CSA Inpainting Studio

CSA Inpainting Studio

Fills in the masked holes of an image and shows two stages of the Coherent Semantic Attention method side by side, a rough result and a refined one. For people studying image inpainting.

Made by KumaPower on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: CSA Inpainting Studio fills in missing regions of an image and lets you compare two stages of the Coherent Semantic Attention method on the same image and mask. One stage is a rough result and the other is a refined one. The Space runs on Hugging Face, so you can try it in a browser with nothing to install.

Who it suits: Researchers and students who study image inpainting, and developers who want to see the method run on their own pictures.

Industry | Who uses it | How they use it | Example
Computer vision | Researchers | Compare a rough stage with its refinement on one mask | Paint a hole and view both results with the wipe slider
Education | Students | See how semantic attention changes a filled-in region | Try the examples in the gallery
Software | Developers | Run a released inpainting bundle on test images | Upload an image and a mask of their own
Image editing | Hobbyists | Remove or repair a patch of a picture | Paint over a flaw and fill it in

Anyone who needs a commercial editing tool can skip it, since the license is non-commercial.

Where to use it: It runs in these places.

• A web browser, on the Space's page at Hugging Face
• Windows, with Python 3.12.12
• macOS, with Python 3.12.12
• Linux, with Python 3.12.12

Use case | Who | What they get back
Compare both stages | Researchers | Two results side by side or under a wipe slider
Try a gallery example | Newcomers | A quick demo without preparing a mask
Use your own mask | Developers | A fill based on a binary mask you upload

How to set it up: Using the Space online needs only a browser. A copy on your own computer needs the following.

• Python 3.12.12 and pip
• The huggingface_hub package, which gives you the hf command
• Disk space for the repository, which holds 7 files

1. Install the huggingface_hub package.
2. Download the Space's files into a local folder.
3. Move into that folder.
4. Install Gradio at the version the Space uses, then install the packages in requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download KumaPower/CSA-Inpainting-Studio --repo-type space --local-dir CSA-Inpainting-Studio
cd CSA-Inpainting-Studio
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page, or run the app on your computer and open the page Gradio shows. Upload an image and paint the holes, or pick an example from the gallery. You may also bring a ready binary mask, or have the app generate one with the training recipe. White mask pixels mark the holes.

python app.py

Tool | What it does
Paint area | Lets you paint the holes on your uploaded image
Gallery | Offers ready examples to start from
Binary mask upload | Takes a mask in which white pixels mark holes
Compare both stages | Shows the rough and refined results side by side or with a wipe slider, and lets you download either PNG

Example: You upload a photo, paint over a small object, and press the compare button. You see the rough fill and the refined fill next to each other, and you can download either one as a PNG.

For newcomers and experts: Inpainting means filling in a missing part of a picture so it blends with the rest. The first small thing to try is a gallery example, so you do not need to make a mask.

The bundle pairs a rough first stage with a refinement stage that uses coherent semantic attention. Inputs are resized to 256 by 256 before inference. Both stages keep the known pixels of that resized input. The example images come from the public CSA and MEDFE releases.

You upload an image and a mask
The image is resized to 256 by 256
The rough stage fills the holes
The refinement stage improves the fill

Good to know: The output size is fixed at 256 by 256, so large photos lose detail.

Title: Training steps of the two released stages
Rough stage 1: 156181 steps
CSA refinement: 97397 steps

• License: cc-by-nc-4
• Cost to run: none named
• Data it reaches: the image and mask you upload, plus the model files that Hugging Face supplies
• Not for: commercial use, which the cc-by-nc-4 license rules out
• Hardware: the Space is set to zero-a10g at Hugging Face, and what your own computer needs is not stated

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: cc-by-nc-4.0

Links

• Space: https://huggingface.co/spaces/KumaPower/CSA-Inpainting-Studio
• Files: https://huggingface.co/spaces/KumaPower/CSA-Inpainting-Studio/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.