InternLM-XComposer2

InternLM-XComposer2 🤗  ｜ InternLM-XComposer2-VL 🤗   | Technical Report 📄

English | 简体中文

Thanks the community for HuggingFace Demo | OpenXLab Demo of InternLM-XComposer2.

👋 join us on Discord and WeChat

Multimodal Projects of Our Team

InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Models

InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition

ShareGPT4V: Improving Large Multi-modal Models with Better Captions

DualFocus: Integrating Macro and Micro Perspectives in Multi-modal Large Language Models

InternLM-XComposer2 is a groundbreaking vision-language large model (VLLM) based on InternLM2-7B excelling in free-form text-image composition and comprehension. It boasts several amazing capabilities and applications:

• Free-form Interleaved Text-Image Composition: InternLM-XComposer2 can effortlessly generate coherent and contextual articles with interleaved images following diverse inputs like outlines, detailed text requirements and reference images, enabling highly customizable content creation.

• Accurate Vision-language Problem-solving: InternLM-XComposer2 accurately handles diverse and challenging vision-language Q&A tasks based on free-form instructions, excelling in recognition, perception, detailed captioning, visual reasoning, and more.

• Awesome performance: InternLM-XComposer2 based on InternLM2-7B not only significantly outperforms existing open-source multimodal models in 13 benchmarks but also matches or even surpasses GPT-4V and Gemini Pro in 6 benchmarks

We release InternLM-XComposer2 series in two versions:

• InternLM-XComposer2-VL-7B 🤗 : The multi-task trained VLLM model with InternLM-7B as the initialization of the LLM for VL benchmarks and AI assistant. It ranks as the most powerful vision-language model based on 7B-parameter level LLMs, leading across 13 benchmarks.

• InternLM-XComposer2-7B 🤗 : The further instruction tuned VLLM for Interleaved Text-Image Composition with free-form inputs.

Please refer to Technical Report for more details.

Demo Video

InternLM-XComposer2_cps.mp4

Please refer to Chinese Demo for the demo of the Chinese version.

News and Updates

• 2024.02.22 🎉🎉🎉 We release DualFocus, a framework for integrating macro and micro perspectives within MLLMs to enhance vision-language task performance.

• 2024.02.06 🎉🎉🎉 InternLM-XComposer2-7B-4bit and InternLM-XComposer-VL2-7B-4bit are publicly available on Hugging Face and ModelScope.

• 2024.02.02 🎉🎉🎉 The finetune code of InternLM-XComposer2-VL-7B are publicly available.
• 2024.01.26 🎉🎉🎉 The evaluation code of InternLM-XComposer2-VL-7B are publicly available.
• 2024.01.26 🎉🎉🎉 InternLM-XComposer2-7B and InternLM-XComposer-VL2-7B are publicly available on Hugging Face and ModelScope.
• 2024.01.26 🎉🎉🎉 We release a technical report for more details of InternLM-XComposer2 series.
• 2023.11.22 🎉🎉🎉 We release the ShareGPT4V, a large-scale highly descriptive image-text dataset generated by GPT4-Vision and a superior large multimodal model, ShareGPT4V-7B.
• 2023.10.30 🎉🎉🎉 InternLM-XComposer-VL achieved the top 1 ranking in both Q-Bench and Tiny LVLM.
• 2023.10.19 🎉🎉🎉 Support for inference on multiple GPUs. Two 4090 GPUs are sufficient for deploying our demo.
• 2023.10.12 🎉🎉🎉 4-bit demo is supported, model files are available in Hugging Face and ModelScope.
• 2023.10.8 🎉🎉🎉 InternLM-XComposer-7B and InternLM-XComposer-VL-7B are publicly available on ModelScope.
• 2023.9.27 🎉🎉🎉 The evaluation code of InternLM-XComposer-VL-7B are publicly available.
• 2023.9.27 🎉🎉🎉 InternLM-XComposer-7B and InternLM-XComposer-VL-7B are publicly available on Hugging Face.
• 2023.9.27 🎉🎉🎉 We release a technical report for more details of our model series.
  

Model Zoo

Model	Usage	Transformers(HF)	ModelScope(HF)	Release Date
InternLM-XComposer2	Text-Image Composition	🤗internlm-xcomposer2-7b	internlm-xcomposer2-7b	2024-01-26
InternLM-XComposer2-VL	Benchmark, VL-Chat	🤗internlm-xcomposer2-vl-7b	internlm-xcomposer2-vl-7b	2024-01-26
InternLM-XComposer2-4bit	Text-Image Composition	🤗internlm-xcomposer2-7b-4bit	internlm-xcomposer2-7b-4bit	2024-02-06
InternLM-XComposer2-VL-4bit	Benchmark, VL-Chat	🤗internlm-xcomposer2-vl-7b-4bit	internlm-xcomposer2-vl-7b-4bit	2024-02-06
InternLM-XComposer	Text-Image Composition, VL-Chat	🤗internlm-xcomposer-7b	internlm-xcomposer-7b	2023-09-26
InternLM-XComposer-4bit	Text-Image Composition, VL-Chat	🤗internlm-xcomposer-7b-4bit	internlm-xcomposer-7b-4bit	2023-09-26
InternLM-XComposer-VL	Benchmark	🤗internlm-xcomposer-vl-7b	internlm-xcomposer-vl-7b	2023-09-26

Evaluation

We evaluate InternLM-XComposer2-VL on 13 multimodal benchmarks: MathVista, MMMU, AI2D, MME, MMBench, MMBench-CN, SEED-Bench, QBench, HallusionBench, ChartQA, MM-Vet, LLaVA-in-the-wild, POPE.

See Evaluation Details here.

Compared with closed-source APIs and previous SOTAs.

	MathVista	AI2D	MMMU	MME	MMB	MMBCN	SEEDI	LLaVAW	QBenchT	MM-Vet	HallB	ChartVQA
Open-source Previous SOTA	SPH-MOE	Monkey	Yi-VL	WeMM	L-Int2	L-Int2	SPH-2	CogVLM	Int-XC	CogVLM	Monkey	CogAgent
	8x7B	10B	34B	6B	20B	20B	17B	17B	8B	30B	10B	18B
	42.3	72.6	45.9	2066.6	75.1	73.7	74.8	73.9	64.4	56.8	58.4	68.4
GPT-4V	49.9	78.2	56.8	1926.5	77	74.4	69.1	93.1	74.1	67.7	65.8	78.5
Gemini-Pro	45.2	73.9	47.9	1933.3	73.6	74.3	70.7	79.9	70.6	64.3	63.9	74.1
QwenVL-Plus	43.3	75.9	46.5	2183.3	67	70.7	72.7	73.7	68.9	55.7	56.4	78.1
InternLM-XComposer2-VL	57.6	78.7	42	2242.7	79.6	77.6	75.9	81.8	72.5	51.2	60.3	72.6

Compared with open-source methods.

Method	LLM	MathVista	MMMU	MMEP	MMEC	MMB	MMBCN	SEEDI	LLaVAW	QBenchT	MM-Vet	HallB	POPE
BLIP-2	FLAN-T5	-	35.7	1,293.8	290.0	-	-	46.4	38.1	-	22.4	-	-
InstructBLIP	Vicuna-7B	25.3	30.6	-	-	36.0	23.7	53.4	60.9	55.9	26.2	53.6	78.9
IDEFICS-80B	LLaMA-65B	26.2	24.0	-	-	54.5	38.1	52.0	56.9	-	39.7	46.1	-
Qwen-VL-Chat	Qwen-7B	33.8	35.9	1,487.5	360.7	60.6	56.7	58.2	67.7	61.7	47.3	56.4	-
LLaVA	Vicuna-7B	23.7	32.3	807.0	247.9	34.1	14.1	25.5	63.0	54.7	26.7	44.1	80.2
LLaVA-1.5	Vicuna-13B	26.1	36.4	1,531.3	295.4	67.7	63.6	68.2	70.7	61.4	35.4	46.7	85.9
ShareGPT4V	Vicuna-7B	25.8	36.6	1,567.4	376.4	68.8	62.2	69.7	72.6	-	37.6	49.8	-
CogVLM-17B	Vicuna-7B	34.7	37.3	-	-	65.8	55.9	68.8	73.9	-	54.5	55.1	-
LLaVA-XTuner	InernLM2-20B	24.6	39.4	-	-	75.1	73.7	70.2	63.7	-	37.2	47.7	-
Monkey-10B	Qwen-7B	34.8	40.7	1,522.4	401.4	72.4	67.5	68.9	33.5	-	33.0	58.4	-
InternLM-XComposer	InernLM-7B	29.5	35.6	1,528.4	391.1	74.4	72.4	66.1	53.8	64.4	35.2	57.0	-
InternLM-XComposer2-VL	InernLM2-7B	57.6	43.0	1,712.0	530.7	79.6	77.6	75.9	81.8	72.5	51.2	59.1	87.7

Requirements

• python 3.8 and above
• pytorch 1.12 and above, 2.0 and above are recommended
• CUDA 11.4 and above are recommended (this is for GPU users)
  

Installation

Before running the code, make sure you have setup the environment and installed the required packages. Make sure you meet the above requirements, and then install the dependent libraries. Please refer to the installation instructions

Quickstart

We provide a simple example to show how to use InternLM-XComposer with 🤗 Transformers.

🤗 Transformers

import torch
from transformers import AutoModel, AutoTokenizer

torch.set_grad_enabled(False)

# init model and tokenizer
model = AutoModel.from_pretrained('internlm/internlm-xcomposer2-vl-7b', trust_remote_code=True).cuda().eval()
tokenizer = AutoTokenizer.from_pretrained('internlm/internlm-xcomposer2-vl-7b', trust_remote_code=True)

text = '<ImageHere>Please describe this image in detail.'
image = 'examples/image1.webp'
with torch.cuda.amp.autocast():
  response, _ = model.chat(tokenizer, query=text, image=image, history=[], do_sample=False)
print(response)
#The image features a quote by Oscar Wilde, "Live life with no excuses, travel with no regret,"
# set against a backdrop of a breathtaking sunset. The sky is painted in hues of pink and orange,
# creating a serene atmosphere. Two silhouetted figures stand on a cliff, overlooking the horizon.
# They appear to be hiking or exploring, embodying the essence of the quote.
# The overall scene conveys a sense of adventure and freedom, encouraging viewers to embrace life without hesitation or regrets.

🤖 ModelScope

import torch
from modelscope import snapshot_download, AutoModel, AutoTokenizer

torch.set_grad_enabled(False)

# init model and tokenizer
model_dir = snapshot_download('Shanghai_AI_Laboratory/internlm-xcomposer2-vl-7b')
model = AutoModel.from_pretrained(model_dir, trust_remote_code=True).cuda().eval()
tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
model.tokenizer = tokenizer

text = '<ImageHere>Please describe this image in detail.'
image = 'examples/image1.webp'
with torch.cuda.amp.autocast():
  response, _ = model.chat(tokenizer, query=text, image=image, history=[], do_sample=False)
print(response)
#The image features a quote by Oscar Wilde, "Live life with no excuses, travel with no regret,"
# set against a backdrop of a breathtaking sunset. The sky is painted in hues of pink and orange,
# creating a serene atmosphere. Two silhouetted figures stand on a cliff, overlooking the horizon.
# They appear to be hiking or exploring, embodying the essence of the quote.
# The overall scene conveys a sense of adventure and freedom, encouraging viewers to embrace life without hesitation or regrets.

Inference on Multiple GPUs

If you have multiple GPUs, but the memory size of each GPU is not enough to accommodate the entire model, you can split the model across multiple GPUs. First, install accelerate using the command: pip install accelerate. Then, execute the follows scripts for chat:

# chat with 2 GPUs
python examples/example_chat.py --num_gpus 2

4-Bit Model

We provide 4-bit quantized models to ease the memory requirement of the models. To run the 4-bit models (GPU memory >= 12GB), you need first install the corresponding dependency, then execute the follows scripts for chat:

🤗 Transformers

import torch, auto_gptq
from transformers import AutoModel, AutoTokenizer 
from auto_gptq.modeling._base import BaseGPTQForCausalLM

auto_gptq.modeling._base.SUPPORTED_MODELS = ["internlm"]
torch.set_grad_enabled(False)

class InternLMXComposer2QForCausalLM(BaseGPTQForCausalLM):
    layers_block_name = "model.layers"
    outside_layer_modules = [
        'vit', 'vision_proj', 'model.tok_embeddings', 'model.norm', 'output', 
    ]
    inside_layer_modules = [
        ["attention.wqkv.linear"],
        ["attention.wo.linear"],
        ["feed_forward.w1.linear", "feed_forward.w3.linear"],
        ["feed_forward.w2.linear"],
    ]
 

=======
# init model and tokenizer
model = InternLMXComposer2QForCausalLM.from_quantized(
  'internlm/internlm-xcomposer2-vl-7b-4bit', trust_remote_code=True, device="cuda:0").eval()
tokenizer = AutoTokenizer.from_pretrained(
  'internlm/internlm-xcomposer2-vl-7b-4bit', trust_remote_code=True)

text = '<ImageHere>Please describe this image in detail.'
image = 'examples/image1.webp'
with torch.cuda.amp.autocast(): 
  response, _ = model.chat(tokenizer, query=text, image=image, history=[], do_sample=False) 
print(response)
#The image features a quote by Oscar Wilde, "Live life with no excuses, travel with no regrets." 
#The quote is displayed in white text against a dark background. In the foreground, there are two silhouettes of people standing on a hill at sunset. 
#They appear to be hiking or climbing, as one of them is holding a walking stick. 
#The sky behind them is painted with hues of orange and purple, creating a beautiful contrast with the dark figures.

Finetune

Please refer to our finetune scripts.

Web UI

Thanks the community for 3rd-party HuggingFace Demo

We provide code for users to build a web UI demo.

Please run the command below for Composition / Chat:

# For Free-form Text-Image Composition
python examples/gradio_demo_composition.py

# For Multimodal Chat
python examples/gradio_demo_chat.py

The user guidance of UI demo is given in HERE. If you wish to change the default folder of the model, please use the --folder=new_folder option.

Citation

If you find our models / code / papers useful in your research, please consider giving stars ⭐ and citations 📝, thx :)

@article{internlmxcomposer2,
      title={InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model},
      author={Xiaoyi Dong and Pan Zhang and Yuhang Zang and Yuhang Cao and Bin Wang and Linke Ouyang and Xilin Wei and Songyang Zhang and Haodong Duan and Maosong Cao and Wenwei Zhang and Yining Li and Hang Yan and Yang Gao and Xinyue Zhang and Wei Li and Jingwen Li and Kai Chen and Conghui He and Xingcheng Zhang and Yu Qiao and Dahua Lin and Jiaqi Wang},
      journal={arXiv preprint arXiv:2401.16420},
      year={2024}
}

@article{internlmxcomposer,
      title={InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition},
      author={Pan Zhang and Xiaoyi Dong and Bin Wang and Yuhang Cao and Chao Xu and Linke Ouyang and Zhiyuan Zhao and Shuangrui Ding and Songyang Zhang and Haodong Duan and Wenwei Zhang and Hang Yan and Xinyue Zhang and Wei Li and Jingwen Li and Kai Chen and Conghui He and Xingcheng Zhang and Yu Qiao and Dahua Lin and Jiaqi Wang},
      journal={arXiv preprint arXiv:2309.15112},
      year={2023}
}

License & Contact Us

The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow free commercial usage. To apply for a commercial license, please fill in the application form (English)/申请表（中文）. For other questions or collaborations, please contact internlm@pjlab.org.cn.