How to Run NVIDIA Komodo Locally for Free and Animate Any 3D Character (6GB VRAM)

AI motion generation is no longer locked behind cloud services or expensive rigs. With NVIDIA’s Komodo text-to-motion model, you can generate realistic 3D character animations from text prompts, run everything locally, and only need around 6GB of VRAM. In this guide, you’ll see how to install Komodo with a one-click script, create custom motions, and retarget them to a 3D character in Blender using a completely free pipeline.

What Is NVIDIA Komodo?

Komodo is NVIDIA’s text-to-motion diffusion model. Instead of generating images or video, it creates 3D human motion directly from text descriptions.

It was trained on around 700 hours of optical motion capture data, which allows it to produce natural, realistic movement. Komodo supports:

• Text prompts (e.g. “a person walking forward confidently”)
• Constraints such as full-body pose keyframes
• Target positions and 2D paths to guide where the character moves

The official setup normally relies on a cloud-based text encoder. However, thanks to a community-made quantized text encoder, you can now run the entire pipeline fully offline on your own hardware, with roughly 6GB of VRAM.

If you’re already experimenting with local 3D AI, Komodo pairs nicely with tools like Trellis 2 for local 3D character generation.

Setting Up Komodo Locally (One-Click Install)

The local setup is wrapped in a simple one-click script. You only need three files:

• install_komodo.bat
• start_komodo.bat
• A small Python wrapper file

Download these files into a folder where you want Komodo to live. Before running anything, make sure you have Python and Git installed on your system. Any recent Python 3 version works (3.11, 3.12, 3.13, etc.).

Install Komodo

1. Run the installer: Double-click install_komodo.bat. This will:

• Create a Python virtual environment
• Install all required dependencies
• Download the Komodo models and the offline text encoder

The process runs through about a dozen steps and can take up to 15 minutes depending on your connection. Just leave the window open until it reports that the installation is finished.

If you have less than 8GB of VRAM, you can use an --offload flag to push more of the model into system RAM and reduce GPU memory usage.

Start the Komodo Web Interface

2. Launch Komodo: Double-click start_komodo.bat. A terminal window will open and Komodo will:

• Detect your system RAM and GPU VRAM
• Decide how much to offload to system RAM
• Load the model and start a local web server

Once it’s ready, the terminal will show a local URL (e.g. http://127.0.0.1:xxxx). Copy that URL, paste it into your browser, and you’ll see the Komodo interface and a quick start guide.

Generating Motion from Text Prompts

Inside the Komodo interface, switch to dark mode if you prefer, then choose your skeleton type on the right. For most use cases, the “SOMA human body” skeleton works best and is the recommended default.

Basic Text-to-Motion

Komodo comes with example scenes you can load. For a simple text prompt example:

• Load the “single text prompt” example
• You’ll see a character and a text box at the bottom
• Scrub the timeline to preview the existing motion

To create your own animation:

1. Double-click the text prompt at the bottom.
2. Always start with “a person …” because the model was trained on that pattern and it yields the best results.
3. For example: “a person walking forward confidently”.
4. Press Enter, then click Generate.

Within a few seconds, Komodo will generate a new motion. You’ll see VRAM and RAM usage in the interface, and you can preview the animation by pressing Spacebar.

You can also adjust:

• Total length (e.g. 5 seconds)
• Number of samples (generate multiple motion variations at once)
• Seed (to reproduce or slightly vary a motion)

Combining Multiple Prompts

Komodo also supports multi-prompt timelines. For example, you can have:

• Frames 0–180: “a person walks forward”
• Frames 180–X: “a person walks in a circle”
• Later frames: “a person does a front flip”

By loading a multi-prompt example, you can see how different actions are chained together. Follow the same pattern in your own prompts: always start with “a person …” and describe each action clearly.

Using Constraints to Control Poses

One of Komodo’s most powerful features is constraints. Constraints let you define specific poses at certain frames, and Komodo will generate motion that respects those keyframes while filling in smooth in-betweens.

Editing a Pose

To add a constraint:

1. Go to the Constraints tab and enter Edit mode.
2. Move to a frame on the timeline (e.g. frame 75).
3. Adjust the character’s pose by rotating arms, legs, feet, etc.
4. Click the buttons to save that pose as a constraint at the current frame.

Initially, you might see the animation “snap” into the pose at that frame, because the motion hasn’t been regenerated yet.

To fix this:

1. Exit Edit mode.
2. Keep the same seed.
3. Click Generate again.

Komodo will re-generate the motion, this time smoothly blending into and out of your constrained pose. This gives you precise control over:

• Where the character is in space at certain frames
• When they kneel, pick something up, or reach a pose
• How complex actions transition from one to another

Preparing a T-Pose for Export

Before exporting, it’s important to have your character in a T-pose at frame 0. This greatly improves retargeting results later in Blender.

To do this:

1. Go to frame 0.
2. Enter Edit mode in Constraints.
3. Manually pose the character into a rough T-pose (arms out, straight body).
4. Save that pose as a constraint at frame 0.

Because you only care about frame 0 for the T-pose, you don’t need to regenerate the whole motion for this step.

Exporting Motion and Retargeting in Blender

Once you’re happy with the motion, it’s time to export it and apply it to a custom 3D character in Blender. The full pipeline uses only free tools:

• Komodo – generate motion
• Trellis 2 – generate a 3D character (optional but recommended)
• Mixamo – auto-rig the character
• Rococo (Rokoko) Blender add-on – retarget the animation
• Blender – final animation and rendering

Exporting from Komodo

In Komodo:

1. Go to the Export section.
2. Choose an output format, typically BVH (you can also use NPC).
3. Click Download and save the file.

This BVH file contains the skeleton and the motion you just generated.

Importing the Motion into Blender

In Blender:

1. Delete the default cube.
2. Go to File > Import > BVH.
3. Select your exported Komodo BVH file.

You’ll see a skeleton with the motion already applied. Scrub the timeline to confirm the animation is working.

Creating and Preparing Your 3D Character

You can generate your 3D character with any tool you like. A powerful option is to use a local 3D model generator such as Trellis 2 running locally on 6GB VRAM, then bring that model into Blender.

Once you have your character model in Blender:

1. Import the character (e.g. as GLB/FBX).
2. Enter Edit Mode on the mesh.
3. Right-click > Merge Vertices > By Distance to clean up duplicate vertices.
4. Press Ctrl+Shift+N to recalculate normals.
5. Optionally, enable Blender’s free 3D Print add-on and run its cleanup tools to remove non-manifold edges and further clean the mesh.

When the mesh is clean, export it as an FBX (e.g. named soldier.fbx) with only the selected objects.

Auto-Rigging with Mixamo

Next, rig the character using Mixamo:

1. Go to the Mixamo website and log in.
2. Upload your character FBX.
3. Place the markers for chin, wrists, elbows, knees, and groin when prompted.
4. Let Mixamo auto-rig the character.
5. Once rigged, download the character as an FBX in “Pose to Pose” mode.

This gives you a Mixamo-compatible rig that works well with many retargeting tools.

Bringing the Rigged Character Back into Blender

In Blender again:

1. Import the Mixamo-rigged FBX.
2. Re-assign materials/textures if needed by selecting the mesh and choosing the correct material slot.
3. Apply transforms with Ctrl+A > All Transforms to clear any unwanted scaling or rotation.

You should now have:

• The Komodo BVH armature with the jump (or other) animation
• The Mixamo-rigged character armature and mesh

Retargeting with the Rococo (Rokoko) Add-on

Install and enable the free Rococo Blender add-on (often referred to as Roco/Rokoko). Once installed, you’ll get a retargeting panel in Blender.

To retarget:

1. Make sure both the source (Komodo BVH) and target (Mixamo character) are in a T-pose at frame 0.
2. Go to frame 0 in the timeline.
3. In the Rococo panel, set the Source to the Komodo armature (e.g. jump).
4. Set the Target to the Mixamo armature (your soldier’s armature).
5. Click Build Bone List so the add-on maps Komodo bones to Mixamo bones.
6. Enable Autoscale and set it to Current so the animation scales correctly between skeleton sizes.
7. Click Retarget Animation.

After processing, you can hide the original Komodo armature. Scrub the timeline and you’ll see your custom character performing the AI-generated motion.

From frame 1 onward, the character should jump, walk, flip, or perform whatever action you generated in Komodo.

Note: In this workflow, hand and finger animation may not be fully detailed by default, since Mixamo’s standard rig and the exported Komodo motion don’t always include fine-grained hand data. You can always add hand keyframes manually or layer additional animation on top.

Fully Local, Free, and Ready for Your Projects

With this setup, you get a complete, fully local animation pipeline:

• Komodo for AI-generated motion from text and constraints
• Trellis 2 (optional) for AI-generated 3D characters
• Mixamo for free auto-rigging
• Rococo’s Blender add-on for retargeting
• Blender for final tweaks, lighting, and rendering

Everything runs on your own hardware, and with offloading you can get it working on GPUs with around 6GB of VRAM. It’s a powerful way to build game-ready or cinematic animations without manual keyframing every move.

Once this pipeline is in place, you can generate essentially unlimited 3D animations for your characters—perfect for indie games, prototypes, previs, or any 3D project where you want fast, controllable motion without cloud costs.