Meta: DINOv2

OpenAI CLIP (Contrastive Language-Image Pretraining) is a vision-language model released in January 2021 by OpenAI. It jointly trains an image encoder and a text encoder to produce matching embeddings for image-caption pairs, using a contrastive objective over WebImageText (WIT), a dataset of 400 million image-text pairs collected from the public web. By learning to associate images with free-form text rather than a fixed set of class labels, CLIP produces a shared embedding space that enables zero-shot classification with arbitrary vocabularies at inference time.CLIP supports zero-shot image classification by embedding candidate class labels as text and selecting the label whose embedding is closest to a given image's embedding. It is also widely used for image-text retrieval, as a frozen backbone in downstream vision-language models, and as a building block for content moderation, similarity search, and generative model guidance — notably as the text conditioning mechanism in early versions of Stable Diffusion. OpenAI released several CLIP variants built on different vision encoders, including ResNet and Vision Transformer backbones at multiple sizes and input resolutions, with ViT-L/14 at 336 pixels being the largest and most widely adopted. CLIP is distributed under the MIT license. The model has been widely influential as the basis for subsequent vision-language work — including SigLIP, OpenCLIP, and MetaCLIP — and remains a common reference baseline despite being released in 2021 and surpassed on many benchmarks by later models.

SigLIP is a vision-language model released in March 2023 by researchers at Google DeepMind. It adapts the CLIP image-text pretraining approach by replacing CLIP's softmax-based contrastive loss with a pairwise sigmoid loss, which operates independently on each image-text pair rather than requiring a global view of all pairs in a batch. This change decouples the loss from batch size, enabling more memory-efficient training and improved performance at smaller batch sizes, a regime where softmax contrastive learning typically struggles. Despite this simplification, SigLIP matches or exceeds CLIP-style models on zero-shot image classification and image-text retrieval benchmarks when trained on comparable data.SigLIP is distributed as an image encoder plus aligned text encoder, supporting zero-shot classification with arbitrary class vocabularies, image-text retrieval, and use as a frozen backbone in downstream vision-language models. Pretrained models are available at multiple Vision Transformer sizes and input resolutions, including 224, 256, 384, and 512 pixel inputs. SigLIP is released under the Apache 2.0 license by Google and is used as the vision encoder in Google's PaliGemma and PaliGemma 2. A successor, SigLIP 2, was released in February 2025 with multilingual support across 109 languages, improvements to localization and dense prediction, and two resolution handling variants (FixRes for backward-compatible fixed resolutions and NaFlex for native aspect ratio with variable sequence length).

Vision Transformer is an image classification model developed by Google Research, first published in October 2020. It applies the transformer architecture directly to sequences of image patches without convolutional layers. Each image is divided into fixed-size patches, linearly projected into embeddings, and processed by a standard transformer encoder with multi-head self-attention. A classification token prepended to the patch sequence aggregates global image information for the final prediction.When pre-trained on large datasets such as JFT-300M and fine-tuned on ImageNet, ViT achieves competitive accuracy with state-of-the-art CNNs of the period. It performs best when pre-training data is abundant, as the lack of convolutional inductive biases makes it less data-efficient than CNN-based classifiers on smaller datasets. ViT established the foundation for transformer-based vision architectures and has influenced a broad range of subsequent models.

ResNet-50 is a deep convolutional neural network architecture introduced in the 2015 paper "Deep Residual Learning for Image Recognition" by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun at Microsoft Research. It is part of the ResNet (Residual Network) family, which introduced residual connections — shortcut paths that allow gradients to bypass layers during training — solving the degradation problem that had previously limited the practical training of very deep networks. ResNet-50 specifically refers to a 50-layer variant with approximately 25.6 million parameters, structured as a sequence of bottleneck residual blocks consisting of 1×1, 3×3, and 1×1 convolutions.ResNet-50 was trained on the ImageNet classification benchmark and achieved leading top-1 accuracy at release. Beyond classification, it became a widely used backbone feature extractor for downstream tasks including object detection (as the base network in Faster R-CNN, Mask R-CNN, and RetinaNet) and semantic and instance segmentation. Most current implementations in PyTorch torchvision, TensorFlow, and NVIDIA NGC use the ResNet-50 v1.5 variant, which relocates the stride-2 downsampling from the first 1×1 convolution to the 3×3 convolution within each bottleneck block, yielding approximately 0.5% higher top-1 accuracy than the original v1 formulation at a small throughput cost. ResNet-50 remains a common reference architecture in computer vision benchmarks and a standard backbone choice in detection and segmentation frameworks. The original Microsoft Research code is released under the MIT license.

MobileNetV2 is a lightweight image classification model developed by Google Research, released in January 2018 under the Apache 2.0 license. It introduces two key architectural innovations: inverted residuals, which expand the channel dimension within each bottleneck block before applying depthwise convolution, and linear bottlenecks, which remove the non-linearity before the projection step to preserve information in low-dimensional spaces.MobileNetV2 achieves competitive top-1 accuracy on ImageNet relative to its computational cost, making it practical for deployment on mobile devices and resource-constrained hardware. It is commonly used as a backbone for classification tasks and as a feature extractor in downstream detection and segmentation models through transfer learning. The architecture scales across a range of width and resolution multipliers, allowing developers to trade accuracy for latency based on deployment requirements.