Towards Suturing World Models (Wan, t2v)

This repository hosts the fine-tuned Wan2.1-T2V-1.3B text-to-video (t2v) diffusion model specialized for generating realistic robotic surgical suturing videos, capturing fine-grained sub-stitch actions including needle positioning, targeting, driving, and withdrawal. The model can differentiate between ideal and non-ideal surgical techniques, making it suitable for applications in surgical training, skill evaluation, and autonomous surgical system development.

Model Details

Base Model: Wan2.1-T2V-1.3B
Resolution: 768×512 pixels (Adjustable)
Frame Length: 49 frames per generated video (Adjustable)
Fine-tuning Method: Low-Rank Adaptation (LoRA)
Data Source: Annotated laparoscopic surgery exercise videos (∼2,000 clips)

Usage Example

import torch
from diffsynth import ModelManager, WanVideoPipeline, save_video, VideoData


model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cpu")
model_manager.load_models([
    "../Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors",
    "../Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth",
    "../Wan2.1-T2V-1.3B/Wan2.1_VAE.pth",
])
model_manager.load_lora("mehmetkeremturkcan/Suturing-Wan2.1-1.3B-T2V", lora_alpha=1.0)
pipe = WanVideoPipeline.from_model_manager(model_manager, device="cuda")
pipe.enable_vram_management(num_persistent_param_in_dit=None)

video = pipe(
    prompt="A needledrivingnonideal clip, generated from a backhand task.",
    num_inference_steps=50,
    tiled=True
)
save_video(video, "video.mp4", fps=30, quality=5)

Applications

Surgical Training: Generate demonstrations of both ideal and non-ideal surgical techniques for training purposes.
Skill Evaluation: Assess surgical skills by comparing actual procedures against model-generated standards.
Robotic Automation: Inform autonomous surgical robotic systems for real-time guidance and procedure automation.

Quantitative Performance

Metric	Performance
L2 Reconstruction Loss	0.0667
Inference Time	~360 seconds per video

Future Directions

Further improvements will focus on increasing model robustness, expanding the dataset diversity, and enhancing real-time applicability to robotic surgical scenarios.