README.md

April 6, 2025 · View on GitHub

Arxiv Hugging Face PWC

Performance

A. MLLMs Evaluation Results

To evaluate MLCD’s performance within multimodal large language models (MLLMs), we replaced the CLIP model in LLaVA-NeXT with the MLCD model. We paired this with the Qwen2.5-7B language model. For reproducibility, we utilized the LLaVA-Pretrain dataset for pre-training and the LLaVA-NeXT-Data for structured fine-tuning. The evaluation results confirm that the MLCD model performs exceptionally well across multiple benchmarks, underscoring its effectiveness in MLLMs.

Vision TowerRoPE2DChartQADocVQAInfoVQAOCRBenchMMMU
CLIP (ViT-L-14-336px)×66.5275.2138.88525.0044.20
SigLIP (ViT-SO400M-384px)×69.2876.7141.38554.0046.78
DFN5B (ViT-H-14-378px)×64.3670.8738.59473.0048.00
HF:MLCD (ViT-L-14-336px)×67.8476.4643.48531.0044.30
HF:MLCD (ViT-bigG-14-336px)71.0779.6344.38572.0046.78
HF:MLCD (ViT-bigG-14-448px)73.8083.3446.59582.0046.00
Vision TowerMLCD (ViT_L_14_336px)CLIP (ViT_L_14_336px)
LLMQwen2.5-7BQwen2.5-7B
AI2D76.9873.15
GQA64.1763.31
ScienceQA-Img78.0976.35
InfoVQA-Val43.4838.88
MMBenchCN-Dev74.8372.51
MMBenchEN-Dev76.3774.57
SeedBench68.2066.80
SeedBench-Img73.7572.72
MMStar50.9848.98
MMMU44.3044.20
POPE88.6988.83
ChartQA67.8466.52
DocVQA-Val76.4675.21
TextVQA-Val61.6962.47
OCRBench531525
MME(cognition)432384
MME(perception)15981512

B. Linear Probe Evaluation Results

This table presents the results of linear probe evaluations comparing CLIP and MLCD models on the ViT_L_14_336px architecture across various datasets. The linear probe test freezes the pre-trained model's weights and trains a linear classifier on top to assess how well the model's representations generalize to different tasks.

The results of the ImageNet linear probe are as follows:

Model NameImageNet Linear ProbeHugging Face
MLCD-ViT-B-32-224px79.1HF:MLCD-ViT-B-32-224px
MLCD-ViT-L-14-336px86.3HF:MLCD-ViT-L-14-336px
MLCD-ViT-bigG-14-224px87.1HF:MLCD-ViT-bigG-14-224px
DatasetMLCD (ViT_L_14_336px)CLIP (ViT_L_14_336px)
Food10196.2195.90
CIFAR-1099.3697.90
CIFAR-10093.6987.40
Birdsnap88.1879.90
SUN39787.9682.20
Stanford Cars95.1691.50
FGVC Aircraft86.3871.60
Describable Textures Dataset86.7083.00
Oxford-IIIT Pets96.2795.10
Caltech-10197.9296.00
Flowers10299.5899.20
ImageNet86.1085.40

convert pytorch2huggingface


python convert_vit_bigG_14_rope2d_to_hf.py \
--pytorch_dump_folder_path mlcd-vit-bigG-patch14-336 \
--checkpoint_path MLCD_ViT_bigG_14_336px_pytorch.pt \
--image_size 336