LLaVA-OneVision

Play with the model on the LLaVA OneVision Chat.

Table of Contents

1. Model Summary
2. Use
3. Limitations
4. Training
5. License
6. Citation

Model Summary

llava-onevision-72b-ov-chat is our latest model specifically designed for chat scenarios. It is built upon llava-onevision-72b-ov and has undergone iterative DPO training with human preference, making it well-suited for chat applications.

Research by Tianyi Xiong indicates that our iterative DPO training method enhances the model's chat capabilities while preserving its instruction-following abilities.

For further details, please refer to our upcoming blog or paper.

• Repository: LLaVA-VL/LLaVA-NeXT
• Project Website: llava-onevision.lmms-lab.com
• Paper: LLaVA-OneVision
• Point of Contact: Tianyi Xiong, Bo Li
• Languages: English, Chinese

Benchmark Performance

To be released

Use

Intended use

The model was trained on LLaVA-OneVision Dataset and have the ability to interact with images, multi-image and videos.

Feel free to share your generations in the Community tab!

Generation

# pip install git+https://github.com/LLaVA-VL/LLaVA-NeXT.git
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from llava.conversation import conv_templates, SeparatorStyle

from PIL import Image
import requests
import copy
import torch

import sys
import warnings

warnings.filterwarnings("ignore")
pretrained = "lmms-lab/llava-onevision-qwen2-0.5b-si"
model_name = "llava_qwen"
device = "cuda"
device_map = "auto"
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map)  # Add any other thing you want to pass in llava_model_args

model.eval()

url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
image = Image.open(requests.get(url, stream=True).raw)
image_tensor = process_images([image], image_processor, model.config)
image_tensor = [_image.to(dtype=torch.float16, device=device) for _image in image_tensor]

conv_template = "qwen_1_5"  # Make sure you use correct chat template for different models
question = DEFAULT_IMAGE_TOKEN + "\nWhat is shown in this image?"
conv = copy.deepcopy(conv_templates[conv_template])
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt_question = conv.get_prompt()

input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)
image_sizes = [image.size]


cont = model.generate(
    input_ids,
    images=image_tensor,
    image_sizes=image_sizes,
    do_sample=False,
    temperature=0,
    max_new_tokens=4096,
)
text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
print(text_outputs)

Training

Model

• Architecture: SO400M + Qwen2
• Pretraining Stage: LCS-558K, 1 epoch, projector
• Mid Stage: A mixture of 4.7M high-quality synthetic data, 1 epoch, full model
• Final-Image Stage: A mixture of 3.6M single-image data, 1 epoch, full model
• OneVision Stage: A mixture of 1.6M single-image/multi-image/video data, 1 epoch, full model
• Critic / Preference Learning Stage: 9.4k question-image input from LLaVA-RLHF with self-generated responses, reward signal from llava-critic-72b, iterative DPO for 3 rounds, full model
• Precision: bfloat16

Hardware & Software

• GPUs: 256 * Nvidia Tesla A100 (for whole model series training)
• Orchestration: Huggingface Trainer
• Neural networks: PyTorch

Citation

@article{li2024llavaonevision,
      title={LLaVA-OneVision: Easy Visual Task Transfer},
      author={Li, Bo and Zhang, Yuanhan and Guo, Dong and Zhang, Renrui and Li, Feng and Zhang, Hao and Zhang, Kaichen and Li, Yanwei and Liu, Ziwei and Li, Chunyuan},
      journal={arXiv preprint arXiv:2408.03326},
      year={2024}
}

@article{xiong2024llavacritic,
  title={LLaVA-Critic: Learning to Evaluate Multimodal Models},
  author={Xiong, Tianyi and Wang, Xiyao and Guo, Dong and Ye, Qinghao and Fan, Haoqi and Gu, Quanquan and Huang, Heng and Li, Chunyuan},
  year={2024},
  eprint={2410.02712},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2410.02712},
}