Yi Vision Language Model

Better Bilingual Multimodal Model

🤗 Hugging Face • 🤖 ModelScope • ✡️ WiseModel

👩‍🚀 Ask questions or discuss ideas on GitHub !

👋 Join us 💬 WeChat (Chinese) !

📚 Grow at Yi Learning Hub !

📕 Table of Contents

What is Yi-VL?
- Overview
- Models
- Features
- Architecture
- Training
- Limitations
Why Yi-VL?
- Benchmarks
- Showcases
How to use Yi-VL?
- Quick start
- Hardware requirements
Misc.
- Acknowledgements and attributions
  - List of used open-source projects
- License

What is Yi-VL?

Overview

Yi Visual Language (Yi-VL) model is the open-source, multimodal version of the Yi Large Language Model (LLM) series, enabling content comprehension, recognition, and multi-round conversations about images.
Yi-VL demonstrates exceptional performance, ranking first among all existing open-source models in the latest benchmarks including MMMU in English and CMMMU in Chinese (based on data available up to January 2024).
Yi-VL-34B is the first open-source 34B vision language model worldwide.

Models

Yi-VL has released the following versions.

Model	Download
Yi-VL-34B	• 🤗 Hugging Face • 🤖 ModelScope
Yi-VL-6B	• 🤗 Hugging Face • 🤖 ModelScope

Features

Yi-VL offers the following features:

Multi-round text-image conversations: Yi-VL can take both text and images as inputs and produce text outputs. Currently, it supports multi-round visual question answering with one image.
Bilingual text support: Yi-VL supports conversations in both English and Chinese, including text recognition in images.
Strong image comprehension: Yi-VL is adept at analyzing visuals, making it an efficient tool for tasks like extracting, organizing, and summarizing information from images.
Fine-grained image resolution: Yi-VL supports image understanding at a higher resolution of 448×448.

Architecture

Yi-VL adopts the LLaVA architecture, which is composed of three primary components:

Vision Transformer (ViT): it's initialized with CLIP ViT-H/14 model and used for image encoding.
Projection Module: it's designed to align image features with text feature space, consisting of a two-layer Multilayer Perceptron (MLP) with layer normalizations.
Large Language Model (LLM): it's initialized with Yi-34B-Chat or Yi-6B-Chat, demonstrating exceptional proficiency in understanding and generating both English and Chinese.

Training

Training process

Yi-VL is trained to align visual information well to the semantic space of Yi LLM, which undergoes a comprehensive three-stage training process:

Stage 1: The parameters of ViT and the projection module are trained using an image resolution of 224×224. The LLM weights are frozen. The training leverages an image caption dataset comprising 100 million image-text pairs from LAION-400M. The primary objective is to enhance the ViT's knowledge acquisition within our specified architecture and to achieve better alignment between the ViT and the LLM.
Stage 2: The image resolution of ViT is scaled up to 448×448, and the parameters of ViT and the projection module are trained. It aims to further boost the model's capability for discerning intricate visual details. The dataset used in this stage includes about 25 million image-text pairs, such as LAION-400M, CLLaVA, LLaVAR, Flickr, VQAv2, RefCOCO, Visual7w and so on.
Stage 3: The parameters of the entire model (that is, ViT, projection module, and LLM) are trained. The primary goal is to enhance the model's proficiency in multimodal chat interactions, thereby endowing it with the ability to seamlessly integrate and interpret visual and linguistic inputs. To this end, the training dataset encompasses a diverse range of sources, totalling approximately 1 million image-text pairs, including GQA, VizWiz VQA, TextCaps, OCR-VQA, Visual Genome, LAION GPT4V and so on. To ensure data balancing, we impose a cap on the maximum data contribution from any single source, restricting it to no more than 50,000 pairs.

Below are the parameters configured for each stage.

Stage	Global batch size	Learning rate	Gradient clip	Epochs
Stage 1, 2	4096	1e-4	0.5	1
Stage 3	256	2e-5	1.0	2

Training resource consumption

The training consumes 128 NVIDIA A800 (80G) GPUs.
The total training time amounted to approximately 10 days for Yi-VL-34B and 3 days for Yi-VL-6B.

Limitations

This is the initial release of the Yi-VL, which comes with some known limitations. It is recommended to carefully evaluate potential risks before adopting any models.

Feature limitation
- Visual question answering is supported. Other features like text-to-3D and image-to-video are not yet supported.
- A single image rather than several images can be accepted as an input.
Hallucination problem
- There is a certain possibility of generating content that does not exist in the image.
- In scenes containing multiple objects, some objects might be incorrectly identified or described with insufficient detail.
Resolution issue
- Yi-VL is trained on images with a resolution of 448×448. During inference, inputs of any resolution are resized to 448×448. Low-resolution images may result in information loss, and more fine-grained images (above 448) do not bring in extra knowledge.
Other limitations of the Yi LLM.

Why Yi-VL?

Benchmarks

Yi-VL outperforms all existing open-source models in MMMU and CMMMU, two advanced benchmarks that include massive multi-discipline multimodal questions (based on data available up to January 2024).

MMMU

CMMMU

Showcases

Below are some representative examples of detailed description and visual question answering, showcasing the capabilities of Yi-VL.

English

Chinese

How to use Yi-VL?

Quick start

You can perform inference using the code from LLaVA. For detailed steps, see simple startup for pretraining.

Notes:

You need to modify the system prompt as follows.

This is a chat between an inquisitive human and an AI assistant. Assume the role of the AI assistant. Read all the images carefully, and respond to the human's questions with informative, helpful, detailed and polite answers. 这是一个好奇的人类和一个人工智能助手之间的对话。假设你扮演这个AI助手的角色。仔细阅读所有的图像，并对人类的问题做出信息丰富、有帮助、详细的和礼貌的回答。

### Human: <image_placeholder>
What is it in the image?
### Assistant:

You need to set the parameter mm_vision_tower in config.json to the local ViT path.

Hardware requirements

For model inference, the recommended GPU examples are:

Yi-VL-6B: RTX 3090, RTX 4090, A10, A30
Yi-VL-34B: 4 × RTX 4090, A800 (80 GB)

Misc.

Acknowledgements and attributions

This project makes use of open-source software/components. We acknowledge and are grateful to these developers for their contributions to the open-source community.

List of used open-source projects

LLaVA

Authors: Haotian Liu, Chunyuan Li, Qingyang Wu, Yuheng Li, and Yong Jae Lee
Source: https://github.com/haotian-liu/LLaVA
License: Apache-2.0 license
Description: The codebase is based on LLaVA code.

OpenClip

Authors: Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt
Source: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K
License: MIT
Description: The ViT is initialized using the weights of OpenClip.

Notes

This attribution does not claim to cover all open-source components used. Please check individual components and their respective licenses for full details.
The use of the open-source components is subject to the terms and conditions of the respective licenses.

We appreciate the open-source community for their invaluable contributions to the technology world.

License

Please refer to the acknowledgments and attributions as well as individual components, for the license of source code.

The Yi series models are fully open for academic research and free for commercial use, permissions of which are automatically granted upon application.

All usage must adhere to the Yi Series Models Community License Agreement 2.1.

For free commercial use, you only need to send an email to get official commercial permission.