# Multimodal Pre-training

YOLOE is a real-time visual omni-model that supports various vision tasks including zero-shot object detection.

YOLOE is a real-time visual omni-model that combines object detection and visual understanding capabilities, suitable for various visual tasks.

YOLOE is a zero-shot object detection model capable of detecting various objects in visual scenes in real-time.

AIM-v2 is an efficient image encoder implemented based on the timm library, suitable for image feature extraction tasks.

AIM-v2 is an efficient image encoder model compatible with the timm framework, suitable for computer vision tasks.

A vision Transformer model based on SigLIP, containing only an image encoder, utilizing the original attention pooling mechanism.

A vision Transformer model based on SigLIP, utilizing global average pooling, suitable for image feature extraction tasks.

Vision Transformer model based on SigLIP, containing only the image encoder part, using original attention pooling mechanism

Vision Transformer model based on SigLIP, containing only the image encoder part, using original attention pooling mechanism

Vision Transformer model based on CLIP architecture, specialized in image feature extraction

AIMv2-Large-Patch14-Native is an adapted image classification model, modified from the original AIMv2 model to be compatible with Hugging Face Transformers' AutoModelForImageClassification class.

A vision-language model trained on the MetaCLIP-400M dataset, supporting zero-shot image classification tasks

Vision Transformer model trained on the LAION-2B dataset, supporting zero-shot image classification tasks

OpenCLIP model based on ResNet50 architecture and trained on the CC12M dataset, supporting zero-shot image classification tasks.

Audio representation learning model based on HuBERT architecture, pre-trained on the complete AudioSet dataset

FoodSeg103 is a dataset containing 7,118 food images, annotated with 104 ingredient categories, with an average of 6 ingredient labels and pixel-level masks per image.

A vision-language model based on the CLIP architecture, supporting zero-shot image classification tasks

EVA02 is a large-scale vision-language model based on the CLIP architecture, supporting zero-shot image classification tasks.

Pix2Struct is an image encoder-text decoder model trained on various image-text pairs for tasks including image captioning and visual question answering.

Chinese CLIP is a simplified implementation of CLIP based on approximately 200 million Chinese image-text pairs, using ViT-L/14@336px as the image encoder and RoBERTa-wwm-base as the text encoder.

The first open-source Chinese-English bilingual Stable Diffusion model, trained on 20 million filtered Chinese image-text pairs

X-CLIP is a minimalist extension of CLIP for general video-language understanding. The model is trained on (video, text) pairs through contrastive learning.

A fine-tuned version of the LayoutLMv3-large model based on the FUNSD dataset, specializing in document intelligence understanding tasks

A document AI model fine-tuned on the FUNSD dataset based on the LayoutLMv3-base model, designed for form understanding tasks

A document visual question answering model fine-tuned based on microsoft/layoutlmv2-large-uncased, suitable for Vietnamese information extraction tasks

BROS is a pre-trained language model focusing on text and layout, designed to better extract key information from documents.

A Chinese classical poetry generation model based on the GPT2 architecture, pre-trained by UER-py, capable of generating Chinese classical poetry.

This model is a question-answering model fine-tuned based on Microsoft's MarkupLM architecture, specifically designed for handling Q&A tasks combining web markup languages (HTML/XML) with text.