
SCAIL-2 Motion Transfer
SCAIL-2 is an open-source model for end-to-end controlled character animation. It uses a driving video to animate a reference character, supporting character replacement and multi-character scenes without relying on intermediate pose representations.

🔎 Overview
Previous character animation methods relied heavily on intermediate representations such as skeleton maps or inpainting masks. These intermediate representations suffer from ambiguity under complex motions, restrict driving sources to human actions, and hinder the development of character replacement and multi-character animation.
SCAIL-2 eliminates this dependency, enabling end-to-end driving. By utilizing multiple off-the-shelf models (SCAIL-Preview, Wan-Animate, MoCha), 60,000 pairs of motion data were synthesized and trained through a unified motion transfer interface (featuring a dedicated mask channel and RoPE design). Combined with a reverse-driving training strategy and a unified architecture, the model learns capabilities that surpass its teacher models, giving rise to the following features:
- Cross-identity character replacement
- Animal-driven scenes
- Zero-shot support for high-level control intermediate representations (such as SAM3D-Body mesh rendering)

📦 Models
| Item | Details |
|---|---|
| Resolution | End-to-end driving supports both 512p and 704p; pose-driven and character replacement perform better at 704p |
| Constraints | Both height H and width W must be divisible by 32 (e.g., 704×1280) |
| Training Setup | Mixed resolution and frame rate |
| Built-in Modules | Wan VAE and T5 are integrated into the Checkpoint for ease of use |
Downloaded file structure:
SCAIL-2/
├── Wan2.1_VAE.pth
├── model
│ ├── 1
│ │ └── fsdp2_rank_0000_Checkpoint.pt
│ └── latest
└── umt5-xxl
└── ...SCAIL 2 Action Transfer
SCAIL-2 Motion Transfer
SCAIL-2 is an open-source model for end-to-end controlled character animation. It uses a driving video to animate a reference character, supporting character replacement and multi-character scenes without relying on intermediate pose representations.

🔎 Overview
Previous character animation methods relied heavily on intermediate representations such as skeleton maps or inpainting masks. These intermediate representations suffer from ambiguity under complex motions, restrict driving sources to human actions, and hinder the development of character replacement and multi-character animation.
SCAIL-2 eliminates this dependency, enabling end-to-end driving. By utilizing multiple off-the-shelf models (SCAIL-Preview, Wan-Animate, MoCha), 60,000 pairs of motion data were synthesized and trained through a unified motion transfer interface (featuring a dedicated mask channel and RoPE design). Combined with a reverse-driving training strategy and a unified architecture, the model learns capabilities that surpass its teacher models, giving rise to the following features:
- Cross-identity character replacement
- Animal-driven scenes
- Zero-shot support for high-level control intermediate representations (such as SAM3D-Body mesh rendering)

📦 Models
| Item | Details |
|---|---|
| Resolution | End-to-end driving supports both 512p and 704p; pose-driven and character replacement perform better at 704p |
| Constraints | Both height H and width W must be divisible by 32 (e.g., 704×1280) |
| Training Setup | Mixed resolution and frame rate |
| Built-in Modules | Wan VAE and T5 are integrated into the Checkpoint for ease of use |
Downloaded file structure:
SCAIL-2/
├── Wan2.1_VAE.pth
├── model
│ ├── 1
│ │ └── fsdp2_rank_0000_Checkpoint.pt
│ └── latest
└── umt5-xxl
└── ...SCAIL-2 Motion Transfer
SCAIL-2 is an open-source model for end-to-end controlled character animation. It uses a driving video to animate a reference character, supporting character replacement and multi-character scenes without relying on intermediate pose representations.

🔎 Overview
Previous character animation methods relied heavily on intermediate representations such as skeleton maps or inpainting masks. These intermediate representations suffer from ambiguity under complex motions, restrict driving sources to human actions, and hinder the development of character replacement and multi-character animation.
SCAIL-2 eliminates this dependency, enabling end-to-end driving. By utilizing multiple off-the-shelf models (SCAIL-Preview, Wan-Animate, MoCha), 60,000 pairs of motion data were synthesized and trained through a unified motion transfer interface (featuring a dedicated mask channel and RoPE design). Combined with a reverse-driving training strategy and a unified architecture, the model learns capabilities that surpass its teacher models, giving rise to the following features:
- Cross-identity character replacement
- Animal-driven scenes
- Zero-shot support for high-level control intermediate representations (such as SAM3D-Body mesh rendering)

📦 Models
| Item | Details |
|---|---|
| Resolution | End-to-end driving supports both 512p and 704p; pose-driven and character replacement perform better at 704p |
| Constraints | Both height H and width W must be divisible by 32 (e.g., 704×1280) |
| Training Setup | Mixed resolution and frame rate |
| Built-in Modules | Wan VAE and T5 are integrated into the Checkpoint for ease of use |
Downloaded file structure:
SCAIL-2/
├── Wan2.1_VAE.pth
├── model
│ ├── 1
│ │ └── fsdp2_rank_0000_Checkpoint.pt
│ └── latest
└── umt5-xxl
└── ...