Visual AI Digest: Spatial Grounding and Compositional Video

ConsistI2V: Temporal-Consistent Image-to-Video Generation

This paper introduces ConsistI2V, a framework for improving temporal consistency in image-to-video generation without finetuning. It's a significant step towards generating smoother, more coherent videos from static images using existing T2V models.

Spatially-Grounded Text-to-Video Generation

Focuses on enabling precise spatial control (like landmarks or edges) in text-to-video models, moving beyond pure text prompts for finer directorial control.

Improving Object Completeness in Text-to-Video Synthesis

Tackles the critical issue of incomplete object parts (like cut-off heads) in generated videos, proposing a method to improve compositional understanding in T2V models.

Complex Textual Prompt Guided Video Generation

Introduces a new benchmark and method for evaluating and improving the semantic alignment between complex text prompts and generated video content.

Example-Guided Video Generation

Explores a method to guide and control video generation using a single example video, offering a more intuitive way to specify desired motion or style than text alone.