New SIRUS framework enables concept removal in text-to-video models

Researchers have developed SIRUS, a novel framework designed to remove specific concepts from text-to-video (T2V) models at inference time without requiring model retraining. This method localizes and suppresses target concepts across frames while preserving non-target elements, temporal coherence, and overall video quality. A new video-centric evaluation framework was also introduced to measure concept forgetting, non-target preservation, and video quality, demonstrating SIRUS's superior performance over existing methods like VideoEraser on models such as CogVideoX and Wan2.2. AI

IMPACT Enables finer control over AI-generated video content, potentially improving safety and customization.

RANK_REASON Research paper detailing a new method for text-to-video models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CogVideoX
• Hugging Face
• SIRUS
• text-to-video (T2V) models
• VideoEraser
• Wan2.2

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →