Supported Models¶

MLV-Bench supports Vision Transformer (ViT) models from multiple sources. This document lists all available models and links to further information. To use a model, call build_model with the <provider>/<name> identifiers listed below.

AIMv2¶

Fini et al. (2024)

Apple ViTs pretrained with a causal autoregressive objective over image patches conditioned on multimodal (image + text) inputs from DFN-2B, with optional teacher distillation (_dist).

AugReg¶

Steiner et al. (2021)

ViTs trained on ImageNet-21K/-1K with carefully tuned augmentation and regularization, matching models trained on an order of magnitude more data.

BEiT¶

Bao et al. (2022)

ViTs pretrained by masked image modeling against discrete dVAE visual tokens on ImageNet-22K, optionally fine-tuned on ImageNet-1K.

BEiT v2¶

Peng et al. (2022)

A BEiT variant that reconstructs CLIP visual features instead of dVAE tokens during masked image modeling.

BEiT-3¶

Wang et al. (2023)

A multimodal foundation model applying masked data modeling across text, images, and image–text pairs with a shared Multiway Transformer, fine-tuned on ImageNet for classification.

CLIP (LAION-2B)¶

Cherti et al. (2023)

OpenCLIP reproduction of CLIP trained on the open LAION-2B dataset, enabling larger model sizes and public reproducibility.

CLIP (OpenAI WIT-400M)¶

Radford et al. (2021)

The original CLIP image encoders, aligned with text by a contrastive objective on OpenAI's private WIT-400M dataset, with optional ImageNet fine-tuning.

DeiT¶

Touvron et al. (2021)

Data-efficient ViTs trained competitively on ImageNet-1K alone via knowledge distillation, with distilled variants adding a dedicated distillation token.

DeiT III¶

Touvron et al. (2022)

A revised DeiT recipe with stronger augmentation and supervised-training best practices, trained on ImageNet-1K or pretrained on ImageNet-21K and fine-tuned on ImageNet-1K.

DINO¶

Caron et al. (2021)

Self-supervised ViTs trained by self-distillation between a student and a momentum teacher, exhibiting emergent object segmentation.

DINOv2¶

Oquab et al. (2024), Darcet et al. (2024)

DINO scaled on the curated LVD-142M dataset for strong transferable dense features, with reg4 variants adding register tokens that remove attention-map artifacts.

DINOv3¶

Siméoni et al. (2025)

A further-scaled DINO with an updated Transformer architecture and larger curated datasets (LVD-1689M and SAT-493M).

EVA¶

Fang et al. (2023)

ViTs pretrained by masked image modeling that reconstructs OpenAI CLIP-L/14 features, included here as ImageNet-fine-tuned checkpoints.

EVA-02¶

Fang et al. (2023)

ViTs pretrained by masked image modeling with an EVA-CLIP teacher on ImageNet-22K or merged vision datasets, available as pretrained-only or ImageNet-fine-tuned checkpoints.

EVA-CLIP¶

Sun et al. (2023)

CLIP scaled to multi-billion-parameter vision encoders via progressive scaling, trained on LAION-400M (EVA01) or a merged 2B dataset (EVA02) and optionally fine-tuned on ImageNet-1K.

FlexiViT¶

Beyer et al. (2023)

A ViT trained with randomly sampled patch sizes so a single checkpoint handles variable patch grids and the full compute–accuracy trade-off, trained on ImageNet-1K or ImageNet-21K.

Franca¶

Venkataramanan et al. (2026)

An open-source DINOv2-based self-supervised framework with a Matryoshka multi-head clustering projector; dinov2 names are baselines and rasa applies a position–semantic disentanglement post-training step.

MAE¶

He et al. (2022)

ViT encoder–decoders pretrained to reconstruct raw pixels from roughly 75% masked image patches, yielding strong transferable representations.

MIIL¶

Ridnik et al. (2021)

ViTs from an efficient large-scale ImageNet-21K pretraining pipeline with a semantic-tree class mapping that eases ImageNet-1K fine-tuning.

Original ViT¶

Dosovitskiy et al. (2021)

The original Vision Transformer checkpoints pretrained on ImageNet-21K and optionally fine-tuned on ImageNet-1K.

Perception Encoder¶

Bolya et al. (2025)

Meta AI's PE-Core vision backbone trained on a large vision–language mix to produce spatially rich representations for dense perception tasks.

RoPE ViT¶

Heo et al. (2024)

ViTs that replace absolute position embeddings with Rotary Position Embeddings (RoPE), improving ImageNet classification and dense prediction.

SAM¶

Chen et al. (2022)

ImageNet-1K ViTs trained with Sharpness-Aware Minimization (Foret et al. 2021) for improved accuracy and generalization.

SigLIP¶

Zhai et al. (2023)

CLIP-style image–text encoders trained with a pairwise sigmoid loss that removes the global normalization step for more efficient training.

SigLIP 2¶

Tschannen et al. (2025)

SigLIP extended with multilingual captions, decoder-based pretraining tasks, and self-supervised objectives for improved spatial understanding and dense features.

Web-SSL¶

Fan et al. (2025)

DINOv2- and MAE-style self-supervised models scaled to web-scale data (up to 8B images) and up to 7B parameters, showing language-free SSL can match language-supervised models.