This repository contains the implementation of the following paper:

ReVersion: Diffusion-Based Relation Inversion from Images
Ziqi Huang^∗, Tianxing Wu^∗, Yuming Jiang, Kelvin C.K. Chan, Ziwei Liu

From MMLab@NTU affiliated with S-Lab, Nanyang Technological University

[Paper] | [Project Page] | [Video] | [Dataset]

We propose a new task, Relation Inversion: Given a few exemplar images, where a relation co-exists in every image, we aim to find a relation prompt <R> to capture this interaction, and apply the relation to new entities to synthesize new scenes. The above images are generated by our ReVersion framework.

• [08/2023] We released the training code for Relation Inversion.
• [04/2023] We released the ReVersion Benchmark.
• [04/2023] Integrated into Hugging Face 🤗 using Gradio. Try out the online Demo:
• [03/2023] Arxiv paper available.
• [03/2023] Pre-trained models with relation prompts released at this link.
• [03/2023] Project page and video available.
• [03/2023] Inference code released.

1. Clone Repo

   git clone https://github.com/ziqihuangg/ReVersion
   cd ReVersion

2. Create Conda Environment and Install Dependencies

   conda create -n reversion
   conda activate reversion
   conda install python=3.8 pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch
   pip install diffusers["torch"]
   pip install -r requirements.txt

Given a set of exemplar images and their entities' coarse descriptions, you can optimize a relation prompt <R> to capture the co-existing relation in these images, namely Relation Inversion.

1. Prepare the exemplar images (e.g., 0.jpg - 9.jpg) and coarse descriptions (text.json), and put them inside a folder. Feel free to use our ReVersion benchmark, or you can also prepare your own images. An example from our ReVersion benchmark is as follows:

   .reversion_benchmark_v1
   ├── painted_on
   │   ├── 0.jpg
   │   ├── 1.jpg
   │   ├── 2.jpg
   │   ├── 3.jpg
   │   ├── 4.jpg
   │   ├── 5.jpg
   │   ├── 6.jpg
   │   ├── 7.jpg
   │   ├── 8.jpg
   │   ├── 9.jpg
   │   └── text.json
   

2. Take the relation painted_on for example, you can start training using this script:

   accelerate launch \
       --config_file="./configs/single_gpu.yml" \
       train.py \
       --seed="2023" \
       --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
       --train_data_dir="./reversion_benchmark_v1/painted_on" \
       --placeholder_token="<R>" \
       --initializer_token="and" \
       --train_batch_size="2" \
       --gradient_accumulation_steps="4" \
       --max_train_steps="3000" \
       --learning_rate='2.5e-04' --scale_lr \
       --lr_scheduler="constant" \
       --lr_warmup_steps="0" \
       --output_dir="./experiments/painted_on" \
       --save_steps="1000" \
       --importance_sampling \
       --denoise_loss_weight="1.0" \
       --steer_loss_weight="0.01" \
       --num_positives="4" \
       --temperature="0.07"
   

   Where train_data_dir is the path to the exemplar images and coarse descriptions. output_dir is the path to save the inverted relation and the experiment logs. To generate relation-specific images, you can follow the next section Generation.

We can use the learned relation prompt <R> to generate relation-specific images with new objects, backgrounds, and style.

1. You can obtain a learned <R> from Relation Inversion using your customized data. You can also download the models from here, where we provide several pre-trained relation prompts for you to play with.

2. Put the models (i.e., learned relation prompt <R>) under ./experiments/ as follows:

   ./experiments/
   ├── painted_on
   │   ├── checkpoint-500
   │   ...
   │   └── model_index.json
   ├── carved_by
   │   ├── checkpoint-500
   │   ...
   │   └── model_index.json
   ├── inside
   │   ├── checkpoint-500
   │   ...
   │   └── model_index.json
   ...
   

3. Take the relation painted_on for example, you can either use the following script to generate images using a single prompt, e.g., "cat <R> stone":

   python inference.py \
   --model_id ./experiments/painted_on \
   --prompt "cat <R> stone" \
   --placeholder_string "<R>" \
   --num_samples 10 \
   --guidance_scale 7.5
   

   Or write a list prompts in ./templates/templates.py with the key name $your_template_name and generate images for every prompt in the list $your_template_name:

   your_template_name='painted_on_examples'
   python inference.py \
   --model_id ./experiments/painted_on \
   --template_name $your_template_name \
   --placeholder_string "<R>" \
   --num_samples 10 \
   --guidance_scale 7.5
   

   Where model_id is the model directory, num_samples is the number of images to generate for each prompt, and guidance_scale is the classifier-free guidance scale.

   We provide several example templates for each relation in ./templates/templates.py, such as painted_on_examples, carved_by_examples, etc.

• We also provide a Gradio Demo to test our method using a UI. This demo supports relation-specific text-to-image generation on the fly. Running the following command will launch the demo:

  python app_gradio.py
  

• Alternatively, you can try the online demo here.

You can also specify diverse prompts with the relation prompt <R> to generate images of diverse backgrounds and style. For example, your prompt could be "michael jackson <R> wall, in the desert", "cat <R> stone, on the beach", etc.

The ReVersion Benchmark consists of diverse relations and entities, along with a set of well-defined text descriptions.

• Relations and Entities. We define ten representative object relations with different abstraction levels, ranging from basic spatial relations (e.g., “on top of”), entity interactions (e.g., “shakes hands with”), to abstract concepts (e.g., “is carved by”). A wide range of entities, such as animals, human, household items, are involved to further increase the diversity of the benchmark.
• Exemplar Images and Text Descriptions. For each relation, we collect four to ten exemplar images containing different entities. We further annotate several text templates for each exemplar image to describe them with different levels of details. These training templates can be used for the optimization of the relation prompt.
• Benchmark Scenarios. We design 100 inference templates composing of different object entities for each of the ten relations.

If you find our repo useful for your research, please consider citing our paper:

@article{huang2023reversion,
     title={{ReVersion}: Diffusion-Based Relation Inversion from Images},
     author={Huang, Ziqi and Wu, Tianxing and Jiang, Yuming and Chan, Kelvin C.K. and Liu, Ziwei},
     journal={arXiv preprint arXiv:2303.13495},
     year={2023}
}

The codebase is maintained by Ziqi Huang and Tianxing Wu.

This project is built using the following open source repositories:

• Stable Diffusion 1.5
• Diffusers