IOPaint - Image inpainting tool powered by SOTA AI Model. Remove any unwanted object, defect, people from your pictures or erase and replace(powered by stable diffusion) any thing on your pictures.

Created at: 2021-11-15 22:16:40
Language: Python
License: Apache-2.0

Lama Cleaner

A free and open-source inpainting tool powered by SOTA AI model.

total download version Open in Colab Hugging Face Spaces python version version



A great introductory youtube video made by ** Aitrepreneur**

1. Remove any unwanted things on the image
Usage Before After
Remove unwanted things unwant_object2 unwant_object2
Remove unwanted person unwant_person unwant_person
Remove text text text
Remove watermark watermark watermark_clean
Remove text balloons on manga manga manga_clean
2. Fix old photo
Usage Before After
Fix old photo oldphoto oldphoto_clean
3. Replace something on the image


Usage Before After
Text Driven Inpainting dog Prompt: a fox sitting on a bench

Paint by Example

Original Image Example Image Result Image

Quick Start

The easiest way to use Lama Cleaner is to install it using pip:

pip install lama-cleaner

# Models will be downloaded at first time used
lama-cleaner --model=lama --device=cpu --port=8080
# Lama Cleaner is now running at http://localhost:8080

If you prefer to use docker, you can check out docker

If you hava no idea what is docker or pip, please check One Click Installer

Available command line arguments:

Name Description Default
--model lama/ldm/zits/mat/fcf/sd1.5/manga/sd2/paint_by_example See details in Inpaint Model lama
--sd-disable-nsfw Disable stable-diffusion NSFW checker.
--sd-cpu-textencoder Always run stable-diffusion TextEncoder model on CPU.
--sd-enable-xformers Enable xFormers optimizations. See: facebookresearch/xformers
--local-files-only Once the model as downloaded, you can pass this arg to avoid diffusers connect to Hugging Face server
--cpu-offload sd/paint_by_example model, offloads all models to CPU, sacrifice speed for reducing vRAM usage.
--no-half Using full precision for sd/paint_by_exmaple model
--device cuda / cpu / mps cuda
--port Port for backend flask web server 8080
--gui Launch lama-cleaner as a desktop application
--gui_size Set the window size for the application 1200 900
--input Path to image you want to load by default None
--debug Enable debug mode for flask web server

Inpainting Model

Model Description Config
cv2 👍 No GPU is required, and for simple backgrounds, the results may even be better than AI models.
LaMa 👍 Generalizes well on high resolutions(~2k)
LDM 👍 Possible to get better and more detail result
👍 The balance of time and quality can be achieved by adjusting steps
😐 Slower than GAN model
😐 Need more GPU memory
Steps: You can get better result with large steps, but it will be more time-consuming
Sampler: ddim or plms. In general plms can get better results with fewer steps
ZITS 👍 Better holistic structures compared with previous methods
😐 Wireframe module is very slow on CPU
Wireframe: Enable edge and line detect
FcF 👍 Better structure and texture generation
😐 Only support fixed size (512x512) input
SD1.5 👍 SOTA text-to-image diffusion model
See model comparison detail

LaMa vs LDM

Original Image LaMa LDM
photo-1583445095369-9c651e7e5d34 photo-1583445095369-9c651e7e5d34_cleanup_lama photo-1583445095369-9c651e7e5d34_cleanup_ldm

LaMa vs ZITS

Original Image ZITS LaMa
zits_original zits_compare_zits zits_compare_lama

Image is from ZITS paper. I didn't find a good example to show the advantages of ZITS and let me know if you have a good example. There can also be possible problems with my code, if you find them, please let me know too!

LaMa vs FcF

Original Image LaMa FcF
texture texture_lama texture_fcf

LaMa vs Manga

Manga model works better on high-quality manga image then LaMa model.

Original Image manga

Model 1080x740 1470x1010
Manga manga_1080x740 manga_1470x1010
LaMa lama_1080x740 lama_1470x1010

Inpainting Strategy

Lama Cleaner provides three ways to run inpainting model on images, you can change it in the settings dialog.

Strategy Description VRAM Speed
Original Use the resolution of the original image High
Resize Resize the image to a smaller size before inpainting. The area outside the mask will not loss quality. Midium
Crop Crop masking area from the original image to do inpainting Low

Download Model Manually

If you have problems downloading the model automatically when lama-cleaner start, you can download it manually. By default lama-cleaner will load model from TORCH_HOME=~/.cache/torch/hub/checkpoints/, you can set TORCH_HOME to other folder and put the models there.


Only needed if you plan to modify the frontend and recompile yourself.


Frontend code are modified from, You can experience their great online services here.

  • Install dependencies:cd lama_cleaner/app/ && yarn
  • Start development server: yarn start
  • Build: yarn build


You can use pre-build docker image to run Lama Cleaner. The model will be downloaded to the cache directory when first time used. You can mount existing cache directory to start the container, so you don't have to download the model every time you start the container.

The cache directories for different models correspond as follows:

  • lama/ldm/zits/mat/fcf: /root/.cache/torch
  • sd1.5: /root/.cache/huggingface

Run Docker (cpu)

docker run -p 8080:8080 \
-v /path/to/torch_cache:/root/.cache/torch \
-v /path/to/huggingface_cache:/root/.cache/huggingface \
--rm cwq1913/lama-cleaner:cpu-0.26.1 \
lama-cleaner --device=cpu --port=8080 --host=

Run Docker (gpu)

  • cuda11.6
  • pytorch1.12.1
  • minimum nvidia driver 510.39.01+
docker run --gpus all -p 8080:8080 \
-v /path/to/torch_cache:/root/.cache/torch \
-v /path/to/huggingface_cache:/root/.cache/huggingface \
--rm cwq1913/lama-cleaner:gpu-0.26.1 \
lama-cleaner --device=cuda --port=8080 --host=

Then open http://localhost:8080

Build Docker image

cpu only

docker build -f --build-arg version=0.x.0 ./docker/CPUDockerfile -t lamacleaner .

gpu & cpu

docker build -f --build-arg version=0.x.0 ./docker/GPUDockerfile -t lamacleaner .