Calluni-Wan 2.2-Animate-14B: Real-Time Spatio-Temporal Diffusion Transformer
Author & Research Lead: Aravindhan (Calluni AI Research Labs)
Organization: Calluni Telepresence & Neural Synthesis
Model Repository: Aravindhan11/calluni-wan2.2-animate-14b
License: Apache 2.0
Associated Paper: Calluni-Wan: Real-Time End-to-End Character Animation via Spatio-Temporal 14B Diffusion Transformers for Universal Holographic Telepresence (IEEE TVCG 2026)
1. Model Overview
Calluni-Wan 2.2-Animate-14B is a 14-billion parameter Spatio-Temporal Diffusion Transformer (DiT) designed for real-time expressive character animation and universal holographic telepresence. By eliminating intermediate structural estimators (such as 2D/3D facial landmarks, SMPL meshes, and blendshapes), Calluni-Wan consumes raw driving video tokens directly through a dual-branch cross-attention mechanism.
Key Highlights
- 14 Billion Parameters: 40 spatio-temporal transformer layers with 3D Rotary Position Embeddings (3D-RoPE).
- Zero-Intermediate Architecture: Direct motion transfer without synthetic keypoints or hand-crafted skeletons.
- Step-Distilled Real-Time Inference: Generates photorealistic frames in only 10–12 ODE sampling steps.
- Ultra-Low Latency: Sub-40ms glass-to-glass latency at 30+ FPS when deployed with TensorRT / BF16 optimization.
- Infinite Streaming Stability: Sliding-window latent cache guarantees continuous multi-hour broadcasts with zero memory drift.
2. Model Weights Inventory (78.4 GB Total)
This repository hosts the complete, uncompressed weights in BF16 and PyTorch formats:
| File Path | Size | Description |
|---|---|---|
wan_animate_2/wan_animate_2_bf16.safetensors |
31.27 GB | Full 14B Spatio-Temporal DiT backbone (BF16) |
wan_animate_2/wan_animate_2_bf16_distillation.safetensors |
31.27 GB | Step-distilled real-time engine (10-step ODE) |
videomodel/Wan-AI/models_t5_umt5-xxl-enc-bf16.pth |
10.84 GB | Text condition encoder (uMT5-XXL BF16) |
videomodel/Wan-AI/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth |
4.55 GB | Visual identity encoder (OpenCLIP ViT-H/14) |
videomodel/Wan-AI/vae.pth |
747.5 MB | 3D Causal Spatio-Temporal VAE |
3. Quickstart & Inference
import torch
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
# Download the 14B step-distilled model
weights_path = hf_hub_download(
repo_id="Aravindhan11/calluni-wan2.2-animate-14b",
filename="wan_animate_2/wan_animate_2_bf16_distillation.safetensors"
)
# Load into GPU memory (BF16)
state_dict = load_file(weights_path)
print(f"Loaded Calluni-Wan 14B weights successfully. Keys: {len(state_dict)}")
4. Citation
@article{aravindhan2026calluniwan,
title={Calluni-Wan: Real-Time End-to-End Character Animation via Spatio-Temporal 14B Diffusion Transformers for Universal Holographic Telepresence},
author={Aravindhan and Calluni AI Research Team},
journal={IEEE Transactions on Visualization and Computer Graphics (TVCG)},
year={2026},
publisher={IEEE}
}
Developed and maintained by Aravindhan at Calluni AI Research Labs.