How to Restore Faces in Stable Diffusion with ADetailer or Alternatives Easily

ADetailer (After Detailer) is an extension for the Stable Diffusion web UI, designed to enhance image details with inpainting, particularly for faces and hands. It can be used to correct common issues like garbled faces.

But before you dive into steps to restore faces with ADetailer, here's the catch: achieving consistent, high-fidelity face restoration isn't always straightforward. It's less convenient if you want to fix existing images, use on SDXL or large images, or run on low-specs computers.

If you prefer to skip these hurdles and go for a more efficient solution, jump to the bonus part below to restore faces with realistic details, and achieve 4K/10K quality.

How to Restore Faces with After Detailer

Ensure you have the Stable Diffusion web UI installed and then follow the steps below to restore faces with ADetailer. Depending on the specific ADetailer models you are using, you can also enhance hands or the entire character for the AI images.

Step 1. Install ADetailer for Stable Diffusion Automatic1111 (A1111) web UI.

• Open the Extensions tab in your A1111 web UI.
• Click on Install from URL.
• Paste the following GitHub repository URL into the URL field: https://github.com/Bing-su/adetailer.git
• Hit the Install button and wait about 5 seconds until you see the confirmation message: "Installed into stable-diffusion-webui\extensions\adetailer. Use Installed tab to restart."
• Go to the Installed tab, click on Check for updates, then click Apply and restart UI.

Step 2. Install the ADetailer Model

• Download the desired model(s) for ADetailer.
• Move the model file into the stable-diffusion-webui\models\adetailer folder.
• Restart A1111 web UI completely, including the terminal.

Step 3. Configure ADetailer Settings and Parameters.

-After installation and restarting, go to the Extensions tab and click on ADetailer to access its general settings.

-Go to the txt2img or img2img section to configure the ADetailer settings (see detailed explanation below) based on your preferences.

These settings will determine how the tool enhances your images, so customize them to fit your needs.

Step 4. Generate an Image with ADetailer ticked in Stable Diffusion A1111.

Once you configured the parameters, tick the box to Enable ADetailer, then use the A1111 interface to generate an image as you normally would.

ADetailer will automatically enhance the image based on the settings you've configured.

It works like this:

As Stable Diffusion generates the image, ADetailer uses a detection model to identify objects and create a mask. Then, it inpaints the image based on the defined masks.

Important Considerations When Using ADetailer:

While ADetailer is a powerful tool, it has certain limitations:

• Resolution Constraints: It may not effectively handle images with very high resolutions.
• Detail Enhancement: In some cases, it might not restore intricate facial details to the desired level.
• Hardware Requirements: Running ADetailer efficiently requires substantial computational resources.

ADetailer Parameters Explained

These settings allow you to customize how the tool detects and processes elements within your images. You can set an initial value, experiment with it, and fine-tune to achieve the desired results.

Tip for beginners: If the technical details are overwhelming, scroll down for an alternative method to restore faces and enhance image quality.

Detection Model: Choose models trained for specific tasks (e.g., face, hand, or full-body detection). Select a model that aligns with your image's focus.

Detection Model Confidence Threshold: It defines the minimum confidence score required for the model to inpaint a detected object.

For instance, when set to 0.8, only objects detected with ≥80% confidence (e.g., clear faces) are processed. Lower this to include blurry or obscured faces; raise it to exclude background faces.

Mask Min/Max Ratio: It sets the mask size limits for detected objects relative to the entire image.

X/Y Offset: It adjusts mask positioning horizontally (X) or vertically (Y) to refine alignment with the target area.

Mask Erosion shrinks mask edges to avoid including unintended background pixels. Mask Dilation does the opposite.

For inpainting parameters, you can adjust inpaint mask blur, denoising strength, padding and more. Detailed explanations can be found in ADetailer repository on Github.

Also, you can use the ControlNet extension if you have ControlNet installed and ControlNet models. ADetailer can be used with ControlNet inpaint, scribble, lineart, openpose, tile, depth controlnet models.

How to Restore Faces with ADetailer Alternative for Better Quality

If you've struggled with ADetailer's technical complexity, hardware demands, or inconsistent facial detailing, consider exploring VideoProc Converter AI, a one-stop video and image enhancer software, with media toolkit built-in.

It is specifically designed to upscale flawed images, restore intricate facial details, and enhance with greater accuracy.

VideoProc Converter AI does all of the following, where ADetailer falls short:

• Improve existing images, batch process images in one-go
• Generate natural details for skin, hair, eyes, pores, etc.
• Restore faces without altering the original facial structure
• Optimized for commercial PC/laptop, low-RAM friendly

How to Use VideoProc Converter AI for Face Restoration

Designed for beginners and professionals, its Image AI module delivers superior face restoration with batch processing, 10K upscaling, and adaptive AI models—all optimized for everyday hardware.

Before getting started, download and install VideoProc Converter AI on your computer.

Step 1. Launch the program and go to the Image AI module.

This dedicated workspace combines super-resolution, face restoration, and colorization tools for professional-grade results.

Step 2. Import flawed images into the working panel.

Unlike ADetailer's manual workflows, VideoProc supports batch processing—restore dozens of photos in one session.

Step 3. Use the Face Restoration feature.

You can also use AI models to upscale the images to 4K/10K upon face restoration.

Select Face Detection Mode:

-High Quality Faces: Prioritizes faces with minimal damage (e.g., slightly blurry portraits).

-Center Face: Focuses on the central subject in group photos.

-All Faces: Restores every detected face, ideal for crowded shots.

💡 Below are before and after examples with face restored in VideoProc Converter AI.

As you can see, those cracked lines and other artifacts in AI generated image is fixed by VideoProc. Better yet, since we enabled the AI upscaling algorithms, the output image has 4K clarity with newly generated details. It won't become blurry or losing details when zooming in.

It works for faces from AI art and photos. You can import existing images into this program to batch enhance the faces and the overall quality of the images.

Step 4. Batch Export enhanced images.

When you are happy with the settings, you can hover over the image thumbnails at the bottom panel, right-click and select Apply Settings to All to replicate settings across multiple images.

Click Run to export the images.

While ADetailer offers valuable enhancements for Stable Diffusion images, it may not meet all user needs, especially concerning high-resolution support, detailed restoration, and hardware requirements. VideoProc Converter AI addresses these limitations, providing a robust solution for AI image generation enthusiasts seeking high-quality face restoration.

Why VideoProc Converter AI Is the Best Bet?

• 4K Face Recovery: Restores faces and the images to 4K/10K resolution with natural textures.
• Enhanced AI Models with Details: Combines proprietary algorithms for lifelike skin, eyes, and hair details, you can even see pores on faces clearly.
• Batch Processing: Efficiently handles multiple images simultaneously.
• Toolkitfor seamless workflow: Besides Image AI, VideoProc also has video enhancer and processing features. You can use it to convert, edit, record, download and restore videos/DVDs/audio files.

Click buttons below to download VideoProc Converter AI to restore faces and generate more details for images:

FAQs

Why are ADetailer files marked as unsafe on HuggingFace?

While not explicitly addressed in the comments, HuggingFace often flags files as "unsafe" due to automated security scanners detecting binary files (e.g., PyTorch model weights) as potential risks. This is a common false positive. Users should verify the source (e.g., trusted repositories like Bingsu/adetailer) and ensure they download files from official channels.

How can I fix only one face in a scene with multiple subjects?

You can experiment with the following settings: 1. Increase the Confidence Threshold (e.g., set to 0.8) to prioritize the most prominent face. 2. Use mask adjustments (Erosion/Dilation or X/Y Offset) to manually refine the area. 3. Run ADetailer multiple times with different settings for each face (requires manual masking).

For easier improvement, you can use the Face Restoration feature in VideoProc Converter AI. It has robust AI models to restore, upscale and enhance faces in 3 steps. You can select to improve center face or all faces.

How do I include an existing character image in the ADetailer workflow?

ADetailer is designed for inpainting generated images, not editing existing ones. To use it with a pre-existing image, though you can generate a rough version of the character in Stable Diffusion with same prompts, seeds, and other settings, such as use ControlNet to guide generation toward your existing character.

What denoising strength is recommended for inpainting faces?

The default values work for average case, or you can experiment with these settings: Low denoising value retains original facial structure, ideal for minor fixes.

Does the ADetailer model work with SDXL?

At the time of writing this blog, there are no ideal models for SDXL. Based on my test, the hand model produces "very strange things" in SDXL. This suggests compatibility issues, as SDXL's architecture and resolutions differ from Stable Diffusion 1.5/2.1.