[CVPR 2025 Highlight] Assessing and Learning Alignment of Unimodal Vision and Language Models

This codebase enables you to train your own CLIP-like model on a single GPU by aligning pretrained vision models, such as DINOv2, with language models like NV-Embed-2. Our approach demonstrates that training only a lightweight alignment layer while keeping the backbones frozen is sufficient to bridge the vision and language representation spaces. Using just 23M web-collected and synthetic image-text pairs, we developed a foundational model called SAIL-L, which surpasses CLIP-L (LAION400M) in vairous retrieval tasks and ImageNet while also serving as a strong vision encoder for building Multimodal LLMs. We hope this codebase serve as a useful testbed for resource-limited community to explore multimodal representation learning in terms of new losses, new data combination as well as new modality-merge strategy.

📯 Release

• [2025/04/04] 🚀 SAIL is selected as CVPR 2025 Highlight !

• [2025/02/27] 🚀 SAIL Paper Accepted to CVPR 2025 !
  Exciting news! This work on assessing and learning cross-modal alignment has been officially accepted by CVPR 2025!

• [2024/11/20] 🔥 SAIL Codebase Open-Sourced!
  The repository now includes the complete pipeline for training data preparation and preprocessing, as well as the full training and evaluation codebase.

🔖 Content

• Quick Usage
• SAIL Training & Evaluation
  • Data Preparation
  • Training
  • Evaluation
• SAIL Enhances SSL Models for MLLMs
  • LLaVA Train with SAIL
  • Evaluation

⚙️ Quick Usage

1. Clone the Repository

   git clone https://github.com/lezhang7/SAIL.git
   pip install -r requirements.txt
   

2. Download the Alignment Layer Checkpoint You can download the pretrained alignment layer checkpoints from the links below:

   Data	Model	Alignment Layer	IN-1K	I2T R@1 (MSCOCO)	T2I R@1 (MSCOCO)	I2T R@1 (Flickr30k)	T2I R@1 (Flickr30k)	Text (Winoground)	Image (Winoground)	Group (Winoground)	Avg. (MMVP)
   23M	SAIL-L (GTE)	download	65.4	54.1	42.7	80.8	68.9	34.0	13.25	8.75	22.2
   23M	SAIL-L (NV2)	download	73.4	62.4	48.6	87.6	75.7	40.25	18.75	15.0	28.9
   LAION-400M	CLIP-L		72.7	59.7	43.0	87.6	70.2	30.5	11.5	8.75	20.0

3. Run the Model

   from model import create_model
   from PIL import Image
   import torch
   
   # Path to the downloaded checkpoint
   checkpoint_path = "checkpoint/sail_dinov2l_nv2.pt"
   
   # Create the model, change the text_model to `Alibaba-NLP/gte-large-en-v1.5` if use sail_dinov2_gte
   model = create_model(
       text_model_name="nvidia/NV-Embed-v2",
       vision_model_name="facebook/dinov2-large",
       head_weights_path=checkpoint_path,
       target_dimension=1024,
   )
   model.eval()  # Set model to evaluation mode
   
   # Prepare images and texts
   image_processor = model.image_processor
   texts = ["a dog", "a cat"]
   dog_image = Image.open("asset/dog.jpg").convert("RGB")
   cat_image = Image.open("asset/cat.jpg").convert("RGB")
   images = image_processor(images=[dog_image, cat_image], return_tensors="pt")
   
   # Generate features and probabilities
   with torch.no_grad():
       image_features = model.encode_image(images, normalize=True)
       text_features = model.encode_text(texts, text_list=texts, normalize=True)
   text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
   
   # Print the label probabilities
   print("Label probs:", text_probs)
   

SAIL Training & Evaluation

The codebase builds upon OpenCLIP (for training SAIL) and LLaVA (for testing SAIL's vision encoder in MLLMs). Please ensure the necessary dependency packages for these frameworks are installed.

Data Preparation

SAIL leverages high-quality, MLLM-enhanced captions for training, using datasets introduced in DreamLIP. To streamline this process, we provide a script for automated dataset preparation. Note that this process is time-intensive, as it involves handling 23M data samples.

cd data_preparation
bash download_mllm_enhanced_data.sh

The scripts run following steps:

1. Downloading Dataset:
   The script downloads .csv files containing image URLs and their corresponding captions for datasets such as CC3M, CC12M, and YFCC15M.

2. Filtering Invalid Data:
   Since some image URLs may have expired or the images may be corrupted, the downloaded images need to be filtered. Invalid samples must also be removed from the .csv files.

3. Manual Processing (Optional):
   As each step can take a significant amount of time, we recommend running the commands manually based on your specific requirements.

Updating Data Paths:

Once the preprocessing is complete, update the dataset paths including annotation and imagedir field in data/data_config.py :

DATADIR = {
'dreamclipcc3m': { 'annotation':f'{SCRATCH}/datasets/DownloadCC3M/cc3m_3long_1raw_captions_filterd.csv',
    'imagedir':f'{SCRATCH}/datasets/DownloadCC3M'
    }
}

---

Training

The training framework of SAIL consists of two main steps: Pre-encoding and Alignment Tuning. This efficient framework allows us to align the representation space of large pretrained unimodal models (e.g., DINOv2 and NV2 models) on a single A100 GPU with a large batch size of 32,768, requiring only approximately ~5 hours of training during the alignment tuning stage.

stage 1. Pre-encoding

We provide scripts to pre-encode image-text pairs into embeddings. The script will automatically download the required model checkpoints from either Hugging Face Transformers or TorchHub. By default, the vision model weights will be stored in the model/backbone_checkpoints directory.

Note: Ensure your transformers library version is >4.38.0 to support SDPA or FlashAttention implementations for some of the models.

Instructions:

• Review scripts/encode.sh for detailed configurations. Choose the text_model or vision_model in scripts/encode.sh.
• Set the encoded domain to either image or text.
  • If encoding text, select the caption source: raw_caption or high-quality captions shortIB_captions, shortSV_captions, shortLLA_captions longIB_captions, longSV_captions, longLLA_captions
• Execute the following command to encode the data. The embeddings will be saved in the data/tensor_data directory.

bash scripts/encode.sh

stage 2. Alignment Tuning

To train SAIL, specify the text and image embedding data by updating the text_embedding_list and image_embedding_list in scripts/sail_train.sh.

• To enable multiple positive captions for contrastive loss, also update the extra_text_embedding_list.
• Important: Ensure embeddings of the same modality are derived from the same model.

Training:

Run the following command to train the alignment layer:

bash scripts/sail_train.sh

Alignment Probing:

To probe the alignment, execute:

bash scripts/alignment_probing.sh

We only save the alignment layer checkpoint at ./logs/${output_name}.

---

Evaluation

We provide comprehensive evaluation scripts for classification and general retrieval tasks in our repository: clip_eval.

We also provide evaluation scripts for MMVP, imagenet, winoground in scripts/sail_eval.sh.

Instructions for MMVP, imagenet, winoground:

• Set the vision_model, text_model, and checkpoint_path in scripts/sail_eval.sh.
  • Ensure that the vision and text models match the embedding data used to train the alignment layers.
• Prepare datasets
  • Download MMVP_VLM and save it to evaluation/MMVP_VLM
  • ImageNet and Winoground will be automatically downloadad and processed
• Specify the task from imagenetv1 winoground MMVP in sail_eval.sh, then evaluate by running:

bash scripts/sail_eval.sh

The evaluation results will be saved to evaluation/eval_result/{task}

Open-vocabulary semantic segmentation Instructions please refer to here

SAIL Enhances SSL Vision Models for MLLMs

SAIL significantly enhances SSL models, such as DINOv2, as vision encoders for MLLMs. Specifically, we replace the vision encoder in LLaVA-1.5 with the SAIL vision encoder, which consists of a DINOv2 backbone combined with an alignment layer. This additional alignment layer dramatically improves DINOv2's performance on MLLM tasks, even surpassing language-supervised CLIP vision encoders in certain tasks! We provide trained checkpoint at le723z/sail-llava-v1.5-7b.

Table: LLaVA-1.5 with Various Vision Models

VTune indicates whether the vision encoder is fine-tuned during the instruction tuning stage.

---

Training LLaVA with SAIL Vision Encoder

We follow the LLaVA-1.5 training process, including pretraining and fine-tuning. To get started, please prepare the data following the instructions in the original codebase:

1. Pretraining data
2. Visual instruction tuning data

Then install the dependency packages following here for training llava-1.5. Recommend using cudatoolkit/12.1.1 for reproducibility.

Pretraining SAIL with LLaVA-1.5

Specify data_path and image_folder for the dataset configuration. Set vlhead_weights_path and vision_tower for SAIL-LLaVA as follows:

deepspeed llava_train/train_mem.py \
    --deepspeed llava_train/zero2.json \
    --model_name_or_path lmsys/vicuna-7b-v1.5 \
    --version plain \
    --data_path blip_laion_cc_sbu_558k.json \
    --target_dimension 1024 \
    --linear_type star \
    --image_folder llava-v1.5-7b/pretrain_data \
    --vision_tower facebook/dinov2-large \
    --mm_projector_type mlp2x_gelu \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --num_train_epochs 1 \
    --per_device_train_batch_size 64 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 4000 \
    --save_total_limit 1 \
    --learning_rate 0.001 \
    --weight_decay 0.0 \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --fp16 False \
    --bf16 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4  \
    --lazy_preprocess True \
    --report_to wandb  \
    --vlhead_weights_path logs/sail_l_nv2_merged23m/checkpoints/64.pt \
    --tune_alignment_layer False \
    --output_dir ./llava_checkpoints/sail_llava_pretrain

Fine-Tuning SAIL with LLaVA-1.5

For fine-tuning, update the data_path and other configurations. Use the following command:

deepspeed llava_train/train_mem.py \
    --deepspeed llava_train/zero3.json \
    --model_name_or_path lmsys/vicuna-7b-v1.5 \
    --version v1 \
    --data_path $dataset/llava-v1.5-7b/instruct_tuning_data/llava_v1_5_mix665k.json \
    --image_folder $dataset/llava-v1.5-7b/instruct_tuning_data/data \
    --target_dimension 1024 \
    --linear_type star \
    --vision_tower facebook/dinov2-large \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 400 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb \
    --vlhead_weights_path llava_checkpoints/sail_llava_pretrain/vlhead.bin \
    --pretrain_mm_mlp_adapter llava_checkpoints/sail_llava_pretrain/mm_projector.bin \
    --tune_alignment_layer True \
    --unlock_vision_tower True \
    --output_dir ./llava_checkpoints/sail_llava_finetune

---

Evaluation

We provide evaluation scripts in scripts/llava_eval_scripts. Download the evaluation dataset following here and update the dataset path for each task bash script, then run the evaluation to test the model's performance.

Acknowledgement

This project is based on open_clip, DreamLIP and LLaVA , we appreicate their great work!

Citation

If you find our work useful, please consider citing:

@inproceedings{zhang2025assessing,
  title={Assessing and Learning Alignment of Unimodal Vision and Language Models},
  author={Zhang, Le and Yang, Qian and Agrawal, Aishwarya},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={14604--14614},
  year={2025}
}