StableDiffusionWebUI: Detail Repair (ADetailer)
Preface
Sometimes SD-generated images have detail issues, such as faces or hands.
A common approach is to add negative prompts and high-quality positive prompts,
but in some scenarios, issues persist.
Scenario
The problematic scenario mentioned above — for example, when the prompt includes “full body”,
SD allocates more effort to the full body, resulting in mediocre facial quality,
even with many high-quality positive prompts and negative prompt embeddings.
Positive Prompt
(best quality:1), (high quality:1), detailed/(extreme, highly, ultra/), realistic, 1girl/(beautiful, delicate, perfect/), full body
Negative Prompt
VeryBadImageNegative_V1.3 NG_DeepNegative_V1_75T NegativeHand FastNegative_V2 EasyNegative BadPrompt_V2 BadHand_V4 BadDream, (worst quality:1), (low quality:1), (normal quality:1), lowres, signature, blurry, watermark, duplicate, bad link, plump, bad anatomy, extra arms, extra digits, missing finger, bad hands, bad feet, deformed, error, mutation, text
Checkpoint
majicmixRealistic_v7.safetensors
Sampler
DPM++ 2M SDE Heun
Steps
20
Seed
402612351
Result
ADetailer
ADetailer can solve this problem.
ADetailer URL: https://github.com/Bing-su/adetailer
After installation, it will automatically download required models from the internet to the local models/adetailer folder.
Usage
Enable ADetailer and use the face full model.
Process
Keep all other parameters unchanged and generate again.
You can see that after generating the image, it automatically detects the face
and performs a second pass optimization on it.
Result
The facial quality is much better now.
Parallel Repair
As you can see, the hands in the result above are not ideal.
ADetailer supports parallel repair.
In the second unit, select the hand model hand_yolov8n.pt, keeping other parameters unchanged.
Generate again. This time it detects both face and hands, and the result is much better.
However, ADetailer’s strongest suit is still facial repair.
Model Comparison
Keeping the above parameters unchanged, let’s compare the effects of different models.
mediapipe_face_full
mediapipe_face_short
mediapipe_face_mesh
mediapipe_face_mesh_eyes_only
person_yolov8s-seg.pt
person_yolov8n-seg.pt
face_yolov8s.pt
face_yolov8n_v2.pt
face_yolov8n.pt
The results look similar for now. The official site has descriptions of what each model excels at.
Summary
ADetailer is a powerful tool for fixing facial details.
SDWebUI
|—URL: https://sdwebui.ai/?from=vq
|—The original Stable Diffusion WebUI V1.10.1
|—Top-tier 4090 GPU, premium members get exclusive single 4090 for image generation
|—Supports Mac client, Windows client, and web usage
|—Innovative instant model transfer feature for local models
|—Supports txt2img, img2img, extras, and image info
|—Supports CLIP interrogation, DeepBooru interrogation
|—Supports Hires.fix upscaling, with 21 built-in upscale models
|—Supports ADetailer detail repair, with 11 built-in repair models
|—Supports ControlNet, with 44 preprocessors and 68 built-in models
|—Supports ultra upscaling (Tiled Diffusion, Demo Fusion, Tiled VAE)
|—Supports old photo restoration, with 21 upscale models, plus GFPGAN and CodeFormer repair models
|—Built-in 2 common Styles, 11 popular Checkpoint models, 7 popular LoRA models, 14 common embeddings
|—Built-in oldsix prompt plugin
|—Built-in 23-lesson beginner tutorial