ADetailer 2026 Fix AI Image Faces Automatically

ADetailer (After Detailer) fixes a common problem in AI image generation. When you create images at 512px or 768px, faces often look soft or wrong even when the rest of the image looks great. ADetailer solves this by finding faces automatically and fixing them right after the main image is created. You do not need to draw any masks by hand.

This guide covers how to install ADetailer, the best settings for adult content, and how to use it with Hires.fix and ControlNet to get the highest quality results.

What you need before you start: A working AUTOMATIC1111 setup with at least one checkpoint model. ADetailer works with any checkpoint including adult-trained models. The extension itself has no content limits. A GPU is recommended because ADetailer adds a quick fix pass after each image.

How to Install ADetailer

Go to Extensions > Available > Load From. Search for “ADetailer.” Find the extension by Bing-su and click Install. When the install finishes, go to the Installed tab and click “Apply and restart UI.” The ADetailer panel will appear below the main settings in both txt2img and img2img. Check the Enable box inside the panel to turn it on for each generation.

clothing remover ai

ADetailer comes with several detection models that download automatically the first time you use them. Here are the key options for adult content generation:

The Most Important Setting: Denoising Strength

Denoising strength is the single most ai porn generation important ADetailer setting. Here is what each range does:

• 0.3 to 0.4: Subtle enhancement. This sharpens details while keeping the original face identity. Use this when the face is mostly correct and just needs clarity.
• 0.45 to 0.55: Moderate regeneration. This improves quality a lot and can fix small proportion problems. There is a small risk that the face might look slightly different on low-detail images.
• 0.6 and above: Aggressive regeneration. This essentially redraws the face. Quality improves greatly but the result may not match the original character.

For most adult content work, 0.4 to 0.5 is the best range.

How to Combine ADetailer with Hires.fix

ADetailer and Hires.fix work great together. Here is the recommended workflow:

1. Generate your image at 512×768 (or the native resolution for your checkpoint).
2. Enable Hires.fix at 1.5x with 0.5 denoising for overall sharpness.
3. Enable ADetailer at 0.45 denoising for face and body detail.

Hires.fix runs first to improve the overall image. Then ADetailer runs on the sharpened result. Total generation time increases by 30 to 60 percent, but the quality improvement is significant for adult content where face and body detail matters.

Using Multiple ADetailer Units

AUTOMATIC1111 supports running multiple ADetailer units one after another. A common workflow for adult content:

• Unit 1: anime_face_yolov8.pt at 0.45 denoising (for face detail).
• Unit 2: hand_yolov8n.pt at 0.4 denoising (for hand correction).

Enable Unit 2 by opening the second ADetailer panel below the first. Each unit runs in order, so face improvement happens before the hand fix pass.

Understanding Each ADetailer Setting

Knowing what each setting does helps you tune ADetailer precisely instead of using it as a black box.

Detection Confidence (0 to 1): This is the minimum confidence score a detected region must reach to trigger fixing. At 0.3 (the default), ADetailer processes most visible faces including partial and angled faces. Raise to 0.5 or higher if ADetailer is incorrectly fixing non-face elements like decorative patterns or background objects. Lower to 0.2 if faces in the background are not being detected.

Mask Erosion and Dilation: Positive dilation values expand the mask outward from the detected face edge. Negative values shrink it inward. For face enhancement, slight dilation (4 to 8 pixels) helps blend the fixed area with the surrounding hair and neck. For targeted fixes like eyes only or mouth only, use negative dilation to shrink the mask to just that feature.

Inpaint Padding: This adds context around the masked area during fixing. Higher padding (32 to 64 pixels) gives the fixing model more surrounding context to blend with, which usually improves edge quality. Lower padding (0 to 16 pixels) speeds up processing. Start at 32 for best quality.

Use Separate Prompt: When checked, you can write prompts specifically for the ADetailer fix pass, separate from the main generation prompt. Use this when you want to give the face fixing specific instructions that do not fit the full image prompt.

Batch Generation with ADetailer

ADetailer runs on every image in a batch. If you generate 8 images with ADetailer enabled, all 8 get the face enhancement pass. This is the most efficient way to use ADetailer. Run your normal batch and get 8 enhanced images instead of 8 rough images that need manual selection and separate fixing.

The time cost per image is modest: approximately 5 to 15 seconds per detected face on a mid-range GPU like the RTX 3080 or 4070.

Production Workflow for Consistent Results

For production workflows where you need large batches of consistent adult character images, use this setup:

• txt2img with your standard prompt.
• ADetailer (anime_face_yolov8 at 0.45 denoising).
• Hires.fix (1.5x at 0.4 denoising).

This single generation pass produces near-final quality images that need very little post-processing. Total generation time per image is 60 to 120 seconds on consumer GPU hardware. This is comparable to generating and then manually fixing separately.

How ADetailer Works with LoRAs

ADetailer uses the active LoRAs from your main generation. If you have a character LoRA active at 0.75 strength in your main prompt, that LoRA stays active during the ADetailer face fix pass. This means the character’s face identity is reinforced during enhancement rather than drifting. The result is sharper, more detailed faces that still look like your specific character instead of a generic improved face.

If you want the ADetailer pass to use a different LoRA strength than the main generation, check “Use Separate LoRA” in the ADetailer settings and set the LoRA strength independently for the fix pass. This is useful when your character LoRA strength is set for body and costume rendering, but you want a slightly different balance for face detail.

Beyond Faces: Hand and Body Detection

ADetailer is not limited to faces. With the right detection model, it can automatically detect and enhance any body area.

Hand detection (hand_yolov8n.pt): Hand anatomy is one of the most common AI generation failure points. Enabling a second ADetailer unit with hand detection at 0.5 denoising greatly reduces malformed fingers. The hand detection model finds hand regions and runs a corrective fix pass with the positive prompt reinforcing correct anatomy. This cannot fix every hand failure, but it eliminates the minor to moderate failures that cause most hand quality issues.

Custom region detection: Community-trained detection models for other body areas are available on Civitai and Hugging Face. Search for “ADetailer model” on Civitai to find detection models for specific body areas. Place downloaded model files in the ADetailer models folder and they will appear in the model dropdown.

Common Problems and Solutions

ADetailer is not running: The Enable checkbox inside the ADetailer panel must be checked each session. It does not persist across restarts unless set in default settings. Check that the panel is expanded and the Enable toggle is on. In AUTOMATIC1111 settings, you can set ADetailer to enable by default under Settings > ADetailer.

Wrong region being detected: ADetailer is detecting a background element as a face. Raise the Detection Confidence threshold from 0.3 to 0.5 or higher to filter out low-confidence detections. Alternatively, mask the generation to keep the main character separate from complex background elements that might confuse the detector.

ADetailer is too slow: Running ADetailer on large batches at high resolution can be slow. Reduce the inpainting resolution to 512px square, which is sufficient for face enhancement, rather than matching the full image resolution. Use the “Inpaint Width” and “Inpaint Height” settings to cap the fix size.

Character face identity keeps changing: Your denoising is too high. Drop from 0.5 to 0.38 to 0.42. At very low denoising (0.35), ADetailer works more like a sharpening pass than a regeneration. For maintaining a distinctive character face across many generations, 0.38 to 0.42 denoising with the character LoRA active is the most reliable setting.

Quick Start Checklist

Here is the fastest way to get ADetailer working with the best default settings for adult anime generation:

1. Install via Extensions > Available, search “ADetailer,” install, and restart.
2. In the ADetailer panel, check Enable.
3. Detection model: anime_face_yolov8.pt for anime checkpoints, face_yolov8n.pt for realistic checkpoints.
4. Confidence: 0.3.
5. Mask blur: 6.
6. Denoising: 0.45.
7. Leave prompts blank to inherit from main generation.

Enable this on every generation. ADetailer adds minimal time (5 to 15 seconds per face) and almost always improves face quality compared to the same generation without it.

For users generating adult content with multiple characters per image: enable “Save detected map” in ADetailer settings to see which faces were detected and processed. This helps diagnose cases where a secondary character’s face was missed.

Final Thoughts

ADetailer is one of the most impactful extensions available for AUTOMATIC1111. It delivers clear quality improvement with minimal setup. The defaults described in this guide (anime_face_yolov8.pt, 0.45 denoising, 6px mask blur) work well across the vast majority of adult anime generation scenarios.

Enable it on every generation session and treat the few seconds of extra processing time as a free quality upgrade. For users who want the complete quality toolkit, ADetailer combined with ControlNet pose control and targeted inpainting covers the three main quality areas in AI image generation: enhancement (ADetailer), structure (ControlNet), and correction (inpainting).