transform / Adjust
↓
Add to Timeline
↓
Generate Next Segment
↓
Review Combined Motion
↓
Create AnimationSequence

This workflow makes it possible to combine authored or captured motion with AI-generated continuation and review the result as one performance.

12. In-Studio Motion Editing

MotionSmith includes practical tools for refining working motion without regenerating everything from scratch.

• Timeline Editing: work with individual motion segments and their placement in the performance.
• Transform Adjustments: reposition or reorient imported and working motion when required.
• Pose Controls: adjust a pose and hold it when the workflow requires a fixed pose decision.
• Mirror Pose: mirror a pose across the character's local sagittal plane.
• Mirror Animation: create a mirrored animation variation while preserving the intended trajectory behavior.
• Root / Path Authoring: refine character travel and root-path decisions using the controls available for the current clip.

The goal is to keep approved animation intact and make targeted corrections instead of returning to full regeneration for every change.

13. Props & Spatial Context

Props can be loaded into the MotionSmith authoring viewport as spatial references while building or reviewing character motion.

Use props as spatial references while authoring and reviewing character motion.

1. Load the prop into the MotionSmith viewport.
2. Position it relative to the character.
3. Generate, import or edit the character motion.
4. Review the spatial relationship in the viewport.
5. Adjust the motion or prop placement as needed.

Props provide authoring context. MotionSmith's primary output remains the character animation; the presence of a prop does not imply automatic hand attachment or guaranteed physical contact solving.

14. Skeleton Retargeting

MotionSmith works with motion generated on the NVIDIA Kimodo SOMA character representation and transfers the finished motion to target Unreal characters through Unreal Engine retargeting workflows.

Retarget generated motion to the target Unreal character without leaving the MotionSmith workflow.

Included Target Workflows

• Epic / Manny-based characters
• MetaHuman
• Character Creator
• Mixamo-based characters

Typical Retarget Workflow

1. Generate or import and approve the source motion.
2. Open the MotionSmith retargeting controls.
3. Select the appropriate target workflow.
4. Select the target character / skeletal mesh.
5. Run the retarget operation.
6. Review the result on the target character before final output.

Characters outside the included profiles may require a separate Unreal IK Rig / IK Retargeter setup and validation.

15. Animation Output

MotionSmith creates standard Unreal Engine animation output so the finished performance can continue through normal production tools.

MotionSmith → AnimationSequence → Unreal Production Pipeline

The resulting animation can be used with standard Unreal workflows such as:

• Sequencer
• Animation editing tools
• Animation Blueprints
• Montages and state-machine workflows
• Further Unreal retargeting or cinematic production

Multiple generated segments, or imported and generated segments, can be reviewed as one continuous performance before creating the final asset.

16. Troubleshooting & FAQ

MotionSmith Runtime Is Not Detected

Confirm that the Runtime was extracted to %LOCALAPPDATA%\MSR\ and that %LOCALAPPDATA%\MSR\runtime.json exists. Then return to Editor Preferences → Plugins → MotionSmith and click Rescan.

Meta Llama 3 Model Is Not Detected

Make sure Meta Llama Folder points to the model root containing the configuration, tokenizer and weight files. Then click Rescan and run Diagnostics / Pre-Flight again.

Backend Is Not Ready

Open Editor Preferences → Plugins → MotionSmith and confirm that the Runtime, Kimodo model, both LLM2Vec adapters and Meta Llama all report READY. Then run Diagnostics / Pre-Flight.

Configuration Changed but the Backend Uses the Old State

Restart the MotionSmith backend from its Lifecycle controls so the updated runtime or model configuration is loaded.

Retargeted Motion Looks Incorrect

Confirm that the selected MotionSmith target workflow matches the target skeleton and that the correct skeletal mesh is selected. For custom setups, validate the Unreal IK Rig / IK Retargeter independently.

The Generated Motion Does Not Match the Prompt

Simplify the prompt and describe one physical action at a time. Split longer choreography into separate timeline segments.

The Next Segment Needs a Different Direction

Keep the approved earlier motion and revise the next segment prompt rather than rebuilding the entire performance.

GPU Memory Is Constrained

Reduce unnecessary GPU-heavy editor workloads while generating and close unrelated GPU-intensive applications. Memory use varies with the current Unreal project and runtime state.

17. Third-Party Licenses & Credits

NVIDIA Kimodo

MotionSmith AI is powered by the open-source NVIDIA Kimodo motion-generation project. MotionSmith AI is an independent product and is not affiliated with or endorsed by NVIDIA Corporation.

NVIDIA Kimodo Documentation

Meta Llama 3

Built with Meta Llama 3. MotionSmith AI does not distribute the Meta Llama 3 model weights. Users are responsible for obtaining and using the model in accordance with Meta's applicable license terms.

Meta Llama 3 8B Instruct

18. Technical Support

For bug reports, setup problems or MotionSmith AI product support:

📧 Support: motionsmithai@gmail.com

When reporting a reproducible issue, include:

• Unreal Engine version
• MotionSmith AI version
• GPU model
• The failed Diagnostics / Pre-Flight item, if any
• A relevant Unreal log excerpt
• Short steps that reproduce the problem

Technical Documentation · MotionSmith AI v1.0.0

From Prompt to In-Engine Animation Pipeline

Complete reference guide for setup, model configuration, motion generation, multi-segment authoring, animation continuation, editing, retargeting and Unreal Engine 5.8 output.

1. Quick Start Guide 2. System & Hardware Requirements 3. Installation & Setup 4. Model & Weights Configuration 5. Backend & Pre-Flight Diagnostics 6. Studio Interface Overview 7. Generation Settings 8. Recipes System 9. Prompting Guide & Syntax 10. Multi-Prompt Timeline & Chaining 11. Import & Continue Workflow 12. In-Studio Motion Editing 13. Props & Spatial Context 14. Skeleton Retargeting 15. Animation Output 16. Troubleshooting & FAQ 17. Third-Party Licenses & Credits 18. Technical Support

1. Quick Start Guide

MotionSmith AI provides a local motion-generation and animation-authoring workflow inside Unreal Engine 5.8. Describe a motion, generate and review the result, extend it with additional segments, refine the performance, retarget it to a supported character workflow and create a standard Unreal animation asset.

Prompt → Generate → Review → Continue → Retarget → AnimationSequence

MotionSmith AI Studio — generation, preview, timeline authoring and retargeting inside Unreal Engine.

First-Time Setup

1. Install and enable MotionSmith AI for Unreal Engine 5.8.
2. Download the separate MotionSmith Runtime package.
3. Extract the Runtime contents to %LOCALAPPDATA%\MSR\.
4. Obtain Meta Llama 3 8B Instruct separately from the official Hugging Face repository.
5. Open Editor Preferences → Plugins → MotionSmith.
6. Select your local Meta Llama 3 8B Instruct model folder.
7. Click Rescan and confirm that the required MotionSmith components report READY.
8. Run Diagnostics / Pre-Flight and confirm all required checks pass.

Create Your First Motion

1. Open MotionSmith AI Studio.
2. Enter a short, focused motion prompt.
3. Generate and review the motion in the viewport.
4. Use Add Segment to continue the performance without rebuilding already approved motion.
5. Retarget to your target character when required.
6. Create the final Unreal AnimationSequence.

Existing animations can also be imported into the MotionSmith workflow, adjusted, placed on the timeline and continued with newly generated motion.

2. System & Hardware Requirements

MotionSmith AI runs its generation workflow locally. Actual memory usage depends on the current Unreal project, viewport workload, runtime configuration and model state, so MotionSmith 1.0.0 does not publish a fixed minimum VRAM figure.

Component	Requirement
Engine	Unreal Engine 5.8
Operating System	Windows 64-bit
GPU	CUDA-capable NVIDIA GPU
Model	Meta Llama 3 8B Instruct stored locally
Storage	Sufficient local SSD space for the MotionSmith Runtime, Meta Llama model and project assets

No Manual Python Environment Required

The separate MotionSmith Runtime package includes the Python environment, PyTorch/CUDA runtime, NVIDIA Kimodo-SOMA-RP-v1.1 model and the required LLM2Vec adapters. You do not need to create a Conda environment or manually install Python, PyTorch, Kimodo or LLM2Vec.

Meta Llama 3 8B Instruct is not included with MotionSmith or the MotionSmith Runtime and must be obtained separately by the user.

If GPU memory becomes constrained, reduce unnecessary GPU-heavy workloads in Unreal while generating and close unrelated GPU-intensive applications.

3. Installation & Setup

Install the MotionSmith Plugin

1. Install MotionSmith AI for Unreal Engine 5.8 through Fab.
2. Open the target Unreal project.
3. Enable the plugin if Unreal asks you to do so.
4. Restart Unreal Editor when required.

Install MotionSmith Runtime

MotionSmith Runtime is distributed separately from the Fab plugin. MotionSmith does not automatically download, install or extract the Runtime.

Download MotionSmith Runtime 1.0.0

1. Download the MotionSmith Runtime ZIP.
2. Open %LOCALAPPDATA% in Windows Explorer.
3. Create an MSR folder if it does not already exist.
4. Extract the contents of the Runtime package into %LOCALAPPDATA%\MSR\.

The Runtime package already includes the required Python environment, PyTorch/CUDA runtime, NVIDIA Kimodo-SOMA-RP-v1.1 model and LLM2Vec adapters.

Configure Meta Llama 3 8B Instruct

1. Obtain access to the official Meta Llama 3 8B Instruct repository on Hugging Face.
2. Download the model to any local folder you prefer.
3. Open Editor Preferences → Plugins → MotionSmith.
4. Set Meta Llama Folder to the downloaded model folder.
5. Click Rescan.

Verify the Setup

The MotionSmith Setup section should report READY for:

• MotionSmith Runtime
• Kimodo-SOMA-RP-v1.1
• LLM2Vec MNTP Adapter
• LLM2Vec Supervised Adapter
• Meta Llama 3 8B Instruct

When all required components are detected, the Backend Setup Summary should report READY.

Run Diagnostics / Pre-Flight and confirm that the release gate reports READY before generating motion.

4. Model & Weights Configuration

MotionSmith Runtime already includes the NVIDIA Kimodo-SOMA-RP-v1.1 model and the required LLM2Vec adapters.

The only model that must be obtained separately by the user is Meta Llama 3 8B Instruct. MotionSmith does not redistribute the Meta Llama model weights.

Download Meta Llama 3 8B Instruct

Obtain access and download the model from the official Hugging Face repository in accordance with Meta's applicable license terms:

Meta Llama 3 8B Instruct on Hugging Face

You may store the downloaded Meta Llama model anywhere on your computer.

Select the Local Model Folder

Open Editor Preferences → Plugins → MotionSmith and set Meta Llama Folder to the downloaded model directory.

A valid model directory contains files such as:

Meta-Llama-3-8B-Instruct/
├── config.json
├── tokenizer_config.json
├── model.safetensors.index.json
├── special_tokens_map.json
├── tokenizer.json
├── model-00001-of-00004.safetensors
├── mode2-00001-of-00004.safetensors
├── mode3-00001-of-00004.safetensors
├── mode4-00001-of-00004.safetensors
└── supporting model configuration files

Click Rescan. When the model is valid, Meta Llama 3 8B Instruct will report READY.

Alternative Download Method — Hugging Face CLI

As an alternative to downloading the model files manually, you can use the official Hugging Face CLI to download the required Meta Llama 3 8B Instruct files directly into a local folder.

First, make sure your Hugging Face account has access to Meta Llama 3 8B Instruct. Then install the Hugging Face Hub CLI:

py -m pip install -U huggingface_hub

Sign in to Hugging Face:

hf auth login

Download the required model files to a folder of your choice:

hf download meta-llama/Meta-Llama-3-8B-Instruct config.json generation_config.json model-00001-of-00004.safetensors model-00002-of-00004.safetensors model-00003-of-00004.safetensors model-00004-of-00004.safetensors model.safetensors.index.json special_tokens_map.json tokenizer.json tokenizer_config.json --local-dir "C:\MotionSmith\Models\Meta-Llama-3-8B-Instruct"

When the download is complete, set Meta Llama Folder in Editor Preferences → Plugins → MotionSmith to the folder you selected, then click Rescan.

5. Backend & Pre-Flight Diagnostics

MotionSmith runs a local backend for motion generation. Setup and validation are available from Editor Preferences → Plugins → MotionSmith.

Setup Areas

• MotionSmith Setup: verifies the installed Runtime, bundled Kimodo model, LLM2Vec adapters and selected Meta Llama folder.
• Environment: validates the local Python, GPU and CUDA environment.
• Lifecycle: starts, stops or restarts the local backend when required.
• Diagnostics / Pre-Flight: performs the final environment and configuration checks before generation.

Run Diagnostics / Pre-Flight after first-time setup and after changing important runtime or model configuration. Required checks should pass before starting production generation.

If configuration has changed while the backend is already running, restart the backend so the new settings are loaded.

6. Studio Interface Overview

MotionSmith AI Studio brings generation, preview, timeline authoring, motion adjustments and retargeting into one Unreal Editor workflow.

• Authoring Viewport: preview the current character motion and spatial relationship to optional props.
• Prompt & Generate Controls: describe the next action and start generation.
• Timeline: review and manage generated or imported motion segments as a continuous performance.
• Details Panel: access generation, motion editing and retargeting controls relevant to the current workflow.
• Retargeting: transfer the working motion to a supported target character workflow without leaving MotionSmith.

7. Generation Settings

MotionSmith exposes the generation controls required for the current build directly in the Studio / Details workflow. Use the visible controls rather than modifying Kimodo or Python configuration manually.

Recommended Workflow

1. Write a focused motion prompt.
2. Select an appropriate Recipe when the current task benefits from one.
3. Review the current generation defaults in the Details panel.
4. Generate the motion.
5. Judge the result in the viewport before changing multiple settings at once.

For predictable iteration, change one meaningful setting at a time. If the result does not match the intended action, simplify or clarify the prompt before making large parameter changes.

8. Recipes System

Recipes provide convenient starting configurations for different motion-generation tasks. They are workflow presets, not replacements for a clear prompt.

Using Recipes

1. Select the recipe that best matches the type of motion you want to create.
2. Write a concise physical-action prompt.
3. Generate and review the result.
4. If the motion intent is wrong, revise the prompt first.
5. If necessary, test another recipe while keeping the prompt focused.

Recipe availability and exact labels are determined by the MotionSmith 1.0.0 build shown in the Studio interface.

9. Prompting Guide & Best Practices

Core Prompting Rules

• Describe motion, not story context: write what the body should physically do.
• Keep prompts focused: one clear action or beat usually produces easier-to-control results.
• Use direction and pace when useful: forward, backward, left, right, slowly, quickly, cautiously, etc.
• Use the timeline for choreography: split longer performances into multiple segments.
• English is recommended: concise English motion descriptions generally provide the most predictable text intent.

Less Useful	Better Motion Prompt
"The soldier is terrified and searches for an escape."	walk forward cautiously while looking from side to side
"Do an awesome fight move."	throw a right high kick and recover to a fighting stance
"Walk around happily."	walk forward with a light energetic step

Build Complex Actions as Segments

idle → walk forward → stop → turn left → step back

10. Multi-Prompt Timeline & Chaining

MotionSmith treats a longer performance as a sequence of editable motion segments. Each prompt can become its own clip on the timeline.

SEGMENT 01  idle and scan the room
SEGMENT 02  walk forward cautiously
SEGMENT 03  stop and turn toward the sound

Add Segment

Add Segment continues from the existing motion tail instead of regenerating the entire approved performance. MotionSmith preserves the earlier animation and uses the previous segment's transition context to generate the next motion.

This makes it practical to approve a section, continue from it, and revise the next beat without discarding earlier work.

11. Import & Continue Workflow

Existing Unreal animation can become part of the same MotionSmith working timeline rather than serving only as an external reference.

Imported animation and generated motion can be combined in the same MotionSmith timeline.

Imported Animation
↓
Transform / Adjust
↓
Add to Timeline
↓
Generate Next Segment
↓
Review Combined Motion
↓
Create AnimationSequence

This workflow makes it possible to combine authored or captured motion with AI-generated continuation and review the result as one performance.

12. In-Studio Motion Editing

MotionSmith includes practical tools for refining working motion without regenerating everything from scratch.

• Timeline Editing: work with individual motion segments and their placement in the performance.
• Transform Adjustments: reposition or reorient imported and working motion when required.
• Pose Controls: adjust a pose and hold it when the workflow requires a fixed pose decision.
• Mirror Pose: mirror a pose across the character's local sagittal plane.
• Mirror Animation: create a mirrored animation variation while preserving the intended trajectory behavior.
• Root / Path Authoring: refine character travel and root-path decisions using the controls available for the current clip.

The goal is to keep approved animation intact and make targeted corrections instead of returning to full regeneration for every change.

13. Props & Spatial Context

Props can be loaded into the MotionSmith authoring viewport as spatial references while building or reviewing character motion.

Use props as spatial references while authoring and reviewing character motion.

1. Load the prop into the MotionSmith viewport.
2. Position it relative to the character.
3. Generate, import or edit the character motion.
4. Review the spatial relationship in the viewport.
5. Adjust the motion or prop placement as needed.

Props provide authoring context. MotionSmith's primary output remains the character animation; the presence of a prop does not imply automatic hand attachment or guaranteed physical contact solving.

14. Skeleton Retargeting

MotionSmith works with motion generated on the NVIDIA Kimodo SOMA character representation and transfers the finished motion to target Unreal characters through Unreal Engine retargeting workflows.

Retarget generated motion to the target Unreal character without leaving the MotionSmith workflow.

Included Target Workflows

• Epic / Manny-based characters
• MetaHuman
• Character Creator
• Mixamo-based characters

Typical Retarget Workflow

1. Generate or import and approve the source motion.
2. Open the MotionSmith retargeting controls.
3. Select the appropriate target workflow.
4. Select the target character / skeletal mesh.
5. Run the retarget operation.
6. Review the result on the target character before final output.

Characters outside the included profiles may require a separate Unreal IK Rig / IK Retargeter setup and validation.

15. Animation Output

MotionSmith creates standard Unreal Engine animation output so the finished performance can continue through normal production tools.

MotionSmith → AnimationSequence → Unreal Production Pipeline

The resulting animation can be used with standard Unreal workflows such as:

• Sequencer
• Animation editing tools
• Animation Blueprints
• Montages and state-machine workflows
• Further Unreal retargeting or cinematic production

Multiple generated segments, or imported and generated segments, can be reviewed as one continuous performance before creating the final asset.

16. Troubleshooting & FAQ

MotionSmith Runtime Is Not Detected

Confirm that the Runtime was extracted to %LOCALAPPDATA%\MSR\ and that %LOCALAPPDATA%\MSR\runtime.json exists. Then return to Editor Preferences → Plugins → MotionSmith and click Rescan.

Meta Llama 3 Model Is Not Detected

Make sure Meta Llama Folder points to the model root containing the configuration, tokenizer and weight files. Then click Rescan and run Diagnostics / Pre-Flight again.

Backend Is Not Ready

Open Editor Preferences → Plugins → MotionSmith and confirm that the Runtime, Kimodo model, both LLM2Vec adapters and Meta Llama all report READY. Then run Diagnostics / Pre-Flight.

Configuration Changed but the Backend Uses the Old State

Restart the MotionSmith backend from its Lifecycle controls so the updated runtime or model configuration is loaded.

Retargeted Motion Looks Incorrect

Confirm that the selected MotionSmith target workflow matches the target skeleton and that the correct skeletal mesh is selected. For custom setups, validate the Unreal IK Rig / IK Retargeter independently.

The Generated Motion Does Not Match the Prompt

Simplify the prompt and describe one physical action at a time. Split longer choreography into separate timeline segments.

The Next Segment Needs a Different Direction

Keep the approved earlier motion and revise the next segment prompt rather than rebuilding the entire performance.

GPU Memory Is Constrained

Reduce unnecessary GPU-heavy editor workloads while generating and close unrelated GPU-intensive applications. Memory use varies with the current Unreal project and runtime state.

17. Third-Party Licenses & Credits

NVIDIA Kimodo

MotionSmith AI is powered by the open-source NVIDIA Kimodo motion-generation project. MotionSmith AI is an independent product and is not affiliated with or endorsed by NVIDIA Corporation.

NVIDIA Kimodo Documentation

Meta Llama 3

Built with Meta Llama 3. MotionSmith AI does not distribute the Meta Llama 3 model weights. Users are responsible for obtaining and using the model in accordance with Meta's applicable license terms.

Meta Llama 3 8B Instruct

18. Technical Support

For bug reports, setup problems or MotionSmith AI product support:

📧 Support: motionsmithai@gmail.com

When reporting a reproducible issue, include:

• Unreal Engine version
• MotionSmith AI version
• GPU model
• The failed Diagnostics / Pre-Flight item, if any
• A relevant Unreal log excerpt
• Short steps that reproduce the problem