---
id: wiki-2026-0508-스테이블-디퓨전을-이용한-오픈소스-기반-정밀-이미지-합성-
title: 스테이블 디퓨전 기반 정밀 이미지 합성 및 해부학적 오류 수정 파이프라인
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [SD Anatomy Fix Pipeline, ControlNet Anatomy, AfterDetailer, ADetailer]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [stable-diffusion, controlnet, adetailer, comfyui, anatomy, inpaint]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: ComfyUI/diffusers/SDXL/FLUX
---

# 스테이블 디퓨전 기반 정밀 합성 + 해부학적 오류 수정

## 매 한 줄
> **"매 anatomy fix 의 본질 은 detection + inpaint 의 loop"**. 매 SDXL/FLUX base 의 raw output 에서 hand/face/text 의 결함 은 inevitable — 매 ADetailer 의 auto-detect, 매 ControlNet 의 pose/depth lock, 매 second-pass inpaint 가 매 결합되어 production-quality 의 reliable pipeline 의 형성. 매 2026 의 open-source workflow 는 ComfyUI 의 graph + node API 가 standard.

## 매 핵심

### 매 problem space
- **Hand 의 6-finger / fused digits**: SD/SDXL 의 chronic. FLUX.1 dev 에서 의 상당 개선, but 의 not perfect.
- **Face degradation at low pixel**: 매 small face crop 에서 의 detail loss → ADetailer 가 face crop → high-res inpaint.
- **Text 의 illegible**: SDXL 의 weak. FLUX 의 strong (그러나 아직 결함 있음).
- **Eye 의 asymmetry**: gaze direction, pupil size 의 mismatch.

### 매 도구 (open source)
- **ComfyUI**: 매 node-based workflow. 매 reproducible JSON.
- **A1111 / Forge / reForge**: 매 web UI. 매 ADetailer extension 의 default.
- **ControlNet**: pose/depth/canny/openpose/MediaPipeFace 의 conditioning.
- **ADetailer**: face/hand 의 auto-detect → inpaint.
- **MeshGraphormer / DWPose**: 매 hand pose 의 estimation → ControlNet hand_refiner.
- **IP-Adapter FaceID Plus v2**: 매 face consistency.

### 매 pipeline 구성
1. **Base generation** (SDXL/FLUX) — high-level composition.
2. **ADetailer face pass** — face crop → upscale → inpaint with same prompt.
3. **ADetailer hand pass** — DWPose 검출 → ControlNet hand_refiner → inpaint.
4. **Manual touch-up** — 잔여 결함 의 mask + inpaint.
5. **Upscale** — Real-ESRGAN / SUPIR / FLUX-Upscale.

### 매 응용
1. Character art (anime, realistic portrait).
2. Fashion editorial (pose-precise).
3. Comic / manga panel.
4. Game asset (consistent character sheets).

## 💻 패턴

### Pattern 1 — ADetailer pipeline (diffusers)
```python
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLInpaintPipeline
from ultralytics import YOLO
import torch

base = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
inpaint = StableDiffusionXLInpaintPipeline.from_pretrained(
    "diffusers/stable-diffusion-xl-1.0-inpainting-0.1", torch_dtype=torch.float16
).to("cuda")
face_det = YOLO("face_yolov8n.pt")

def adetailer_face(img, prompt):
    boxes = face_det(img)[0].boxes.xyxy.cpu().numpy()
    for x1,y1,x2,y2 in boxes:
        crop = img.crop((x1,y1,x2,y2)).resize((1024,1024))
        mask = make_face_mask(crop)
        fixed = inpaint(prompt=prompt, image=crop, mask_image=mask,
                        num_inference_steps=25, strength=0.5).images[0]
        img.paste(fixed.resize((int(x2-x1),int(y2-y1))), (int(x1),int(y1)))
    return img
```

### Pattern 2 — Hand fix with DWPose + ControlNet
```python
from controlnet_aux import DWposeDetector
from diffusers import StableDiffusionXLControlNetInpaintPipeline, ControlNetModel

dwpose = DWposeDetector.from_pretrained("yzd-v/DWPose")
hand_cn = ControlNetModel.from_pretrained("hr16/ControlNet-HandRefiner-pruned",
                                          torch_dtype=torch.float16)

def fix_hands(img, prompt):
    pose = dwpose(img, hand_only=True)            # depth-style hand map
    mask = hand_mask_from_dwpose(pose)
    pipe = StableDiffusionXLControlNetInpaintPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0",
        controlnet=hand_cn, torch_dtype=torch.float16
    ).to("cuda")
    return pipe(prompt=prompt, image=img, mask_image=mask,
                control_image=pose, num_inference_steps=30,
                controlnet_conditioning_scale=0.9).images[0]
```

### Pattern 3 — ControlNet OpenPose lock
```python
from controlnet_aux import OpenposeDetector
from diffusers import StableDiffusionXLControlNetPipeline

op = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
pose_map = op(reference_image, hand_and_face=True)

cn = ControlNetModel.from_pretrained("xinsir/controlnet-openpose-sdxl-1.0",
                                     torch_dtype=torch.float16)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=cn, torch_dtype=torch.float16
).to("cuda")
out = pipe(prompt="cinematic shot of a woman dancing", image=pose_map,
           num_inference_steps=30, guidance_scale=6.5).images[0]
```

### Pattern 4 — Multi-ControlNet (depth + canny + pose)
```python
from diffusers import MultiControlNetModel
controlnets = MultiControlNetModel([
    ControlNetModel.from_pretrained("xinsir/controlnet-depth-sdxl-1.0",  torch_dtype=torch.float16),
    ControlNetModel.from_pretrained("xinsir/controlnet-canny-sdxl-1.0",  torch_dtype=torch.float16),
    ControlNetModel.from_pretrained("xinsir/controlnet-openpose-sdxl-1.0", torch_dtype=torch.float16),
])
result = pipe(prompt=p,
              image=[depth_map, canny_map, pose_map],
              controlnet_conditioning_scale=[0.6, 0.4, 0.9]).images[0]
```

### Pattern 5 — Differential diffusion (soft mask)
```python
# Soft strength per-pixel mask: 0.0 keep, 1.0 fully regen
import numpy as np
from PIL import Image

def soft_mask(face_box, img_size, edge_blur=24):
    m = np.zeros(img_size, dtype=np.float32)
    x1,y1,x2,y2 = face_box
    m[y1:y2, x1:x2] = 1.0
    from scipy.ndimage import gaussian_filter
    return Image.fromarray((gaussian_filter(m, edge_blur)*255).astype(np.uint8))

# Pass into pipe.diff_diffusion_strength=
```

### Pattern 6 — Face restore (CodeFormer / GFPGAN)
```python
from codeformer import CodeFormer
cf = CodeFormer(weight_path="codeformer.pth", device="cuda")
restored = cf.enhance(img_np, w=0.7)  # 0=keep id, 1=full repair
```

### Pattern 7 — ComfyUI workflow JSON snippet
```json
{
  "1": {"class_type": "KSampler",
        "inputs": {"steps": 30, "cfg": 6.5, "sampler_name": "dpmpp_2m_sde",
                   "scheduler": "karras", "model": ["4",0], "positive":["6",0],
                   "negative":["7",0], "latent_image":["5",0]}},
  "20":{"class_type":"FaceDetailer",
        "inputs":{"image":["1",0],"model":["4",0],
                  "bbox_detector":"face_yolov8m.pt",
                  "wildcard":"perfect symmetric eyes, sharp focus",
                  "guide_size":768,"max_size":1024,"steps":20}}
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Anime portrait, hand 결함 | A1111 + ADetailer (face_yolov8n + hand_yolov8n) |
| Realistic full body | SDXL + Multi-ControlNet (pose+depth) |
| Reproducible production | ComfyUI workflow JSON + Git |
| Maximum quality, slow | FLUX.1 dev + Differential Diffusion |
| Real-time iteration | SDXL Lightning / Turbo |

**기본값**: 매 SDXL base + ADetailer face/hand pass + 1 manual inpaint round.

## 🔗 Graph
- 부모: [[AI 이미지 생성 (AI Image Generation)]] · [[Stable Diffusion]]
- 변형: [[ControlNet]] · [[ADetailer]] · [[IP-Adapter]]
- 응용: [[사후 편집 (Post-editing)]] · [[Brand Consistency Maintenance|Character Consistency]]
- Adjacent: [[FLUX]] · [[ComfyUI]]

## 🤖 LLM 활용
**언제**: 매 ComfyUI workflow JSON 의 authoring/debugging, 매 prompt + negative-prompt 의 systematic generation.
**언제 X**: 매 fine pixel-level inpaint 결정 — 매 visual judgment 가 필요.

## ❌ 안티패턴
- **Single-pass at full res**: 매 face/hand 의 detail starvation. 매 second-pass crop+upscale 의 mandatory.
- **Wrong ControlNet 조합**: 매 pose + canny 동시에 high weight → 매 over-constrain → composition 의 collapse.
- **Hand_refiner 없는 SDXL hand**: 매 5% 미만 의 success rate.
- **ADetailer denoise 1.0**: 매 identity 의 destroyed. 매 0.4–0.6 의 sweet spot.

## 🧪 검증 / 중복
- Verified (Mikubill/sd-webui-controlnet, Bing-su/adetailer; HuggingFace diffusers docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — ADetailer + ControlNet hand_refiner pipeline |