DARK DOJO STUDIOS

Documentation: https://darkdojostudios.com/pages/dds-motion-kimodo-installation-guide

DDS Motion is an Unreal Engine 5.8 plugin that lets you author animation by describing it.
It runs NVIDIA's Kimodo motion model locally, on your own GPU, with no cloud service, no
subscription and no per-generation cost. Your prompts and your animation never leave your
machine.

It is not a one-shot generator that hands you a clip and wishes you luck. Generation is the
first step of a real authoring workflow that ends with a retarget-ready animation on your
own character.

What it does

Write it. Describe the motion you want. Build a timeline of several prompts with
individual durations and the plugin blends between them, so one clip can walk, plant its
feet and stop.

Generate it. Runs locally on your GPU. Control seed, sample count, diffusion steps and
how strongly the result follows your text against your posed constraints. Generate several
samples at once and pick the best.

Pose it. AutoPose gives you hand control over the result. Pose bones directly with full
FK and IK, key what you want held, and let generation respect those keys. Make Loop turns a
clip into a seamless cycle.

Ground it. Attach a prop to the hand or place a chair, crate or ledge in the scene, then
pose against something real instead of guessing where the character's hands should be.

Ship it. Trim to the frames you want, choose root motion or in place, and export. Then
retarget onto Manny, Quinn, a MetaHuman or your own skeleton.

Prove it. Every generation writes an immutable result asset recording the recipe hash,
the resolved seed, the model revision and the artifact checksums. You can always show where
an animation came from and reproduce it exactly.

Built for locomotion sets

Prompt timelines, fixed seeds and consistent settings make it practical to build a set of
clips that blend with each other, which is what Motion Matching and blendspace-based
locomotion actually need. Consistency across a set matters more than any single clip, and
the reproducibility model is designed for exactly that.

Runs on a normal graphics card

Kimodo needs roughly 17 GB of video memory to run everything on the GPU. One setting moves
prompt encoding to system RAM and brings measured peak video memory down to about 2 GB.

Measured on an AMD Radeon RX 7900 XTX under ROCm, through the plugin's own code path:

| Text encoder | Clip | Peak video memory | Time |
|---|---|---|---|
| GPU | 2 seconds, 1 sample | 15.93 GiB | 45 seconds |
| CPU | 2 seconds, 1 sample | 1.82 GiB | 39 to 52 seconds |
| CPU | 6 seconds, 1 sample | 1.84 GiB | 41 seconds |
| CPU | 6 seconds, 4 samples | 2.01 GiB | 51 seconds |

Video memory barely moves with longer clips or more samples. Tripling the duration and
quadrupling the sample count added under 0.2 GiB. The setting costs you time, not quality:
the motion model is unchanged and still runs on your GPU.

System RAM is a separate budget and it does not change between the two paths. Measured peak
RAM was 15.7 GiB with the encoder on CPU and 16.4 GiB with it on the GPU, so plan for 32 GB
regardless of which card you have.

Because Unreal also needs video memory, plan for about 6 GB as a practical floor and 8 GB
for comfort. These figures were measured on AMD under ROCm. NVIDIA is expected to behave
the same way but has not yet been measured on NVIDIA hardware.

Example content included

The plugin ships with finished work you can open immediately: five recipes, five generated
animations and two exported clips, playing on the included preview body. You can read every
prompt, timeline and pose key before you install anything else. No model, GPU or account
required just to look.

---

What you need

- Unreal Engine 5.8, Windows
- A GPU with about 6 GB of video memory or more
- 32 GB of system RAM. This is not a low video memory requirement, it applies whatever your
  card is: measured peak RAM was 15.7 GiB with the text encoder on CPU and 16.4 GiB with it
  on the GPU, because the model weights pass through system memory either way. 16 GB is not
  enough once Windows and Unreal are also resident.
- About 25 GB of disk for model files and the Python environment
- A free Hugging Face account, and Meta's approval for Llama 3, which is used to encode your
  prompt text

The motion model, text encoder and GPU runtime are not redistributed with the plugin.
You download them yourself under your own accounts and license acceptances. The included
installation guide walks through every step, including Meta's approval process.

Status

DDS Motion is in beta. It is in active development and shipping real work. The included user
and installation guides are complete and current.