lucataco/florence-2-large

# lucataco/florence-2-large – Replicate Source: [https://replicate.com/lucataco/florence-2-large](https://replicate.com/lucataco/florence-2-large) ## 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: \+[Florence\-2 technical report](https://arxiv.org/abs/2311.06242)\. \+[Jupyter Notebook for inference and visualization of Florence\-2\-large](https://huggingface.co/microsoft/Florence-2-large/blob/main/sample_inference.ipynb) ModelModel sizeModel DescriptionFlorence\-2\-base[\[HF\]](https://huggingface.co/microsoft/Florence-2-base)0\.23BPretrained model with FLD\-5BFlorence\-2\-large[\[HF\]](https://huggingface.co/microsoft/Florence-2-large)0\.77BPretrained model with FLD\-5BFlorence\-2\-base\-ft[\[HF\]](https://huggingface.co/microsoft/Florence-2-base-ft)0\.23BFinetuned model on a colletion of downstream tasksFlorence\-2\-large\-ft[\[HF\]](https://huggingface.co/microsoft/Florence-2-large-ft)0\.77BFinetuned model on a colletion of downstream tasks## Tasks This model is capable of performing different tasks through changing the prompts\. ### 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](https://huggingface.co/microsoft/Florence-2-large/blob/main/sample_inference.ipynb)</details\> ## 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\#paramsCOCO Cap\. test CIDErNoCaps val CIDErTextCaps val CIDErCOCO Det\. val2017 mAPFlamingo80B84\.3\-\-\-Florence\-2\-base0\.23B133\.0118\.770\.134\.7Florence\-2\-large0\.77B135\.6120\.872\.837\.5The following table continues the comparison with performance on other vision\-language evaluation tasks\. MethodFlickr30k test R@1Refcoco val AccuracyRefcoco test\-A AccuracyRefcoco test\-B AccuracyRefcoco\+ val AccuracyRefcoco\+ test\-A AccuracyRefcoco\+ test\-B AccuracyRefcocog val AccuracyRefcocog test AccuracyRefcoco RES val mIoUKosmos\-278\.752\.357\.447\.345\.550\.742\.260\.661\.7\-Florence\-2\-base83\.653\.958\.449\.751\.556\.447\.966\.365\.134\.6Florence\-2\-large84\.456\.361\.651\.453\.657\.949\.968\.067\.035\.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\# ParamsCOCO Caption Karpathy test CIDErNoCaps val CIDErTextCaps val CIDErVQAv2 test\-dev AccTextVQA test\-dev AccVizWiz VQA test\-dev Acc**Specialist Models**CoCa2\.1B143\.6122\.4\-82\.3\-\-BLIP\-27\.8B144\.5121\.6\-82\.2\-\-GIT25\.1B145\.0126\.9148\.681\.767\.371\.0Flamingo80B138\.1\-\-82\.054\.165\.7PaLI17B149\.1127\.0160\.0▲84\.358\.8 / 73\.1▲71\.6 / 74\.4▲PaLI\-X55B149\.2126\.3147\.0 / 163\.7▲86\.071\.4 / 80\.8▲70\.9 / 74\.6▲**Generalist Models**Unified\-IO2\.9B\-100\.0\-77\.9\-57\.4Florence\-2\-base\-ft0\.23B140\.0116\.7143\.979\.763\.663\.6Florence\-2\-large\-ft0\.77B143\.3124\.9151\.181\.773\.572\.6Method\# ParamsCOCO Det\. val2017 mAPFlickr30k test R@1RefCOCO val AccuracyRefCOCO test\-A AccuracyRefCOCO test\-B AccuracyRefCOCO\+ val AccuracyRefCOCO\+ test\-A AccuracyRefCOCO\+ test\-B AccuracyRefCOCOg val AccuracyRefCOCOg test AccuracyRefCOCO RES val mIoU**Specialist Models**SeqTR\-\-\-83\.786\.581\.271\.576\.364\.974\.974\.2\-PolyFormer\-\-\-90\.492\.987\.285\.089\.878\.085\.885\.976\.9UNINEXT0\.74B60\.6\-92\.694\.391\.585\.289\.679\.888\.789\.4\-Ferret13B\-\-89\.592\.484\.482\.888\.175\.285\.886\.3\-**Generalist Models**UniTAB\-\-\-88\.691\.183\.881\.085\.471\.684\.684\.7\-Florence\-2\-base\-ft0\.23B41\.484\.092\.694\.891\.586\.891\.782\.289\.882\.278\.0Florence\-2\-large\-ft0\.77B43\.485\.293\.495\.392\.088\.392\.983\.691\.291\.780\.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} } ``` Model createdover 1 year ago