koriweb
d8a80f6272
이름만 다른(표기 변형) [[위키링크]]를 대상 문서의 canonical 제목으로 치환해 끊겼던 1,200개 링크를 연결. 제목/파일명 정규화 일치만 적용하고 별칭 매칭은 과병합 위험으로 제외(애매성 가드). 원본은 _link_reconcile_backup/ 에 백업. 도구: Datacollect/scripts/link_reconcile_apply.mjs Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
178 lines
6.7 KiB
Markdown
178 lines
6.7 KiB
Markdown
---
|
|
id: wiki-2026-0508-image-prompt-작성-방법
|
|
title: Image Prompt 작성 방법
|
|
category: 10_Wiki/Topics
|
|
status: verified
|
|
canonical_id: self
|
|
aliases: [Image Prompt Engineering, T2I Prompting, Diffusion Prompting]
|
|
duplicate_of: none
|
|
source_trust_level: A
|
|
confidence_score: 0.9
|
|
verification_status: applied
|
|
tags: [image-generation, prompting, diffusion, midjourney, flux]
|
|
raw_sources: []
|
|
last_reinforced: 2026-05-10
|
|
github_commit: pending
|
|
tech_stack:
|
|
language: text
|
|
framework: FLUX/Midjourney/SDXL
|
|
---
|
|
|
|
# Image Prompt 작성 방법
|
|
|
|
## 매 한 줄
|
|
> **"매 image prompt 의 핵심 = 'Subject + Style + Composition + Quality modifier' 의 명시적 ordering."**. 2022 SD 1.5 era 의 keyword soup 에서 출발 → 2026 FLUX.1.1 Pro / Midjourney v7 / Imagen 4 era 는 natural-language 의 long-form description 의 superior 을 increasingly 한다. 매 prompt structure 의 deterministic 한 control 의 사용자 의 intent 의 model 의 latent space 의 mapping 을 direct.
|
|
|
|
## 매 핵심
|
|
|
|
### 매 Prompt anatomy (4-part structure)
|
|
- **Subject**: "a red fox", "an elderly Japanese ceramicist" — concrete noun + descriptors.
|
|
- **Style**: "oil painting", "cinematic photography, 35mm film", "Studio Ghibli" — medium / artist / movement.
|
|
- **Composition**: "low-angle shot", "rule of thirds", "shallow depth of field f/1.4" — framing 의 explicit specification.
|
|
- **Quality / technical**: "8k, sharp focus, golden hour lighting" — modifiers 의 final boost.
|
|
|
|
### 매 Token weight (Midjourney/SDXL)
|
|
- `(keyword:1.3)` — 매 emphasis 의 multiplier.
|
|
- `[keyword]` — 매 de-emphasis (SD WebUI).
|
|
- Midjourney `--stylize 250` (default 100) — 매 artistic license 의 control.
|
|
- Midjourney `::` — 매 multi-prompt weight: `red fox::2 forest::0.5`.
|
|
|
|
### 매 Negative prompt
|
|
1. SDXL/SD3: `negative_prompt="blurry, low quality, deformed hands, watermark"`.
|
|
2. Midjourney: `--no text logo` — exclusion list.
|
|
3. FLUX: 매 native negative prompt support 의 X — instead 매 positive 의 rephrase.
|
|
|
|
## 💻 패턴
|
|
|
|
### Pattern 1: FLUX.1.1 Pro 의 long-form natural prompt
|
|
```text
|
|
A photorealistic portrait of an elderly Japanese ceramicist, weathered hands
|
|
shaping a clay tea bowl on a wooden wheel. Soft window light from the left,
|
|
shallow depth of field. Earthy tones, traditional workshop with shelves of
|
|
glazed pottery in the background. Shot on Hasselblad H6D, 80mm lens, f/2.8.
|
|
```
|
|
|
|
### Pattern 2: Midjourney v7 structured prompt
|
|
```text
|
|
/imagine prompt: cyberpunk samurai standing in neon-lit Tokyo alleyway,
|
|
holographic katana, rain-soaked street reflections, cinematic composition,
|
|
volumetric lighting, by Syd Mead and Moebius --ar 16:9 --stylize 350 --v 7
|
|
```
|
|
|
|
### Pattern 3: SDXL 의 weighted prompt + negative
|
|
```python
|
|
from diffusers import StableDiffusionXLPipeline
|
|
import torch
|
|
|
|
pipe = StableDiffusionXLPipeline.from_pretrained(
|
|
"stabilityai/stable-diffusion-xl-base-1.0",
|
|
torch_dtype=torch.float16,
|
|
).to("cuda")
|
|
|
|
prompt = "(masterpiece:1.2), red fox in autumn forest, golden hour, " \
|
|
"(detailed fur:1.3), cinematic photography, 8k"
|
|
negative = "blurry, low quality, deformed, watermark, text"
|
|
|
|
image = pipe(
|
|
prompt=prompt,
|
|
negative_prompt=negative,
|
|
num_inference_steps=30,
|
|
guidance_scale=7.5,
|
|
).images[0]
|
|
```
|
|
|
|
### Pattern 4: ControlNet 의 pose + prompt
|
|
```python
|
|
from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel
|
|
from controlnet_aux import OpenposeDetector
|
|
|
|
openpose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
|
|
pose_image = openpose(input_image)
|
|
|
|
controlnet = ControlNetModel.from_pretrained("xinsir/controlnet-openpose-sdxl-1.0")
|
|
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
|
|
"stabilityai/stable-diffusion-xl-base-1.0",
|
|
controlnet=controlnet,
|
|
)
|
|
|
|
result = pipe(
|
|
prompt="a knight in shining armor, dramatic lighting",
|
|
image=pose_image,
|
|
controlnet_conditioning_scale=0.8,
|
|
).images[0]
|
|
```
|
|
|
|
### Pattern 5: LLM 의 prompt expansion
|
|
```python
|
|
import anthropic
|
|
|
|
client = anthropic.Anthropic()
|
|
|
|
def expand_prompt(short_idea: str) -> str:
|
|
msg = client.messages.create(
|
|
model="claude-opus-4-7",
|
|
max_tokens=300,
|
|
messages=[{
|
|
"role": "user",
|
|
"content": f"Expand this image idea into a detailed FLUX prompt "
|
|
f"(subject, style, composition, lighting, camera): {short_idea}"
|
|
}],
|
|
)
|
|
return msg.content[0].text
|
|
|
|
prompt = expand_prompt("lonely lighthouse at dusk")
|
|
```
|
|
|
|
### Pattern 6: Imagen 4 의 text rendering
|
|
```text
|
|
A vintage diner sign at night, neon letters reading "MOON CAFÉ — OPEN 24 HOURS",
|
|
warm pink and cyan glow, rain reflections on asphalt, 1950s Americana aesthetic,
|
|
cinematic wide shot
|
|
```
|
|
(Imagen 4 / FLUX 의 accurate text rendering 의 strong — 매 quote-wrapped text 의 use.)
|
|
|
|
### Pattern 7: Aspect ratio + seed control
|
|
```text
|
|
/imagine prompt: misty mountain temple at sunrise --ar 21:9 --seed 42 --v 7
|
|
```
|
|
- `--ar 21:9` — cinematic widescreen.
|
|
- `--seed 42` — 매 reproducibility.
|
|
|
|
## 매 결정 기준
|
|
| 상황 | Approach |
|
|
|---|---|
|
|
| Photorealistic portrait | FLUX.1.1 Pro + long natural prompt |
|
|
| Stylized art / illustration | Midjourney v7 + artist references |
|
|
| Precise pose / composition | SDXL + ControlNet (openpose/canny/depth) |
|
|
| Text in image | Imagen 4 / FLUX (quote-wrapped) |
|
|
| Local / private | SDXL or FLUX.1-dev on RTX 4090 |
|
|
| Batch / API workflow | Replicate / fal.ai / Together AI |
|
|
|
|
**기본값**: FLUX.1.1 Pro (photoreal) or Midjourney v7 (stylized) + 4-part structured prompt.
|
|
|
|
## 🔗 Graph
|
|
- 부모: [[Generative-AI]] · [[Diffusion-Models]]
|
|
- 변형: [[Vary Region (인페인팅)]] · [[ControlNet]] · [[LoRA-Fine-Tuning]]
|
|
- Adjacent: [[Prompt_Engineering|Prompt-Engineering]]
|
|
|
|
## 🤖 LLM 활용
|
|
**언제**: prompt expansion / variation generation / style transfer description / negative prompt brainstorming.
|
|
**언제 X**: 매 final image quality 의 judgment 의 X — 매 human-in-the-loop 의 visual evaluation 의 always 필요.
|
|
|
|
## ❌ 안티패턴
|
|
- **Keyword soup**: "8k, masterpiece, best quality, ultra detailed, ..." — 매 modern model 의 dilution. Natural language 의 prefer.
|
|
- **Contradictory modifiers**: "photorealistic anime" — 매 latent space 의 confusion.
|
|
- **Negative prompt overload**: 매 30+ negative tokens 의 generation 의 degrade.
|
|
- **No seed control**: 매 reproducibility 의 X — 매 iteration / A/B 의 impossible.
|
|
- **Ignoring aspect ratio**: 매 default square 의 composition 의 mismatch (cinematic shot 의 16:9 의 필요).
|
|
|
|
## 🧪 검증 / 중복
|
|
- Verified: Midjourney v7 docs (2026), Black Forest Labs FLUX docs, Stability AI SDXL paper.
|
|
- 신뢰도 A.
|
|
|
|
## 🕓 Changelog
|
|
| 날짜 | 변경 |
|
|
|---|---|
|
|
| 2026-05-08 | Phase 1 |
|
|
| 2026-05-10 | Manual cleanup — full image prompt engineering guide |
|