Joint-GRPO
A method that orchestrates the collaboration between a Vision-Language Model and a Video Diffusion Model to optimize their outputs based on a shared reward.
Category: General AI/ML
hallucination behaviors
The phenomenon where models generate outputs that are not grounded in the input data or reality.
Token Efficiency
A measure of how effectively a model utilizes tokens (or inputs) to produce outputs, impacting computational resource usage.
Video-Next-Event Prediction
A task that involves predicting the next event in a video given a procedural or predictive question, requiring dynamic video responses.
Dual-Mode Thinking
A cognitive strategy that allows a model to switch between quick, heuristic-based decision-making and slower, analytical reasoning.
XHuman
A dataset designed for testing human reconstruction algorithms.
HuGe100K
A large-scale dataset for human avatar reconstruction tasks.
THuman2.0
A dataset used for benchmarking human avatar reconstruction methods.
Ground-Truth Poses
The accurate and true poses of a subject, used as a reference in various computer vision tasks.
Sparse Input Reconstruction
A method of reconstructing models using a limited number of input images.