VideoGPT+ is a novel video conversation model that integrates image and video encoders to enhance video understanding. The model combines the strengths of image encoders for detailed spatial information and video encoders for global temporal context. It processes videos by dividing them into smaller segments and applying an adaptive pooling strategy on features extracted by both encoders. VideoGPT+ achieves improved performance across multiple video benchmarks, including VCGBench, MVBench, and Zero-shot question-answering. Additionally, the model uses a novel semi-automatic annotation pipeline to develop a 112K video-instruction set, further improving performance. The paper also introduces VCGBench-Diverse, a benchmark covering 18 broad video categories, which evaluates the generalization of video LMMs on dense video captioning, spatial and temporal understanding, and complex reasoning. The model's architecture includes a dual-encoder design, visual adapters that project vision features to the language domain, and a large language model for generating comprehensive responses. The model is trained on a combination of video-instruction datasets and achieves superior performance on various video benchmarks. The paper also presents ablation studies on different aspects of the model, including the type of vision encoder, feature pooling strategies, and LLM types. The results show that VideoGPT+ outperforms previous state-of-the-art approaches in multiple video understanding tasks.VideoGPT+ is a novel video conversation model that integrates image and video encoders to enhance video understanding. The model combines the strengths of image encoders for detailed spatial information and video encoders for global temporal context. It processes videos by dividing them into smaller segments and applying an adaptive pooling strategy on features extracted by both encoders. VideoGPT+ achieves improved performance across multiple video benchmarks, including VCGBench, MVBench, and Zero-shot question-answering. Additionally, the model uses a novel semi-automatic annotation pipeline to develop a 112K video-instruction set, further improving performance. The paper also introduces VCGBench-Diverse, a benchmark covering 18 broad video categories, which evaluates the generalization of video LMMs on dense video captioning, spatial and temporal understanding, and complex reasoning. The model's architecture includes a dual-encoder design, visual adapters that project vision features to the language domain, and a large language model for generating comprehensive responses. The model is trained on a combination of video-instruction datasets and achieves superior performance on various video benchmarks. The paper also presents ablation studies on different aspects of the model, including the type of vision encoder, feature pooling strategies, and LLM types. The results show that VideoGPT+ outperforms previous state-of-the-art approaches in multiple video understanding tasks.