🛠️ Method:

이미지 생성 프로세스

어떠한 사진들을 만들 것인가?

중립(Neutral) 사진 - 기준 이미지를 제작
기타 표정으로 변환 - Angry, Sad, Fear 등의 표정으로 변환.

사용 도구

어떻게 만들 것인가?

온라인 Tools: 인터넷에서 제공하는 전용 툴들을 활용
로컬: 직접 Stable Diffusion을 로컬에서 실행
nano-banana

🔧 Procedure:

1. Stable Diffusion Settings

얼굴 표정을 변환하기 위한 모델

Model: Juggernaut XL Ragnarok.safetensors (dd08fa32f9)
Mode: Inpainting
LoRA: Tuning Face Detailer v2.0
LoRA Weight: 0.5

Prompt

📸 Stable Diffusion prompt 스크린샷

Positive-Prompt (Token: ??/150)

A close-up portrait of the same male with a mildly angry expression. The person is frowning slightly, eyebrows gently furrowed, lips pressed, showing a hint of annoyance. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that looks like an authentic subtle frown. <lora:Tuning Face Detailer v2.0 Rank 256:1>

Negative-Prompt (Token: 44~46/75)

smile, neutral expression, lowres, blurry, cartoon, painting, overexposed, underexposed, background distortion, identity change, extra artifacts, watermark, text, deformed eyebrows, plastic skin, exaggerated wrinkles, different eye colors, mismatched eye colors, averted gaze, looking away

Caution!

성별 정보 가 담긴 부분에서 주의!!
ex) his, male, etc
또, smile 표정을 만들 때에는 neg-prompt에서 smile을 제거.

server option : —opt-sdp-attention

1. Angry face

Positive-Prompt (Token: 94/150)

A close-up portrait of the same ==[male] showing a moderate expression of anger==. [His] eyebrows are drawn downward and inward with a noticeable crease between them. [His] eyes are slightly narrowed, adding tension to his gaze. The lips are pressed firmly together, and the jaw appears set, suggesting controlled frustration. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that captures a restrained but clearly visible anger. <lora:Tuning Face Detailer v2.0 Rank 256:0.5>

Positive-Prompt (Token: ??/150) - refining

A close-up portrait of the same ==[male] showing a moderate expression of anger==. [His] eyebrows are drawn downward and inward with a noticeable crease between them. [His] eyes are slightly narrowed, adding tension to his gaze. The lips are pressed firmly together, and the jaw appears set, suggesting controlled frustration. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that captures a restrained but clearly visible anger. <lora:Tuning Face Detailer v2.0 Rank 256:0.5>

2. Happy face

Positive-Prompt (Token: 94/150)

A close-up portrait of the same [male] with a moderately intense smile. The corners of his lips are clearly lifted into a confident curve, and the muscles around his eyes engage enough to cause slight crow’s feet. [His] cheeks are noticeably raised, giving his face a warm, energized appearance. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that conveys a sincere and emotionally present smile. <lora:Tuning Face Detailer v2.0 Rank 256:1>

3. Sad face

Positive-Prompt (Token: 93/150)

A close-up portrait of the same [male] with a moderate expression of sadness. The inner corners of his eyebrows are raised and knitted slightly, forming a soft wrinkle at the top of his nose. His eyelids droop subtly, and the corners of his lips are turned down with visible tension in the mouth. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that communicates genuine emotional heaviness. <lora:Tuning Face Detailer v2.0 Rank 256:1>

1. A close-up portrait of the same male frozen in a moment of intense fear. His eyebrows are raised high and pulled together tightly, forming strong vertical wrinkles above the nose. His upper eyelids are stretched wide, fully exposing the whites of his eyes, and the lower lids are tensed. His mouth is open and jaw dropped, as if gasping or about to scream. The entire face is taut with alarm. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that conveys a sudden, powerful surge of panic.

4. Fear face

Positive-Prompt (Token: 117/150)

A close-up portrait of the same male with a moderately fearful expression. His eyebrows are lifted and drawn together, forming vertical lines between them. His upper eyelids are raised enough to expose more of the whites of his eyes, and the lower eyelids are slightly tensed. His lips are stretched back horizontally with a slight parting of the mouth, suggesting a controlled but noticeable startle response. All other features remain unchanged – same person, same background, same lighting – resulting in a photorealistic image that conveys alertness and unease without exaggeration. <lora:Tuning Face Detailer v2.0 Rank 256:1>

📸 Stable Diffusion parameter setting 스크린샷

Inpaint area: Only masked(mask된 영역만 SD로 수정하겠다.)

Mask blur: 4
Sampling Method: DPM++ 2M
Schedule type: Karras
Sampling step: 30
CFG Scale: 7
Denoising strength: 0.55
Batch size: 5
Batch count: 2

Env Setting:
Nvidia L4 GPU 1장
Mask blur: 4
Sampling Method: DPM++ 2M
Schedule type: Karras
Sampling step: 30
CFG Scale: 7
Denoising strength: 0.6
Batch size: 5
Batch count: 2
OpenPose → IP-Adapter
openpose_face_only
—xformers: On(0.0.23.post1) / python: 3.10.18 / torch:2.1.2+cu121

Sampling Step

SD는 기본적으로 노이즈(noise)로 가득 찬 이미지를 점진적으로 디노이징(denoising)하여 최종 이미지를 만든다.
이 denoising 과정은 여러 step에 걸쳐 이루어지는데, sampling step은 이 단계를 얼마나 많이 거칠지를 결정하는 값.

노이즈 제거 : 각 스텝마다 모델은 이미지에 남아 있는 노이즈를 조금씩 덜어낸다.

반복 횟수 : 이 과정을 여러 번 반복할수록 이미지는 더 선명해지고 디테일이 살아난다.

DPM++ 2M Karras는 적은 step으로도 좋은 결과를 내는 경향이 있다.
일반적으로 step은 20~30정도 가져가는데, 표정 같은 디테일을 변화시키는 것이 목표이므로, ~~30으로 가져감.~~

너무 오래 걸린다. → 30에서 20으로 변경(한 사람 당 15분에서 10분으로 단축.)

2. ControlNet Settings

얼굴 구조와 포즈를 원본(Neutral)과 동일하게 유지하기 위해 ControlNet 적용.

📸 ControlNet Unit0: IP-Adapter

Unit0-preprocessor: ip-adapter_clip_sdxl_plus_vith

Unit0: ip-adapter-plus-face_sdxl_vit-h

Control Weight: 0.9

Weight Type: Normal (0.9, 0.9, 0.9 →)

Control Step(Starting, Ending): 0.0~1.0

Resolution: 512

SD에 의해 원본과 다른 사람의 얼굴이 되는 것을 막아줌.

📸 ControlNet Unit1: OpenPose

Unit1-preprocessor: openpose_faceonly

Unit1: OpenPose XL2

Control Weight: 0.5

Control Step(Starting, Ending): 0.0~0.7

Resolution: 512
원본과 다른 눈 방향이나 얼굴 윤곽을 보정해줌.

📊 분석 및 시사점

Angry Face

neg-prompt가 매우매우 중요하다.
- angry를 만들 때, smile을 neg로 줄 때와 아닐 때 차이가 있다.

const SDXL_path = "assets/Expression-changed";
const expression = "ANG";
const nums  = [1]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const withNegPath = `${SDXL_path}/${expression}/interesting/neg-prompt-comp/w-neg-smile.jpeg`;
  const withoutNegPath = `${SDXL_path}/${expression}/interesting/neg-prompt-comp/wo-neg-smile.jpeg`;
 
  const withNeg  = toUrl(withNegPath);
  const withoutNeg  = toUrl(withoutNegPath);
  
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😠 ANG w/ smile-neg prompt</h3>
    <div class="image-container">
      <img src="${withNeg}" alt="Angry expression w/ neg prompt 'smile'" class="clickable-image" data-full="${withNeg}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">neg prompt "smile"</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG w/o smile-neg prompt</h3>
    <div class="image-container">
      <img src="${withoutNeg}" alt="Angry expression w/ neg prompt 'smile'" class="clickable-image" data-full="${withoutNeg}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">neg prompt "smile" x</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

📊 결과

Angry Face

Caucasian Male (CM)

const base_path = "assets/LCBL/AI-Generated";
const SDXL_path = "assets/LCBL/Expression-changed";
const expression = "ANG";
const race = "Caucasian";
const gender = "Man";
const code  = race[0]+gender[0];
const nums  = [1,2,3]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${base_path}/${race}-${gender}/${code}${n}_Neu.jpeg`;
  const angPath  = `${base_path}/${race}-${gender}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${SDXL_path}/${expression}/${code}/${code}${n}/${code}${n}-${expression}-SDXL.jpeg`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Caucasian Female (CW)

const group = "assets/LCBL/AI-Generated/Caucasian-Woman";
const code  = "CW";
const nums  = [1,2]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${group}/${code}${n}_Neu.jpeg`;
  const angPath  = `${group}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${group}/${code}-SDXL/${code}${n}_ANG_SDXL.png`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Korean Male (KM)

const group = "assets/LCBL/AI-Generated/Korean-Man";
const code  = "KM";
const nums  = [1,2]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${group}/${code}${n}_Neu.jpeg`;
  const angPath  = `${group}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${group}/${code}-SDXL/${code}${n}_ANG_SDXL.png`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Korean Female (KW)

const group = "assets/LCBL/AI-Generated/Korean-Woman";
const code  = "KW";
const nums  = [1,2]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${group}/${code}${n}_Neu.jpeg`;
  const angPath  = `${group}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${group}/${code}-SDXL/${code}${n}_ANG_SDXL.png`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Black Male (BM)

const group = "assets/LCBL/AI-Generated/Black-Man";
const code  = "BM";
const nums  = [1,2]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${group}/${code}${n}_Neu.jpeg`;
  const angPath  = `${group}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${group}/${code}-SDXL/${code}${n}_ANG_SDXL.png`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Black Female (BW)

const group = "assets/LCBL/AI-Generated/Black-Woman";
const code  = "BW";
const nums  = [1,2]; // 필요 시 변경
//const nums = Array.from({ length: 80 }, (_, i) => i + 1);
 
// 현재 노트 기준으로 링크 -> TFile -> 리소스 URL로 변환
const srcPath = dv.current().file.path;
const toUrl = (vaultPath) => {
  const file = app.metadataCache.getFirstLinkpathDest(vaultPath, srcPath);
  if (!file) {
    console.warn("[이미지 찾을 수 없음]", vaultPath);
    return "";
  }
  return app.vault.getResourcePath(file);
};
 
for (const n of nums) {
  const neuPath  = `${group}/${code}${n}_Neu.jpeg`;
  const angPath  = `${group}/${code}${n}_ANG.jpeg`;
  const sdxlPath = `${group}/${code}-SDXL/${code}${n}_ANG_SDXL.png`;
 
  const neu  = toUrl(neuPath);
  const ang  = toUrl(angPath);
  const sdxl = toUrl(sdxlPath);
 
  const html = `
<div class="image-comparison">
  <div class="image-column">
    <h3>😐 Neutral</h3>
    <div class="image-container">
      <img src="${neu}" alt="Neutral expression ${code+n}" class="clickable-image" data-full="${neu}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">기준 중립 표정</p>
  </div>
  <div class="image-column">
    <h3>😠 ANG (OpenArt)</h3>
    <div class="image-container">
      <img src="${ang}" alt="Angry expression OpenArt ${code+n}" class="clickable-image" data-full="${ang}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">OpenArt로 생성된 분노 표정</p>
  </div>
  <div class="image-column">
    <h3>😤 ANG (SDXL)</h3>
    <div class="image-container">
      <img src="${sdxl}" alt="Angry expression SDXL ${code+n}" class="clickable-image" data-full="${sdxl}">
      <div class="image-overlay"><span class="zoom-icon">🔍</span></div>
    </div>
    <p class="image-description">SDXL로 생성된 분노 표정</p>
  </div>
</div>`;
  dv.container.innerHTML += html;
}

Juhyeon's Blog

탐색기

Stable Diffusion

🛠️ Method:

이미지 생성 프로세스

어떠한 사진들을 만들 것인가?

사용 도구

어떻게 만들 것인가?

🔧 Procedure:

1. Stable Diffusion Settings

Prompt

1. Angry face

2. Happy face

3. Sad face

4. Fear face

2. ControlNet Settings

📊 분석 및 시사점

Angry Face

📊 결과

Angry Face

Caucasian Male (CM)

Caucasian Female (CW)

Korean Male (KM)

Korean Female (KW)

Black Male (BM)

Black Female (BW)

그래프 뷰

목차

Properties

백링크