The Imp project aims to provide a family of highly capable yet lightweight LMMs. Our Imp-v1.5-3B-Phi2 is a strong lightweight LMMs with only 3B parameters, which is build upon Phi-2 (2.7B) and a powerful visual encoder SigLIP (0.4B), and trained on 1M mixed dataset.

As shown in the Table below, Imp-v1.5-3B-Phi2 significantly outperforms the counterparts of similar model sizes, and even achieves slightly better performance than the strong LLaVA-7B model on various multimodal benchmarks.

We release our model weights and provide an example below to run our model . Detailed technical report and corresponding training/evaluation code will be released soon on our GitHub repo. We will persistently improve our model and release the next versions to further improve model performance :)

How to use

Install dependencies

bash

pip install transformers # latest version is ok, but we recommend v4.37.0
pip install -q pillow accelerate einops

You can use the following code for model inference. The format of text instruction is similar to LLaVA. Note that the example can only be run on GPUs currently.

Python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image

torch.set_default_device("cuda")

#Create model
model = AutoModelForCausalLM.from_pretrained(
    "MILVLG/Imp-v1.5-3B-Phi2/", 
    torch_dtype=torch.float16, 
    device_map="auto",
    trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("MILVLG/Imp-v1.5-3B-Phi2", trust_remote_code=True)

#Set inputs
text = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat are the colors of the bus in the image? ASSISTANT:"
image = Image.open("images/bus.jpg")

input_ids = tokenizer(text, return_tensors='pt').input_ids
image_tensor = model.image_preprocess(image)

#Generate the answer
output_ids = model.generate(
    input_ids,
    max_new_tokens=100,
    images=image_tensor,
    use_cache=True)[0]
print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())

Model evaluation

We conduct evaluation on 9 commonly-used benchmarks, including 5 academic VQA benchmarks and 4 popular MLLM benchmarks, to compare our Imp model with LLaVA (7B) and existing lightweight LMMs of similar model sizes.

Models	Size	VQAv2	GQA	SQA(IMG)	TextVQA	POPE	MME(P)	MMB	MMBCN	MM-Vet
LLaVA-v1.5-lora	7B	79.1	63.0	68.4	58.2	86.4	1476.9	66.1	-	30.2
TinyGPT-V-3B	3B	-	38.9	-	-	-	-	-	-	-
LaVA-Phi-3B	3B	71.4	-	68.4	48.6	85.0	1335.1	59.8	-	28.9
MobileVLM-3B	3B	-	59.0	61.0	47.5	84.9	1288.9	59.6	-	-
MiniCPM-V-3B	3B	-	-	-	-	-	1452.0	67.9	65.3	-
Bunny-3B	3B	79.8	62.5	70.9	-	86.8	1488.8	68.6	-	-
Imp-v1.5-3B-Phi2	3B	81.2	63.5	72.8	59.8	88.9	1446.4	72.9	46.7	43.3

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Citation

If you use our model or refer our work in your studies, please cite:

bibtex

@article{imp2024,
  title={Imp: Highly Capable Large Multimodal Models for Mobile Devices},
  author={Shao, Zhenwei and Yu, Zhou and Yu, Jun and Ouyang, Xuecheng and Zheng, Lihao and Gai, Zhenbiao and Wang, Mingyang and Ding, Jiajun},
  journal={arXiv preprint arXiv:2405.12107},
  year={2024}
}