Research map / Video / Video generation
SurgVista: Long-Horizon Surgical World Modeling with Plausible Instrument-Tissue Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SurgVista, a surgical world model built on Wan2.1-1.3B that adds deformation consistency regularization and drift adaptation training for long-horizon action-conditioned surgical video prediction, plus the SurgWorld-Bench benchmark.
Problem
Surgical world models generate action-conditioned surgical video but suffer from spatial interaction incoherence (tissue fails to deform consistently under instrument contact) and temporal fidelity collapse (errors compound over autoregressive long rollouts).
Approach
Enforces cross-frame coherence by contrastive learning over tracked scene-point features in latent space (DCR), and simulates inference-time drift during training via online self-prediction residuals plus photometric augmentations on the conditioning frame (DAT).
Key results
- Short-horizon (80 frames): SurgVista reaches instrument AJ 45.76 and tissue AJ 41.68 vs. 39.34 and 36.19 for the best baseline Wan2.1-Control, with PSNR 19.85 vs 19.08.
- Long-horizon (800 frames): SurgVista achieves instrument AJ 27.13 and tissue AJ 24.15 vs. 18.92 and 19.91 for Wan2.1-Control; LPIPS 0.37 vs 0.42.
- Ablation (long-horizon): DAT alone raises instrument AJ from 19.45 to 26.54; adding DCR on top gives 27.13.
Surprising findings
- Error recycling from prior work gives only limited gains over clean-only training in the drift-strategy ablation (instrument AJ 24.15 vs 22.11), while the online self-prediction approach performs better.
Limitations
- Struggles with large-scale tissue deformations
- Struggles with extremely fine-grained manipulations
Experimental setup
Datasets: GraSP SurgToolLoc22 CholecTrack20
Benchmarks: SurgWorld-Bench
Metrics: AJ delta_avg_vis PSNR SSIM LPIPS Motion Smoothness Score
Baselines: SVD CogVideoX1.5-5B HunyuanVideo-1.5 Wan2.2-TI2V-5B Wan2.1-Control SurgSora
Builds on: Wan2.1 flow matching InfoNCE Error recycling Self-Forcing
Closest papers by meaning
- SurgicAI: A Hierarchical Platform for Fine-Grained Surgical Policy Learning and Benchmarking (NeurIPS 2024 · similarity 0.87)
- Bridging Vision and Language for Robust Context-Aware Surgical Point Tracking: The VL-SurgPT Dataset and Benchmark (AAAI 2026 · similarity 0.86)
- Multi-scale Temporal Prediction via Incremental Generation and Multi-agent Collaboration (NeurIPS 2025 · similarity 0.86)
- LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE (NeurIPS 2026 · similarity 0.86)
- SurgCoT: Advancing Spatiotemporal Reasoning in Surgical Videos through a Chain-of-Thought Benchmark (CVPR 2026 · similarity 0.85)
- CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.