MIUProject/ARDY-Core-RP-20FPS-Horizon40-int4

Llama 3 ARDY Core RP 20FPS Horizon40 BF16 / INT4 Encoder

Built with Meta Llama 3. This is a reduced-size ARDY bundle intended for fast inference and lower memory use in ComfyUI VNCCS Utils. The motion denoiser and motion tokenizer use BF16; the LLM2Vec text encoder uses ConvRot W4A4 INT4. No additional training was performed. Motion weights are not INT4 or INT8.

Contents

File	Contents	Bytes
motion.bf16.safetensors	Consolidated ARDY denoiser and FSQ motion tokenizer; all 428 tensors in BF16	382,564,544
text_encoder/model.safetensors	Consolidated Llama encoder with both adapters merged; 225 ConvRot INT4 matrices	3,758,959,600
config.yaml	Architecture; both motion components select the BF16 checkpoint	—
text_encoder/*.json	Encoder configuration and tokenizer	—
stats/	Original motion and tokenizer normalization statistics	—
provenance.json	Source repositories, encoder revision pins and precision policy	—
LICENSE, LICENSES.md, NOTICE, licenses/	Component licenses and attribution	—

The two weight files total 4,141,524,144 bytes (3.86 GiB). Statistics remain in original precision. BF16 reduces motion checkpoint storage by approximately 50% relative to the consolidated FP32 checkpoint. Size on disk is not runtime RAM or VRAM.

Precision and runtime

Motion parameters are assigned in BF16 without expanding them to FP32. Neural motion computation uses CUDA BF16 autocast; sampling state, statistics and skeleton reconstruction remain FP32. Encoder execution disables the caller's autocast.

Eligible encoder matrices use native signed ConvRot W4A4 INT4 with FP32 row scales. Both LLM2Vec adapters were merged in MNTP then supervised order. The unused language-model output head is omitted. Other encoder parameters use BF16. Packed embeddings decode only requested vocabulary rows; the encoder is offloaded between prompts. Compilation is disabled.

The encoder safetensors identifier is vnccs-convrot-w4a4-v1. This bundle requires the VNCCS Utils compact loader and compatible comfy-kitchen ConvRot kernels. It is not a drop-in stock Transformers checkpoint or a generic ComfyUI diffusion model. The motion runtime also requires CUDA BF16 support.

Installation

Copy the runtime files and subdirectories to:

ComfyUI/models/text_to_motion/ARDY-Core-RP-20FPS-Horizon40-int4/

Select ARDY Core RP — BF16 motion / INT4 encoder in Pose Studio's Motion panel. Its stable model ID is ardy-core-rp-20fps-h40. The existing directory name retains int4 for compatibility; this describes the encoder, not the motion weights.

The node loads ARDY only from the ComfyUI models directory. This publication folder inside the project is never an inference source. Missing BF16 files produce a setup error; ordinary ARDY weights are not a fallback. Kimodo remains supported separately.

No Hugging Face repository has been created or published by the node. Automatic download will be configured after the owner publishes this folder and supplies the repository ID. No conversion scripts or validation outputs are required for inference.

Recorded speed and memory

RTX 5070 Ti, Windows ComfyUI, PyTorch 2.11.0+cu130, comfy-kitchen 0.2.35. Walking prompt, seed 42, two seconds / 40 frames, no initial pose, full 10-step schedule. One warmup and one measured generation; encoder execution and transfers are included.

Measurement	BF16 motion / INT4 encoder
Warm generation	1.929 s
Generation allocated VRAM peak	4.18 GiB
Generation reserved VRAM peak	4.31 GiB
Generation process RAM peak	31.12 GiB
Loading/generation process RAM peak	32.46 GiB

RAM is whole-process sampled RSS, including other resident ComfyUI components. These measurements are not model-only memory requirements. Duration, concurrent nodes and hardware affect speed and memory. Numerical motion differs slightly from FP32; the BF16 variant was selected after reviewing the generated animation.

Motion scope

ARDY generates motion for one 27-joint Core human skeleton at 20 FPS. This integration supports 1–10-second clips, a seed, guidance and an optional initial-pose constraint. It does not model scene objects, multi-character interactions or individual finger motion. Upstream foot-skate post-processing is not included.

Pose Studio projects source rotation changes relative to frame zero onto the target skeleton through its shared rotation-import mechanism. Target bone offsets and lengths are preserved. Different proportions can change hand and foot contacts. This precision policy does not guarantee numerically identical results to the original FP32 model.

Licenses and attribution

The bundle combines the NVIDIA Open Model License for ARDY, the Meta Llama 3 Community License and Acceptable Use Policy for the encoder, and MIT terms for LLM2Vec adapters. Full texts and attribution are included. Preserve the Llama 3 model title and Built with Meta Llama 3 notice when redistributing.

Source repositories and pinned encoder revisions are recorded in provenance.json. The historical ARDY source cache used main; its exact commit SHA was not recorded. NVIDIA, Meta, NousResearch and McGill-NLP did not publish or endorse this derivative.