Ovis2Translation site

What's Ovis2?

Ovis2是阿里巴巴国际化团队于2025年2月21日开源发布的新一代多模态macrolanguage model系列。作为Ovis1.6的继任者，Ovis2在数据构建和训练方法上进行了显著改进，旨在提升模型在多模态任务中的表现。

Ovis2 Key Features:

• High performance for small models: By optimizing the training strategy, Ovis2's small-scale model achieves higher capacity density and excels in cross-layer performance.
• Enhanced reasoning skills: The use of instruction fine-tuning and preference learning techniques substantially improves the chain-of-thinking (CoT) reasoning of the model.
• Video and multi-image processing: Video and multi-image processing capabilities have been introduced to enhance the model's understanding of dynamic and complex visual information.
• Multi-language OCR support: Expanded multi-language optical character recognition (OCR) capabilities to improve text extraction performance in complex scenarios.

Ovis2 model version:

The Ovis2 family consists of six versions, 1B, 2B, 4B, 8B, 16B and 34B, all of which are state-of-the-art (SOTA) at the same parameter scale. Among them, Ovis2-34B is ranked second among all open source models in the multimodal general capability list of OpenCompass, the authoritative review platform, surpassing many open source flagship models with 70B parameters at less than half the parameter scale.

Ovis2 architecture design:

Ovis2 utilizes an innovative architectural design aimed at structurally aligning visual and textual embeddings. Its core components include a Visual Segmenter, a Visual Embedding Table and a Large Language Model (LLM). The Visual Segmenter segments the input image into multiple image chunks, extracts features and maps them to "visual words" to generate probabilized visual tokens, the Visual Embedding Table stores the embedding vectors corresponding to each visual word, and the LLM splices together the visual and textual embedding vectors to generate textual outputs, thus completing the multimodal task.

Ovis2 training strategy:

Ovis2 uses a four-stage training method:

• Visual module training: Freeze most of the LLM and Visual Transformer (ViT) parameters and train the vision module to learn the visual feature-to-embedding transformation.
• Feature Extraction Enhancement: Further improve the feature extraction capability of the vision module, enhance high-resolution image understanding, multi-language support and OCR capability.
• Visual-Text Alignment: Aligning visual embedding with LLM's dialog format by visually describing data in dialog form.
• Multimodal instruction training and preference learning: Enhancing the model's ability to follow user commands and the quality of its output in multimodal situations.

Ovis2 keyframe selection algorithm:

To improve video comprehension, Ovis2 has developed an innovative key frame selection algorithm. The algorithm selects the most useful video frames based on the relevance of the frames to the text, the diversity between frames, and the sequentiality. Through high-dimensional conditional similarity computation, Deterministic Point Process (DPP) and Markov Decision Process (MDP), key frames are efficiently selected in limited visual contexts to enhance video comprehension performance.

Ovis2 open source information:

The code of Ovis2 has been open-sourced on GitHub, and the model is available on the Hugging Face and ModelScope platforms with online demos for user experience. Related research papers have also been published on arXiv for developers and researchers.

Ovis2 Related Links:

• GitHub code repository:https://github.com/AIDC-AI/Ovis
• Hugging Face model:https://huggingface.co/AIDC-AI/Ovis2-34B
• ModelScope model:https://modelscope.cn/collections/Ovis2-1e2840cb4f7d45
• Online Demo:https://huggingface.co/spaces/AIDC-AI/Ovis2-16B
• The arXiv paper:https://arxiv.org/abs/2405.20797

With these resources, developers and researchers can gain insights into Ovis2's architecture, training methods, and application scenarios, further advancing themultimodal macromodeldevelopment and innovation.