Zhipu AI / KJ Quantized Version

Wan21 14B SCAIL Preview

Action Transfer Model - Achieving Studio-Grade Character Animation

Model File: Wan21 14B SCAIL preview_fp8_scaled_mixed.safetensors

SCAIL (Studio-Grade Character Animation via In-Context Learning) model weights, a framework capable of achieving high-fidelity character animation under diverse and challenging conditions, including large motion variations, stylized characters, and multi-character interactions.

SCAIL: Towards Studio-Grade Character Animation via In-Context Learning of 3D Consistent Pose Representations

SCAIL: Towards Studio Grade Character Animation via In Context Learning of 3D Consistent Pose Representations

Paper link: teal024.github.io/SCAIL/

SCAIL Pose Project Summary

1. Core Project Positioning

SCAIL Pose is the companion codebase of the SCAIL (Studio Grade Character Animation via In-Context Learning) framework, focusing on 3D pose extraction and rendering. It enables high-fidelity character animation production in complex scenarios such as large motion variations, stylized characters, and multi-character interactions.

2. Core Features and Technical Characteristics

1. Core Capabilities

3D Pose Extraction: Estimates 3D human key points through skeletal topology connections, represents the skeleton as spatial cylinders, and rasterizes them to obtain 2D motion guidance signals;

Multi-Character Processing: Segments individual characters, extracts respective poses, and then renders them jointly, supporting multi-character pose extraction;

3D Pose Retargeting: Provides pose retargeting functionality to balance "identity anonymization" and "rich motion information retention" in pose representation;

Adaptable to Various Scenarios: Compatible with single/multi-character, full-body/portrait needs, supporting camera control (zooming, moving, principal point adjustment).

2. Technical Advantages

Leverages NLFPose for reliable depth estimation, enhancing pose estimation robustness in multi-character interaction scenarios;

Utilizes 3D consistent pose representation to avoid 2D-3D mismatches and noise issues, while achieving identity-independent motion retention and precise occlusion handling.