Stephanie Fu* $^{1}$, Netanel Tamir* $^{2}$, Shobhita Sundaram* $^{1}$, Lucy Chai $^1$, Richard Zhang $^3$, Tali Dekel $^2$, Phillip Isola $^1$.
(*equal contribution, order decided by random seed)
$^1$ MIT, $^2$ Weizmann Institute of Science, $^3$ Adobe Research.

Summary

Current metrics for perceptual image similarity operate at the level of pixels and patches. These metrics compare images in terms of their low-level colors and textures, but fail to capture mid-level differences in layout, pose, semantic content, etc. Models that use image-level embeddings such as DINO and CLIP capture high-level and semantic judgements, but may not be aligned with human perception of more finegrained attributes.

DreamSim is a new metric for perceptual image similarity that bridges the gap between "low-level" metrics (e.g. LPIPS, PSNR, SSIM) and "high-level" measures (e.g. CLIP). Our model was trained by concatenating CLIP, OpenCLIP, and DINO embeddings, and then finetuning on human perceptual judgements. We gathered these judgements on a dataset of ~20k image triplets, generated by diffusion models. Our model achieves better alignment with human similarity judgements than existing metrics, and can be used for downstream applications such as image retrieval.

10/14/24: We released 4 new variants of DreamSim! These new checkpoints are:

• DINOv2 B/14 and SynCLR B/16 as backbones
• Models trained with the original contrastive loss on both CLS and dense features.

These models (and the originals) are further evaluated in our new NeurIPS 2024 paper, When Does Perceptual Alignment Benefit Vision Representations?

We find that our perceptually-aligned models outperform the baseline models on a variety of standard computer vision tasks, including semantic segmentation, depth estimation, object counting, instance retrieval, and retrieval-augmented generation. These results point towards perceptual alignment being a useful task for learning general-purpose vision representations. See the paper and our blog post for more details.

Here's how they perform on NIGHTS:

9/14/24: We released new versions of the ensemble and single-branch DreamSim models compatible with peft>=0.2.0.

We also released the entire 100k (unfiltered) NIGHTS dataset and the JND (Just-Noticeable Difference) votes.

• Requirements
• Setup
• Usage
  • Quickstart
  • Single-branch models
  • Feature extraction
  • Image retrieval
  • Perceptual loss function
• NIGHTS Dataset
• Experiments
• Citation

• Linux
• Python 3

Option 1: Install using pip:

pip install dreamsim

The package is used for importing and using the DreamSim model.

Option 2: Clone our repo and install dependencies. This is necessary for running our training/evaluation scripts.

python3 -m venv ds
source ds/bin/activate
pip install -r requirements.txt
export PYTHONPATH="$PYTHONPATH:$(realpath ./dreamsim)"

To install with conda:

conda create -n ds
conda activate ds
conda install pip # verify with the `which pip` command
pip install -r requirements.txt
export PYTHONPATH="$PYTHONPATH:$(realpath ./dreamsim)"

For walk-through examples of the below use-cases, check out our Colab demo.

The basic use case is to measure the perceptual distance between two images. A higher score means more different, lower means more similar.

The following code snippet is all you need. The first time that you run dreamsim it will automatically download the model weights. The default model settings are specified in ./dreamsim/config.py.

from dreamsim import dreamsim
from PIL import Image

device = "cuda"
model, preprocess = dreamsim(pretrained=True, device=device)

img1 = preprocess(Image.open("img1_path")).to(device)
img2 = preprocess(Image.open("img2_path")).to(device)
distance = model(img1, img2) # The model takes an RGB image from [0, 1], size batch_sizex3x224x224

To run on example images, run demo.py. The script should produce distances (0.4453, 0.2756).

By default, DreamSim uses an ensemble of CLIP, DINO, and OpenCLIP (all ViT-B/16). If you need a lighter-weight model you can use single-branch versions of DreamSim where only a single backbone is finetuned. The single-branch models provide a ~3x speedup over the ensemble.

The available options are OpenCLIP-ViTB/32, DINO-ViTB/16, CLIP-ViTB/32, in order of performance. To load a single-branch model, use the dreamsim_type argument. For example:

dreamsim_dino_model, preprocess = dreamsim(pretrained=True, dreamsim_type="dino_vitb16")

To extract a single image embedding using dreamsim, use the embed method as shown in the following snippet:

img1 = preprocess(Image.open("img1_path")).to("cuda")
embedding = model.embed(img1)

The perceptual distance between two images is the cosine distance between their embeddings. If the embeddings are normalized (true by default) L2 distance can also be used.

Our model can be used for image retrieval, and plugged into existing such pipelines. The code below ranks a dataset of images based on their similarity to a given query image.

To speed things up, instead of directly calling model(query, image) for each pair, we use the model.embed(image) method to pre-compute single-image embeddings, and then take the cosine distance between embedding pairs.

import pandas as pd
from tqdm import tqdm
import torch.nn.functional as F

# let query be a sample image.
# let images be a list of images we are searching.

# Compute the query image embedding
query_embed = model.embed(preprocess(query).to("cuda"))
dists = {}

# Compute the (cosine) distance between the query and each search image
for i, im in tqdm(enumerate(images), total=len(images)):
   img_embed = model.embed(preprocess(im).to("cuda"))
   dists[i] = (1 - F.cosine_similarity(query_embed, img_embed, dim=-1)).item()

# Return results sorted by distance
df = pd.DataFrame({"ids": list(dists.keys()), "dists": list(dists.values())})
return df.sort_values(by="dists")

Our model can be used as a loss function for iterative optimization (similarly to the LPIPS metric). These are the key lines; for the full example, refer to the Colab.

for i in range(n_iters):
    dist = model(predicted_image, reference_image)
    dist.backward()
    optimizer.step()

DreamSim is trained by fine-tuning on the NIGHTS dataset. For details on the dataset structure and creation, refer to the dataset page.

Run ./dataset/download_dataset.sh to download and unzip the NIGHTS dataset into ./dataset/nights. The unzipped dataset size is 58 GB.

Having trouble with the large file sizes? Run ./dataset/download_chunked_dataset.sh to download the NIGHTS dataset split into 200 smaller zip files. The output of this script is identical to download_dataset.sh.

We only use the 20k unanimous triplets for training and evaluation, but release all 100k triplets (many with few and/or split votes) for research purposes. Run ./dataset/download_unfiltered_dataset.sh to download and unzip this unfiltered version of the NIGHTS dataset into ./dataset/nights_unfiltered. The unzipped dataset size is 289 GB.

Download the just-noticeable difference (JND) votes by running ./dataset/download_jnd_dataset.sh. The CSV will be downloaded to ./dataset/jnd_votes.csv. Check out the Colab for an example of loading a JND trial.

Run ./training/download_models.sh to download and unzip necessary ViT checkpoints (for CLIP, OpenCLIP, and MAE) into ./models.

To finetune a perceptual model on the dataset, run ./training/train.py. For example, to finetune an ensemble of DINO, CLIP, and OpenCLIP using LoRA, run:

python ./training/train.py --config ./configs/train_ensemble_model_lora.yaml

We provide more sample configs in ./configs, including examples of finetuning with LoRA and with an MLP head. See ./training/train.py for a full list and description of training options.

To evaluate a perceptual model on the dataset, run ./training/evaluate.py. For example, to evaluate dreamsim on the dataset, run:

python ./training/evaluate.py --config ./configs/eval_baseline.yaml

For an example of evaluating using a trained checkpoint, refer to ./configs/eval_checkpoint.yaml. See ./training/evaluate.py for a full list and description of evaluation options.

If you find our work or any of our materials useful, please cite our paper:

@misc{fu2023dreamsim,
      title={DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data}, 
      author={Stephanie Fu and Netanel Tamir and Shobhita Sundaram and Lucy Chai and Richard Zhang and Tali Dekel and Phillip Isola},
      year={2023},
      eprint={2306.09344},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Our code borrows from the "Deep ViT Features as Dense Visual Descriptors" repository for ViT feature extraction, and takes inspiration from the UniverSeg respository for code structure.