Home

Awesome

DreamCom-Image-Composition

This is the official repository for the following research paper:

DreamCom: Finetuning Text-guided Inpainting Model for Image Composition [arXiv]<br>

Lingxiao Lu, Bo Zhang, Li Niu <br>

We divide generative composition methods into token-to-object methods and object-to-object methods (e.g., ObjectStitch, ControlCom). DreamCom belongs to token-to-object methods. Currently, token-to-object methods have not shown clear advantage over object-to-object methods.

Task

Given a few (3~5) reference images for a subject, we aim to customize a generative model, which can insert this subject into an arbitrary background image. In the obtained composite image, the foreground subject should be compatible with the background in terms of illumination, geometry, and semantics.

Technically, we finetune a text-guided inpainting model based on the reference images of one subject, during which a special token is associated with this subject. Then, we apply the finetuned model to a new background image.

<p align='center'> <img src='./figures/task.jpg' width=60% /> </p>

Code and Model

Our code is based on the basic code from the diffusers library. Also, our model's format follows the rules set by the diffusers framework.

  1. Dependencies

    • Python == 3.8

    • Pytorch == 1.11.0

    • Run

       conda env create -f environment.yml

  2. Download Models

    Please download the files and models from https://huggingface.co/runwayml/stable-diffusion-inpainting/tree/main, and put them in the diffusers-v1-5-inpaint/ folder.

  3. Dataset

    You can download our MureCom Dataset, and put it under the folder MureCom/ for training and testing.

  4. Train

    You can execute the following code snippet to initiate the training process for a model. This model is specifically designed to establish a correlation between the unique train located in the MureCom/Train/fg1 directory and the concept represented by "a photo of sks train". The trained model will be subsequently stored within the models/ directory.

    accelerate launch --num_processes=1 train.py --fg_name fg1 --image_num 5 --package_name="Train" --class_name="train" --pretrained_model_name_or_path="diffusers-v1-5-inpaint"  --instance_data_dir="MureCom/" --output_dir="models/" --instance_prompt="a photo of sks " --background_prompt="background" --resolution=512 --train_batch_size=1 --learning_rate=5e-6 --lr_scheduler="constant" --lr_warmup_steps=0 --max_train_steps=400 --gradient_accumulation_steps=1

  5. Test

    You can execute the code provided below to utilize our recently trained model for generating sks train against the backgrounds found in the MureCom/Train/bg directory. The generated results will be appropriately stored in the corresponding location within the MureCom/Train/result/fg1/ directory.

    python test.py --background_prompt="background" --do_crop --package_name="Train" --fg_name fg1 --image_num 5 --class_name="train"

  6. Train and test for all instances in MureCom

    bash train_and_test.sh

Experiments

We show our results compared with some baselines.

<p align='center'> <img src='./figures/result.png' width=90% /> </p>

Our MureCom Dataset

Our MureCom dataset is available in the /MureCom folder. This folder consists of 32 subject subfolders, where each subfolder contains the following data:

Other Resources