image-reward

ImageReward

🤗 HF Repo • 🐦 Twitter • 📃 Paper • 🖼 Dataset • 🌐 中文博客

ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation

ImageReward is the first general-purpose text-to-image human preference RM, which is trained on in total 137k pairs of expert comparisons, outperforming existing text-image scoring methods, such as CLIP (by 38.6%), Aesthetic (by 39.6%), and BLIP (by 31.6%), in terms of understanding human preference in text-to-image synthesis.

Additionally, we introduce Reward Feedback Learning (ReFL) for direct optimizing a text-to-image diffusion model using ImageReward. ReFL-tuned Stable Diffusion wins against untuned version by 58.4% in human evaluation.

Both ImageReward and ReFL are all packed up to Python image-reward package now!

Try image-reward package in only 3 lines of code for ImageReward scoring!

# pip install image-reward
import ImageReward as RM
model = RM.load("ImageReward-v1.0")

rewards = model.score("<prompt>", ["<img1_obj_or_path>", "<img2_obj_or_path>", ...])

Try image-reward package in only 4 lines of code for ReFL fine-tuning!

# pip install image-reward
# pip install diffusers==0.16.0 accelerate==0.16.0 datasets==2.11.0
from ImageReward import ReFL
args = ReFL.parse_args()
trainer = ReFL.Trainer("CompVis/stable-diffusion-v1-4", "data/refl_data.json", args=args)
trainer.train(args=args)

If you find ImageReward's open-source effort useful, please 🌟 us to encourage our following developement!

• ImageReward
  • Quick Start
    • Install Dependency
    • Example Use
  • ReFL
    • Install Dependency
    • Example Use
  • Demos of ImageReward and ReFL
  • Training code for ImageReward
  • Integration into Stable Diffusion Web UI
    • Features
      • Score generated images and append to image information
        • Usage:
        • Demo video:
      • Automatically filter out images with low scores
        • Usage:
        • Demo video:
      • View the scores of images that have been scored
        • Usage:
        • Example:
      • Other Features
        • Memory Management
  • Reproduce Experiments in Table 1
  • Reproduce Experiments in Table 3
  • Citation

Quick Start

Install Dependency

We have integrated the whole repository to a single python package image-reward. Following the commands below to prepare the environment:

# Clone the ImageReward repository (containing data for testing)
git clone https://github.com/THUDM/ImageReward.git
cd ImageReward

# Install the integrated package `image-reward`
pip install image-reward

Example Use

We provide example images in the assets/images directory of this repo. The example prompt is:

a painting of an ocean with clouds and birds, day time, low depth field effect

Use the following code to get the human preference scores from ImageReward:

import os
import torch
import ImageReward as RM

if __name__ == "__main__":
    prompt = "a painting of an ocean with clouds and birds, day time, low depth field effect"
    img_prefix = "assets/images"
    generations = [f"{pic_id}.webp" for pic_id in range(1, 5)]
    img_list = [os.path.join(img_prefix, img) for img in generations]
    model = RM.load("ImageReward-v1.0")
    with torch.no_grad():
        ranking, rewards = model.inference_rank(prompt, img_list)
        # Print the result
        print("\nPreference predictions:\n")
        print(f"ranking = {ranking}")
        print(f"rewards = {rewards}")
        for index in range(len(img_list)):
            score = model.score(prompt, img_list[index])
            print(f"{generations[index]:>16s}: {score:.2f}")

The output should be like as follow (the exact numbers may be slightly different depending on the compute device):

Preference predictions:

ranking = [1, 2, 3, 4]
rewards = [[0.5811622738838196], [0.2745276093482971], [-1.4131819009780884], [-2.029569625854492]]
          1.webp: 0.58
          2.webp: 0.27
          3.webp: -1.41
          4.webp: -2.03

ReFL

Install Dependency

pip install diffusers==0.16.0 accelerate==0.16.0 datasets==2.11.0

Example Use

We provide example dataset for ReFL in the data/refl_data.json of this repo. Run ReFL as following:

bash scripts/train_refl.sh

Demos of ImageReward and ReFL

Training code for ImageReward

1. Download data: 🖼 Dataset.

2. Make dataset.

cd train
python src/make_dataset.py

1. Set training config: train/src/config/config.yaml

2. One command to train.

bash scripts/train_one_node.sh

Integration into Stable Diffusion Web UI

We have developed a custom script to integrate ImageReward into SD Web UI for a convenient experience.

The script is located at sdwebui/image_reward.py in this repository.

The usage of the script is described as follows:

1. Install: put the custom script into the stable-diffusion-webui/scripts/ directory
2. Reload: restart the service, or click the "Reload custom script" button at the bottom of the settings tab of SD Web UI. (If the button can't be found, try clicking the "Show all pages" button at the bottom of the left sidebar.)
3. Select: go back to the "txt2img"/"img2img" tab, and select "ImageReward - generate human preference scores" from the "Script" dropdown menu in the lower left corner.
4. Run: the specific usage varies depending on the functional requirements, as described in the "Features" section below.

Features

Score generated images and append to image information

Usage:

1. Do not check the "Filter out images with low scores" checkbox.
2. Click the "Generate" button to generate images.
3. Check the ImageReward at the bottom of the image information below the gallery.

Demo video:

https://github.com/THUDM/ImageReward/assets/98524878/9d8a036d-1583-4978-aac7-4b758edf9b89

Automatically filter out images with low scores

Usage:

1. Check the "Filter out images with low scores" checkbox.
2. Enter the score lower limit in "Lower score limit". (ImageReward roughly follows the standard normal distribution, with a mean of 0 and a variance of 1.)
3. Click the "Generate" button to generate images.
4. Images with scores below the lower limit will be automatically filtered out and will not appear in the gallery.
5. Check the ImageReward at the bottom of the image information below the gallery.

Demo video:

https://github.com/THUDM/ImageReward/assets/98524878/b9f01629-87d6-4c92-9990-fe065711b9c6

View the scores of images that have been scored

Usage:

1. Upload the scored image file in the "PNG Info" tab
2. Check the image information on the right with the score of the image at the bottom.

Example:

Other Features

Memory Management

• ImageReward model will not be loaded until first script run.
• "Reload UI" will not reload the model nor unload it, but reuses the currently loaded model (if it exists).
• A "Unload Model" button is provided to manually unload the currently loaded model.

Reproduce Experiments in Table 1

Note: The experimental results are produced in an environment that satisfies:

• (NVIDIA) Driver Version: 515.86.01
• CUDA Version: 11.7
• torch Version: 1.12.1+cu113 According to our own reproduction experience, reproducing this experiment in other environments may cause the last decimal place to fluctuate, typically within a range of ±0.1.

Run the following script to automatically download data, baseline models, and run experiments:

bash ./scripts/test-benchmark.sh

Then you can check the results in benchmark/results/ or the terminal.

If you want to check the raw data files individually:

• Test prompts and corresponding human rankings for images are located in benchmark/benchmark-prompts.json.
• Generated outputs for each prompt (originally from DiffusionDB) can be downloaded from Hugging Face or Tsinghua Cloud.
  • Each <model_name>.zip contains a directory of the same name, in which there are in total 1000 images generated from 100 prompts of 10 images each.
  • Every <model_name>.zip should be decompressed into benchmark/generations/ as directory <model_name> that contains images.

Reproduce Experiments in Table 3

Run the following script to automatically download data, baseline models, and run experiments:

bash ./scripts/test.sh

If you want to check the raw data files individually:

• Test prompts and corresponding human rankings for images are located in data/test.json.
• Generated outputs for each prompt (originally from DiffusionDB) can be downloaded from Hugging Face or Tsinghua Cloud. It should be decompressed to data/test_images.

Citation

@misc{xu2023imagereward,
      title={ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation},
      author={Jiazheng Xu and Xiao Liu and Yuchen Wu and Yuxuan Tong and Qinkai Li and Ming Ding and Jie Tang and Yuxiao Dong},
      year={2023},
      eprint={2304.05977},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}