LivePortrait Models (safetensors)

Mirror of the LivePortrait model weights, converted from .pth to safetensors format for safe, pickle-free loading. Hosted by Æmotion Studio for use with ComfyUI-FFMPEGA.

What is LivePortrait?

LivePortrait is an AI model for portrait animation — it transfers head pose, facial expressions, eye gaze, and lip movements from a driving video onto a source face image. It produces high-quality results in real-time.

Paper: LivePortrait: Efficient Portrait Animation with Stitching and Retargeting Control

Files

File Size Params Description
appearance_feature_extractor.safetensors 3.2 MB 836K Extracts 3D appearance features from face crops
motion_extractor.safetensors 107.3 MB 28.1M Extracts keypoints, head pose, and expressions
spade_generator.safetensors 211.5 MB 55.4M SPADE decoder — generates output face images
warping_module.safetensors 173.7 MB 45.5M Warps appearance features via dense motion
stitching_retargeting_module.safetensors 0.9 MB 227K Stitching + lip/eye retargeting networks
Total 496.6 MB 130M

⚠️ Conversion Note

Weights were converted from the official .pth checkpoints in KlingTeam/LivePortrait using:

from safetensors.torch import save_file
import torch

# For base models (direct state dict)
sd = torch.load("model.pth", map_location="cpu")
save_file(sd, "model.safetensors")

# For stitching module (nested sub-dicts → flat prefixed keys)
# retarget_shoulder → stitching.*
# retarget_mouth   → retarget_lip.*
# retarget_eye     → retarget_eye.*

The stitching/retargeting module uses flat prefixed keys in safetensors format. The module. DDP prefix is stripped from all keys.

Usage

These models are automatically downloaded by ComfyUI-FFMPEGA when the Animate Portrait skill is used. No manual setup required if allow_model_downloads is enabled.

Manual Installation

  1. Download all .safetensors files from this repo
  2. Place them in ComfyUI/models/liveportrait/

Why This Mirror?

  • Pickle-free: safetensors format eliminates arbitrary code execution risks
  • Supply chain resilience: first-party mirror ensures availability
  • Faster loading: safetensors loads faster than pickle-based .pth

License

LivePortrait is released under the MIT License — see LICENSE.md. See the original repository for details.

Credits

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for AEmotionStudio/liveportrait-models