Hugging Face's logo

OpenFace-CQUPT
/
Human_LLaVA

Visual Question Answering
Transformers
Safetensors
llava
image-text-to-text
AIGC
LLaVA
Inference Endpoints
Model card Files Community
3
Human_LLaVA / README.md
ponytail's picture
ponytail
Update README.md
a112ff5 verified
preview code
|
raw
history blame
5.37 kB
metadata
license: llama3
base_model: meta-llama/Meta-Llama-3-8B-Instruct
library_name: transformers
tags:
  - AIGC
  - LLaVA
datasets:
  - OpenFace-CQUPT/FaceCaption-15M
metrics:
  - accuracy
pipeline_tag: visual-question-answering

Human-LLaVA-8B

DEMO

image/png

Introduction

Human-related vision and language tasks are widely applied across various social scenarios. The latest studies demonstrate that the large vision-language model can enhance the performance of various downstream tasks in visual-language understanding. Since, models in the general domain often not perform well in the specialized field. In this study, we train a domain-specific Large Language-Vision model, Human-LLaVA, which aim to construct an unified multimodal Language-Vision Model for Human-related tasks.

Specifically, (1) we first construct a large-scale and high-quality human-related image-text (caption) dataset extracted from Internet for domain-specific alignment in the first stage (Coming soon); (2) we also propose to construct a multi-granularity caption for human-related images (Coming soon), including human face, human body, and whole image, thereby fine-tuning a large language model. Lastly, we evaluate our model on a series of downstream tasks, our Human-LLaVA achieved the best overall performance among multimodal models of similar scale. In particular, it exhibits the best performance in a series of human-related tasks, significantly surpassing similar models and ChatGPT-4o. We believe that the Huaman-LLaVA model and a series of datasets presented in this work can promote research in related fields.

Result

human-llava has a good performance in both general and special fields

image/png

News and Update 🔥🔥🔥

🤗 Transformers

To use Human-LLaVA for the inference, all you need to do is to input a few lines of codes as demonstrated below. However, please make sure that you are using latest code.

import requests
from PIL import Image

import torch
from transformers import AutoProcessor, AutoModelForPreTraining


model_id = "OpenFace-CQUPT/Human_LLaVA"
cuda = 0
model = AutoModelForPreTraining.from_pretrained("OpenFace-CQUPT/Human_LLaVA", torch_dtype=torch.float16).to(cuda)

processor = AutoProcessor.from_pretrained(model_id,trust_remote_code=True)


text = "Please describe this picture"
prompt = "USER: <image>\n" + text + "\nASSISTANT:"
image_file = "./test1.jpg"
raw_image = Image.open(image_file)
# raw_image = Image.open(requests.get(image_file, stream=True).raw)
inputs = processor(images=raw_image, text=prompt, return_tensors='pt').to(cuda, torch.float16)

output = model.generate(**inputs, max_new_tokens=400, do_sample=False)
predict = processor.decode(output[0][:], skip_special_tokens=True)
print(predict)

Our training code have been released publicly on github.ddw2AIGROUP2CQUPT/Human-LLaVA-8B(github.com)

Get the Dataset

Dataset Example

image/png

Domain Alignment Stage

HumanCaption-10M(self construct): is released!

Instruction Tuning Stage

All public data sets have been filtered, and we will consider publishing all processed text in the future

HumanCaption-HQ(self construct): is released!

FaceCaptionA(self construct): is released!

CelebA: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

ShareGPT4V:https://github.com/InternLM/InternLM-XComposer/blob/main/projects/ShareGPT4V/docs/Data.md

LLaVA-Instruct_zh : https://huggingface.co/datasets/openbmb/llava_zh

verified_ref3rec: https://huggingface.co/datasets/lucasjin/refcoco/blob/main/ref3rec.json

verified_ref3reg: https://huggingface.co/datasets/lucasjin/refcoco/blob/main/ref3rec.json

verified_shikra: https://github.com/shikras/shikra

Citation 

@misc{dai2024humanvlmfoundationhumanscenevisionlanguage,
      title={HumanVLM: Foundation for Human-Scene Vision-Language Model}, 
      author={Dawei Dai and Xu Long and Li Yutang and Zhang Yuanhui and Shuyin Xia},
      year={2024},
      eprint={2411.03034},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2411.03034}, 
}

contact

mailto: S230201133@stu.cqupt.edu.cn or dw_dai@163.com