Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks

Model Summary

This Hub repository contains a HuggingFace's transformers implementation of Florence-2 model from Microsoft.

Florence-2 is an advanced vision foundation model that uses a prompt-based approach to handle a wide range of vision and vision-language tasks. Florence-2 can interpret simple text prompts to perform tasks like captioning, object detection, and segmentation. It leverages our FLD-5B dataset, containing 5.4 billion annotations across 126 million images, to master multi-task learning. The model's sequence-to-sequence architecture enables it to excel in both zero-shot and fine-tuned settings, proving to be a competitive vision foundation model.

Resources and Technical Documentation:

Model	Model size	Model Description
Florence-2-base[HF]	0.23B	Pretrained model with FLD-5B
Florence-2-large[HF]	0.77B	Pretrained model with FLD-5B
Florence-2-base-ft[HF]	0.23B	Finetuned model on a colletion of downstream tasks
Florence-2-large-ft[HF]	0.77B	Finetuned model on a colletion of downstream tasks

How to Get Started with the Model

Use the code below to get started with the model.

import requests

from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM 


model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)

prompt = "<OD>"

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(text=prompt, images=image, return_tensors="pt")

generated_ids = model.generate(
    input_ids=inputs["input_ids"],
    pixel_values=inputs["pixel_values"],
    max_new_tokens=1024,
    num_beams=3,
    do_sample=False
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]

parsed_answer = processor.post_process_generation(generated_text, task="<OD>", image_size=(image.width, image.height))

print(parsed_answer)

Tasks

This model is capable of performing different tasks through changing the prompts.

First, let's define a function to run a prompt.

Click to expand

import requests

from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM 


model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

def run_example(task_prompt, text_input=None):
    if text_input is None:
        prompt = task_prompt
    else:
        prompt = task_prompt + text_input
    inputs = processor(text=prompt, images=image, return_tensors="pt")
    generated_ids = model.generate(
      input_ids=inputs["input_ids"],
      pixel_values=inputs["pixel_values"],
      max_new_tokens=1024,
      num_beams=3
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]

    parsed_answer = processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))

    print(parsed_answer)

Here are the tasks Florence-2 could perform:

Click to expand

Caption

prompt = "<CAPTION>"
run_example(prompt)

Detailed Caption

prompt = "<DETAILED_CAPTION>"
run_example(prompt)

More Detailed Caption

prompt = "<MORE_DETAILED_CAPTION>"
run_example(prompt)

Caption to Phrase Grounding

caption to phrase grounding task requires additional text input, i.e. caption.

Caption to phrase grounding results format: {'<CAPTION_TO_PHRASE_GROUNDING>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['', '', ...]}}

task_prompt = "<CAPTION_TO_PHRASE_GROUNDING>"
results = run_example(task_prompt, text_input="A green car parked in front of a yellow building.")

Object Detection

OD results format: {'<OD>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['label1', 'label2', ...]} }

prompt = "<OD>"
run_example(prompt)

Dense Region Caption

Dense region caption results format: {'<DENSE_REGION_CAPTION>' : {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['label1', 'label2', ...]} }

prompt = "<DENSE_REGION_CAPTION>"
run_example(prompt)

Region proposal

Dense region caption results format: {'<REGION_PROPOSAL>': {'bboxes': [[x1, y1, x2, y2], ...], 'labels': ['', '', ...]}}

prompt = "<REGION_PROPOSAL>"
run_example(prompt)

OCR

prompt = "<OCR>"
run_example(prompt)

OCR with Region

OCR with region output format: {'<OCR_WITH_REGION>': {'quad_boxes': [[x1, y1, x2, y2, x3, y3, x4, y4], ...], 'labels': ['text1', ...]}}

prompt = "<OCR_WITH_REGION>"
run_example(prompt)

for More detailed examples, please refer to notebook

Benchmarks

Florence-2 Zero-shot performance

The following table presents the zero-shot performance of generalist vision foundation models on image captioning and object detection evaluation tasks. These models have not been exposed to the training data of the evaluation tasks during their training phase.

Method	#params	COCO Cap. test CIDEr	NoCaps val CIDEr	TextCaps val CIDEr	COCO Det. val2017 mAP
Flamingo	80B	84.3	-	-	-
Florence-2-base	0.23B	133.0	118.7	70.1	34.7
Florence-2-large	0.77B	135.6	120.8	72.8	37.5

The following table continues the comparison with performance on other vision-language evaluation tasks.

Method	Flickr30k test R@1	Refcoco val Accuracy	Refcoco test-A Accuracy	Refcoco test-B Accuracy	Refcoco+ val Accuracy	Refcoco+ test-A Accuracy	Refcoco+ test-B Accuracy	Refcocog val Accuracy	Refcocog test Accuracy	Refcoco RES val mIoU
Kosmos-2	78.7	52.3	57.4	47.3	45.5	50.7	42.2	60.6	61.7	-
Florence-2-base	83.6	53.9	58.4	49.7	51.5	56.4	47.9	66.3	65.1	34.6
Florence-2-large	84.4	56.3	61.6	51.4	53.6	57.9	49.9	68.0	67.0	35.8

Florence-2 finetuned performance

We finetune Florence-2 models with a collection of downstream tasks, resulting two generalist models Florence-2-base-ft and Florence-2-large-ft that can conduct a wide range of downstream tasks.

The table below compares the performance of specialist and generalist models on various captioning and Visual Question Answering (VQA) tasks. Specialist models are fine-tuned specifically for each task, whereas generalist models are fine-tuned in a task-agnostic manner across all tasks. The symbol "▲" indicates the usage of external OCR as input.

Method	# Params	COCO Caption Karpathy test CIDEr	NoCaps val CIDEr	TextCaps val CIDEr	VQAv2 test-dev Acc	TextVQA test-dev Acc	VizWiz VQA test-dev Acc
Specialist Models
CoCa	2.1B	143.6	122.4	-	82.3	-	-
BLIP-2	7.8B	144.5	121.6	-	82.2	-	-
GIT2	5.1B	145.0	126.9	148.6	81.7	67.3	71.0
Flamingo	80B	138.1	-	-	82.0	54.1	65.7
PaLI	17B	149.1	127.0	160.0▲	84.3	58.8 / 73.1▲	71.6 / 74.4▲
PaLI-X	55B	149.2	126.3	147.0 / 163.7▲	86.0	71.4 / 80.8▲	70.9 / 74.6▲
Generalist Models
Unified-IO	2.9B	-	100.0	-	77.9	-	57.4
Florence-2-base-ft	0.23B	140.0	116.7	143.9	79.7	63.6	63.6
Florence-2-large-ft	0.77B	143.3	124.9	151.1	81.7	73.5	72.6

Method	# Params	COCO Det. val2017 mAP	Flickr30k test R@1	RefCOCO val Accuracy	RefCOCO test-A Accuracy	RefCOCO test-B Accuracy	RefCOCO+ val Accuracy	RefCOCO+ test-A Accuracy	RefCOCO+ test-B Accuracy	RefCOCOg val Accuracy	RefCOCOg test Accuracy	RefCOCO RES val mIoU
Specialist Models
SeqTR	-	-	-	83.7	86.5	81.2	71.5	76.3	64.9	74.9	74.2	-
PolyFormer	-	-	-	90.4	92.9	87.2	85.0	89.8	78.0	85.8	85.9	76.9
UNINEXT	0.74B	60.6	-	92.6	94.3	91.5	85.2	89.6	79.8	88.7	89.4	-
Ferret	13B	-	-	89.5	92.4	84.4	82.8	88.1	75.2	85.8	86.3	-
Generalist Models
UniTAB	-	-	-	88.6	91.1	83.8	81.0	85.4	71.6	84.6	84.7	-
Florence-2-base-ft	0.23B	41.4	84.0	92.6	94.8	91.5	86.8	91.7	82.2	89.8	82.2	78.0
Florence-2-large-ft	0.77B	43.4	85.2	93.4	95.3	92.0	88.3	92.9	83.6	91.2	91.7	80.5

BibTex and citation info

@article{xiao2023florence,
  title={Florence-2: Advancing a unified representation for a variety of vision tasks},
  author={Xiao, Bin and Wu, Haiping and Xu, Weijian and Dai, Xiyang and Hu, Houdong and Lu, Yumao and Zeng, Michael and Liu, Ce and Yuan, Lu},
  journal={arXiv preprint arXiv:2311.06242},
  year={2023}
}