Introduction

What if you could direct a character's blocking the way you'd play a game, by clicking where they should walk, and then hand that exact motion to a video model? That's what ComfyUI-Ardy does. It's a fresh open-source node pack from boricuapab, first pushed at 05:52 UTC on November 1 (13:52 MYT).

It wraps NVIDIA's ARDY, an autoregressive diffusion model for interactive human motion generation (SIGGRAPH 2026, Apache-2.0 code), and runs it inside ComfyUI. The wrapper is MIT licensed. You steer a live 3D body, record the take, and export it as a GLB or as an ordered frame batch plus silhouette mask sized for MiniMax H3 reference-video inputs.

What Shipped

  • ARDY Live Preview: a real-time viewport inside the node. Click the ground grid to add waypoints, or hold the arrow keys for camera-relative steering. You can change the text prompt and speed while it runs.
  • Chunked motion: the Core checkpoint generates 40 frames at 20 FPS per chunk, so a steering change appears about two seconds later. Live runs default to 120 seconds, and 0 runs until you stop.
  • Record and save: Stop & save take writes a recording to output/ardy/recordings. Saved takes are reused across restarts without reloading ARDY.
  • ARDY Motion to GLB: export as mesh, mesh plus skeleton, or skeleton only, then send it to ComfyUI's native Preview 3D (Advanced) and 3D save nodes.
  • ARDY Preview Advanced · Driving Frames: renders the take as an IMAGE batch plus a body-silhouette MASK, in solid, normal or depth shading, at your resolution and camera angle.
  • ARDY Camera Recorder: orbit and zoom while the take plays back to record a camera path. Save it as JSON and reuse it across render nodes.
  • H3-ready timing: output_fps defaults to 24 for MiniMax H3 and resamples ARDY's 20 FPS motion without changing its duration.
  • Leaner text encoder: an optional INT8 ConvRot Llama-3-8B encoder applies the LLM2Vec adapters without merging. The original BF16 encoder needs about 16 GB just for its weights, or you can run it on the CPU.

How to Use It

From your ComfyUI/custom_nodes folder:

git clone https://github.com/boricuapab/ComfyUI-Ardy.git
python -m pip install -r ComfyUI-Ardy/requirements.txt

The requirements build a pinned NVIDIA ARDY from source, which needs Git, CMake and a C++17 toolchain. On Windows that means Visual Studio Build Tools. The nodes never download weights, so place them yourself:

  • ComfyUI/models/ardy/ARDY-Core-RP-20FPS-Horizon40 (the full checkpoint, including its stats folder)
  • ComfyUI/models/ardy/text_encoder/ with Meta-Llama-3-8B-Instruct (gated on Hugging Face) plus the two LLM2Vec MNTP adapter folders

A basic motion-to-3D graph:

Load ARDY Text Encoder -> Load ARDY Model -> Generate ARDY Motion -> ARDY Motion to GLB -> Preview 3D (Advanced)

For video, record a take in ARDY Live Preview. Then send recorded_motion to ARDY Preview Advanced · Driving Frames, and pass its images into your H3 reference-video input or into Create Video.

Good to Know

  • Licenses: the wrapper is MIT and the ARDY code is Apache-2.0. The ARDY checkpoints ship under the NVIDIA Open Model agreement, and the Llama 3 text encoder is gated by Meta.
  • Very early: the repo had 0 stars and a single commit when we checked. It does include unit tests and says the recording workflow was verified with ComfyUI's dynamic VRAM.
  • Queue use: live generation occupies the ComfyUI queue until you stop it. Use Stop & save take, not a queue cancel, to keep the recording.
  • Memory: ARDY and its encoder offload after running, so your downstream H3 pass doesn't keep their weights in VRAM.

Why It Matters

Pose and motion control for video models usually starts with a reference clip you filmed or found. ComfyUI-Ardy lets you generate that clip from text, steer it like a game character and choose the camera, all in the same graph that feeds H3. It's a neat bridge between open motion research and everyday AI filmmaking.

Conclusion

This one feels like a toy in the best way. Click a path across the grid, watch the body walk it and save the take. Then let H3 dress that motion in any world you can prompt. It's early and has a few setup steps, but it's a lovely new way to block a shot.

Want to see how H3 moves before you wire up the motion pass? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-ardy-nvidia-motion-steering-minimax-h3-driving-frames

Original Source

Click a path, and happy directing!

—Aurelia ♡