SAMA-14B Masters Video Editing While Preserving Motion

By vramkickedin | | 2 min read
Close up shot of a digital video editing interface

SAMA-14B is a new open-source AI model designed for instruction-guided video editing. It allows users to modify videos using text instructions while keeping the original motion and temporal details intact.

The model was developed by a team from Baidu Inc, Tsinghua University, City University of Hong Kong, and Zhejiang University. It addresses a common challenge in video editing where models struggle to balance making precise changes while preserving the natural movement in footage.

Model Size: 14B parameters / 28.6 GB & VRAM GPU: requirements vary

Key features of SAMA-14B

  • Semantic Anchoring for instruction-aware structural planning.
  • Motion Alignment to preserve temporal dynamics.
  • Two-stage training pipeline for better semantic-motion representation.
  • Strong zero-shot video editing.
  • Official ComfyUI integration.

Video editors and content creators who need to make specific changes to footage will find this tool useful. It enables modifications through simple text prompts while maintaining the original motion, which helps with tasks like object replacement, style changes, or scene adjustments without needing to reshoot.

Technical approach and requirements

The developers designed SAMA to avoid relying on external features or structural conditions that can limit how well a model works across different scenarios. Instead, the framework splits video editing into two separate processes: semantic anchoring and motion modeling. The research paper notes that:

'the factorized pre-training alone already yields strong zero-shot video editing ability, validating the proposed factorization.'

Users will need the Wan2.1-T2V-14B base model for SAMA-14B to work. The installation requires Linux with an NVIDIA GPU running CUDA 12.1 and Python 3.10. Input frames are automatically padded to meet the 4k+1 frame requirement, and output videos use the source FPS when available.

Model access and further details