AZKimodo — Kimodo text-to-motion for Windows and Maya

Type a sentence, get a character animation — on your own GPU, baked onto your own rig.

AZKimodo is the complete build around kimodo.cpp, the community C++/GGML port of NVIDIA's Kimodo text-to-motion diffusion model. Three parts, one author:

Built by Alexander Antonov — AzRigTool. Not affiliated with NVIDIA or LocalAI.

1. Download kimodo.bat from the latest release.
2. Put it anywhere and double-click. Press Enter at the folder prompt. Say Yes to the Windows permission prompts.
3. Wait — the first run installs what is missing and downloads about 19 GB of weights. Expect the best part of an hour.
4. The browser opens on the demo when it is ready. Every later run opens it in seconds.

Leave the window open while you use the demo; it is the server.

Requirements: Windows 10/11 x64 with winget, ~30 GB free, internet, a Vulkan-capable GPU for speed (CPU works, slowly). Non-ASCII user names are fine.

Usage: kimodo.bat — or kimodo.bat D:\folder 8095 to pick the folder and port.

Prebuilt binaries: kimodo-win64-vulkan.zip in the release holds kmd-generate, kmd-constrain, kmd-encode, kmd-inspect and the ggml DLLs, if you would rather not compile. Unpack into your checkout; weights still come from kimodo.bat.

azigo45/kimodo.cpp is upstream main plus one commit, four files:

• kmd-constrain (new) — generate through poses supplied from outside. Reads a .kmdp file: frames, world rotations per joint, root, and a per-element mask. Because the mask is per element, a pin can be a whole pose, the root row alone (follow a path, invent the gait), the feet, or an arm chain. Writes the text embedding beside the take so the next solve with the same prompt skips the 15 GB encoder.
• Hard pinning — sample_motion_from_noise_conditioned gains project_observed: masked channels of the clean prediction are forced back onto their observed values every step. Conditioning alone is a soft pull; projection makes a pin exact. Measured: a pinned key pose reproduced at 0.000 cm.
• Weighted masks — the mask is a float, not a flag. A fractional weight blends the pinned value with the prediction, which is what lets a held span ease in and out instead of switching on between two frames.
• One denoiser graph — the transformer used to build 18 per-layer GGML graphs on every denoising step. They are one graph now: ~5× faster on Vulkan (RTX 4080), output bit-identical.

Apache-2.0, as upstream.

maya/azkimodo/ — two files, Maya 2025, PySide6/PySide2.

Install: copy the azkimodo folder into Documents\maya\2025\scripts\ (or anywhere on Maya's Python path). Then in the Script Editor:

from azkimodo import az_kimodo_gen_qt
az_kimodo_gen_qt.show()

The tool finds the engine on its own: $AZ_KIMODO_ROOT if set, otherwise the first kimodo.cpp checkout on any drive — which is wherever kimodo.bat put it.

What it does

• Generate from a text prompt (several segments with transitions), at your scene's frame rate, non-blocking — Maya stays responsive while the GPU solves.
• Retarget the take onto a rig. Ships with maps for the UE5 Mannequin skeleton and for an Advanced Skeleton rig built on it: animation is baked onto the FK controls, root travel goes on the Main control, limbs are switched to FK for the bake and switched back when you clear.
• Pose pins. Pose the rig at a frame, press Pose, and the next generation has to pass through that pose. Modes: whole pose, exact, root only, feet, hands, hands locked, upper body, lower body. A pin can hold across a span — a hand grabs something and stays. A ghost skeleton shows each pin on its frame.
• Regenerate as many times as you like: edit a pin, re-read the pose, solve again, the take replaces the previous one on the rig. Round-trip measured at 0.0000° across all mapped joints — reading a pose off the rig and writing it back does not drift.
• Honest defaults, measured. Hard pinning is off unless a mode needs it, because on a 20-frame hold it held the arm no better than soft pinning and lurched 130 cm in one frame where soft moved 36. Eased holds fade in and out over three frames. Travel between pins is filled with root-only pins so the character walks the distance instead of loitering and bolting.

Versioning: the tool prints its version in the window title and in every log line ([AZ Kimodo Motion 1.29.0]).

• Installer: the window stays open with the reason; run the file again and it resumes. Weights failed after 12 attempts → Hugging Face throttling, run again later. Running on the CPU → install the Vulkan SDK, delete <folder>\build, run again. "vcvars64.bat missing" → Visual Studio Installer → Modify → Desktop development with C++.
• Maya tool: no bind pose was recorded → put the rig at bind and press Remember bind pose. the rig carries animation, so its bind pose cannot be measured → retarget a take once, or Clear rig animation while at bind. Pins captured with an older version carry that version's reference; re-read them with Pose.

Full notices and license texts: THIRD_PARTY.md and licenses/.

• Installer and Maya tool: MIT (LICENSE), © Alexander Antonov (AzRigTool). The skeleton tables inside the Maya tool are copied from kimodo.cpp and stay Apache-2.0.
• Engine: kimodo.cpp by LocalAI contributors, Apache-2.0; the fork keeps its license and NOTICE, and marks the files it changed. The prebuilt binaries carry the Apache-2.0 and ggml (MIT) license texts inside the archive.
• Models — downloaded by the installer from their publishers, not redistributed here: Kimodo SOMA / G1 by NVIDIA under the NVIDIA Open Model License (commercially usable, outputs are yours); the SMPL-X model is non-commercial and is not downloaded; the text encoder is built with Meta Llama 3 under the Meta Llama 3 Community License.
• Running kimodo.bat accepts third-party license terms on your behalf — Microsoft Visual Studio Build Tools, LunarG Vulkan SDK, Git for Windows, Python and Go are installed unattended, exactly as if you had clicked through their installers.
• "Kimodo" is NVIDIA's model name, used here only to say what this installs. Not affiliated with NVIDIA, Meta or LocalAI.
• The .bat is ASCII with CRLF on purpose — cmd cannot find labels in a file saved with LF.

• 2.2.0 — the complete build: engine fork published, installer clones it (so kmd-constrain is built), Maya tool published as the azkimodo package (1.29.0), prebuilt Windows binaries attached.
• 2.1.x — installer: fixes from an adversarial review (bracket in the compiler path aborted the file; Python scripts folder unreadable on a Cyrillic user name; CPU-build remedy; Git Bash location), drive-root prompt, comments stripped, author line.
• 2.0.x — installer and launcher merged into one file; fast path; window kept open on error; folder remembered.

Полная сборка text-to-motion Kimodo под Windows: один установщик kimodo.bat (двойной клик — через час GPU-сборка и демо в браузере, потом за секунды), форк движка с генерацией через заданные позы (kmd-constrain), жёстким пиннингом и слитым графом денойзера (~5× быстрее, бит в бит), и инструмент для Maya: генерация из текста, ретаргет на риг Advanced Skeleton с запеканием в контролы, пины поз, снятые прямо с рига, холды, перегенерация без дрейфа.

Установка Maya-части: папку maya/azkimodo в Documents\maya\2025\scripts\, в Script Editor — from azkimodo import az_kimodo_gen_qt; az_kimodo_gen_qt.show().

Сборка: Alexander Antonov — AzRigTool.