Inference-Time Concept Suppression and Video-Centric Evaluation for Text-to-Video Models
Researchers propose SIRUS, a framework for suppressing target concepts in text-to-video models. This allows for more controlled video generation without retraining the model.
- SIRUS is a training-free inference-time framework for concept-level unlearning in text-to-video models
- It allows for the suppression of target concepts across frames while preserving other aspects of the video
- SIRUS has the potential to improve the control and usability of text-to-video models
Text-to-video models have made significant progress in generating realistic videos, but controlling the output to exclude specific concepts has been a challenge.
The SIRUS framework addresses this by providing a training-free inference-time solution for concept-level unlearning. It works by localizing target-related prompt elements and suppressing the target concept across frames, while preserving other aspects of the video.
This development is important for applications where video content needs to be carefully controlled, such as in advertising, education, or social media. By enabling the suppression of specific concepts, SIRUS can help reduce the risk of unwanted or offensive content being generated.
The SIRUS framework has the potential to improve the overall quality and usability of text-to-video models, making them more suitable for a wide range of applications. Its ability to operate without requiring retraining of the model makes it a practical solution for real-world use cases.
enables more controlled video generation
improves video content control for applications like advertising and education
enhances video generation capabilities
- concept-level unlearning
- the ability to remove a specific concept from a generated video
- inference-time framework
- a system that operates during the generation phase, without requiring retraining of the model
Xi Jinping sets out China’s goal to be global AI leader - Financial Times
QFireNet: A Quantum-Enhanced U-Net for Wildfire Segmentation from Sentinel-2 Imagery
Certified Domain Consistency for Multi-Domain Retrieval: Label-Free Per-Domain Contamination Control with Conformal Risk Guarantees
TEDDY: A Pediatric Foundation Model for Risk Forewarning from ICD-Coded Diagnostic Histories
KeyFrame-Compass: Towards Comprehensive Evaluation of Keyframe-Conditioned Video Generation
China’s Xi says AI ‘should not be a solo performance by a single country’ - Al Jazeera
Chinese President Xi Jinping stated that artificial intelligence development should not be dominated by a single country, calling for broader international cooperation.
Organize your curiosity: Generative AI tools prove adept at structuring volumes of information - Editor and Publisher
Generative AI tools are effective in structuring large volumes of information, helping to organize and make sense of complex data. This development has significant implications for various industries and applications.
SeeSE3: Emergence of 3D Space in Vision Features
A new paper titled SeeSE3 investigates whether vision foundation models inherently represent 3D Euclidean space. The authors introduce probes to measure the relationship between visual features and Euclidean transformations.
3D Lane Detection with Odometry for High-Speed Vehicle Racing
Researchers introduce a new dataset for 3D lane detection in high-speed racing and compare various approaches to this challenging problem.
SD-MAR: Multi-image Analytical Reasoning via Synthetic Data and Reinforcement Learning
Researchers introduce SD-MAR, a method using synthetic data and reinforcement learning to enhance multi-image analytical reasoning in Vision Language Models.
MixCompress: Mixture of Experts for Variable Rate Learned Image Compression
Researchers propose MixCompress, a new framework for variable rate learned image compression, addressing limitations of existing methods.