TencentARC/MasaCtrl

MasaCtrl: Tuning-free Mutual Self-Attention Control for Consistent Image Synthesis and Editing

Pytorch implementation of MasaCtrl: Tuning-free Mutual Self-Attention Control for Consistent Image Synthesis and Editing

Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, Yinqiang Zheng

MasaCtrl enables performing various consistent non-rigid image synthesis and editing without fine-tuning and optimization.

Updates

• [2023/5/13] The inference code of MasaCtrl with T2I-Adapter is available.
• [2023/4/28] Hugging Face demo released.
• [2023/4/25] Code released.
• [2023/4/17] Paper is available here.

Introduction

We propose MasaCtrl, a tuning-free method for non-rigid consistent image synthesis and editing. The key idea is to combine the contents from the source image and the layout synthesized from text prompt and additional controls into the desired synthesized or edited image, with Mutual Self-Attention Control.

Main Features

1 Consistent Image Synthesis and Editing

MasaCtrl can perform prompt-based image synthesis and editing that changes the layout while maintaining contents of source image.

The target layout is synthesized directly from the target prompt.

Consistent synthesis results Real image editing results

2 Integration to Controllable Diffusion Models

Directly modifying the text prompts often cannot generate target layout of desired image, thus we further integrate our method into existing proposed controllable diffusion pipelines (like T2I-Adapter and ControlNet) to obtain stable synthesis and editing results.

The target layout controlled by additional guidance.

Synthesis (left part) and editing (right part) results with T2I-Adapter

3 Generalization to Other Models: Anything-V4

Our method also generalize well to other Stable-Diffusion-based models.

Results on Anything-V4

4 Extension to Video Synthesis

With dense consistent guidance, MasaCtrl enables video synthesis

Video Synthesis Results (with keypose and canny guidance)

Usage

Requirements

We implement our method with diffusers code base with similar code structure to Prompt-to-Prompt. The code runs on Python 3.8.5 with Pytorch 1.11. Conda environment is highly recommended.

pip install -r requirements.txt

Checkpoints

Stable Diffusion: We mainly conduct expriemnts on Stable Diffusion v1-4, while our method can generalize to other versions (like v1-5).

You can download these checkpoints on their official repository and Hugging Face.

Personalized Models: You can download personlized models from CIVITAI or train your own customized models.

Demos

Notebook demos

To run the synthesis with MasaCtrl, single GPU with at least 16 GB VRAM is required.

The notebook playground.ipynb and playground_real.ipynb provide the synthesis and real editing samples, respectively.

Online demos

We provide with Gradio app. Note that you may copy the demo into your own space to use the GPU. Online Colab demo is also available.

Local Gradio demo

You can launch the provided Gradio demo locally with

CUDA_VISIBLE_DEVICES=0 python app.py

MasaCtrl with T2I-Adapter

Install T2I-Adapter and prepare the checkpoints following their provided tutorial. Assuming it has been successfully installed and the root directory is T2I-Adapter.

Thereafter copy the core masactrl package and the inference code masactrl_w_adapter.py to the root directory of T2I-Adapter

cp -r MasaCtrl/masactrl T2I-Adapter/
cp MasaCtrl/masactrl_w_adapter/masactrl_w_adapter.py T2I-Adapter/

Last, you can inference the images with following command (with sketch adapter)

python masactrl_w_adapter.py \
--which_cond sketch \
--cond_path_src SOURCE_CONDITION_PATH \
--cond_path CONDITION_PATH \
--cond_inp_type sketch \
--prompt_src "A bear walking in the forest" \
--prompt "A bear standing in the forest" \
--sd_ckpt models/sd-v1-4.ckpt \
--resize_short_edge 512 \
--cond_tau 1.0 \
--cond_weight 1.0 \
--n_samples 1 \
--adapter_ckpt models/t2iadapter_sketch_sd14v1.pth

NOTE: You can download the sketch examples here.

For real image, the DDIM inversion is performed to invert the image into the noise map, thus we add the inversion process into the original DDIM sampler. You should replace the original file T2I-Adapter/ldm/models/diffusion/ddim.py with the exteneded version MasaCtrl/masactrl_w_adapter/ddim.py to enable the inversion function. Then you can edit the real image with following command (with sketch adapter)

python masactrl_w_adapter.py \
--src_img_path SOURCE_IMAGE_PATH \
--cond_path CONDITION_PATH \
--cond_inp_type image \
--prompt_src "" \
--prompt "a photo of a man wearing black t-shirt, giving a thumbs up" \
--sd_ckpt models/sd-v1-4.ckpt \
--resize_short_edge 512 \
--cond_tau 1.0 \
--cond_weight 1.0 \
--n_samples 1 \
--which_cond sketch \
--adapter_ckpt models/t2iadapter_sketch_sd14v1.pth \
--outdir ./workdir/masactrl_w_adapter_inversion/black-shirt

NOTE: You can download the real image editing example here.

Acknowledgements

We thank the awesome research works Prompt-to-Prompt, T2I-Adapter.

Citation

@misc{cao2023masactrl,
      title={MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing}, 
      author={Mingdeng Cao and Xintao Wang and Zhongang Qi and Ying Shan and Xiaohu Qie and Yinqiang Zheng},
      year={2023},
      eprint={2304.08465},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
    }

Contact

If your have any comments or questions, please open a new issue or feel free to contact Mingdeng Cao and Xintao Wang.