Stable Diffusion 모델에서 좋은 Negative prompts를 사용하면 이상한 이미지가 생성되는 것을 막을 수 있다.
요즘 모델들은 negative prompt 없이도 좋은 이미지를 생성하지만 더욱 더 정교한 이미지를 위해 negative prompt가 필요한 경우가 있다.
Stable Diffusion은 text-to-image 모델이며 텍스트를 기반으로 이미지를 생성하는 모델이다.
Stable Diffusion을 사용하여 이미지를 만들려면 Negative prompts를 포함한 특정 설정을 이해해야 한다. 이 글에서는 Stable Diffusion에서 상황 별로 추천하는 Negative prompts와 지정된 Prompts 설정이 이미지에 어떤 영향을 미치는 지 예시로 보여준다.
해당 Negative prompts를 그대로 사용해도 좋지만 출력하고자 하는 이미지에 따라 조절할 필요가 있다.
♻️ 보편적인 이미지를 위한 Negative Prompts
Poor anatomy, Unclear, Cut-off, Distorted, Copy, Mistake, Additional arms, Extra legs, Unpleasant proportions, Long neck, Low-grade, Low resolution, Lack of arms legs, Unhealthy, Genetic variation, Beyond the frame, Inserted text, Unattractive, Lowest quality
⚡ 고품질 이미지를 위한 Negative Prompts
Signature, Poor body structure, Low-quality drawing, Incorrect size, Outside the edges, Unclear, Dull background, Logo, Cropped, Trimmed, Body parts separated, Uneven size, Twisted, Copy, Duplicated elements, Additional arms, fingers, hands, legs, Additional body parts, Flaw, Imperfection, Joined fingers, Unpleasant size, Identifying sign, Incorrect structure, Wrong proportion, Tacky, Poor quality, Poor clarity, Spot, Absent arms, fingers, hands, legs, Error, Damaged, Beyond the image, Badly drawn face, feet, hands, Text on paper, Repulsive, Unpleasant size, Shortened, Narrow eyes, Visual plan, Arrangement, Cut off, Unpleasant, Blurry, Unattractive, Awkward position, Imaginary framework, Watermark
🍳 음식 및 음료 이미지을 위한 Negative Prompts
Uneven, Poorly, Prepared, Dark, Plus, Light, Fuzzy, Charred, Comic, Animated, Chilly, Dim, Twisted, Spoiled, Sandy, People, Pieces, Unfinished, Small, Clarity, quality, Rotten, Altered, Overdone, Overcolored, Spoiled, Artwork, Drawing, Wet, Tart, Old, Raw, Blurry, Imaginary, Inverted
💡 추상적인 그림을 위한 Negative Prompts
Body structure-wise, Accurate, Creatures, Structures, Animated, Urban view, Solid, 3D, Representational, Person, Scenery, Exact, Blurry, Low-quality, Individuals, Realistic, Portrait, Render, Realistic, Landscape, Crisp lines, Drawing, Balanced, Words, Characters
🐯 야생동물 이미지를 위한 Negative Prompts
Creature, Expression, Unbalanced, Replicated, Additional, Fiction, Unclear, Textured, Individual, Expression, Unfinished, Poor quality, Low resolution, Lacking components, Intense, Too bright, Edited, Scary, Outline, Drawing, Unreal, Twisted, Dim, Blurry, Virtual engine
📺 가구나 가전제품 이미지를 위한 Negative Prompts
Missing, Included, Parts, Layout, Damaged, Exaggerated, Duplicated, Fallen, Backdrop, Viewpoint, Additional, Faulty, Form, Reversed, Incorrect, Ratio, Unfinished, Wrong, Shape, Reversed, Blurry, Crooked, Deformed, Altered, Off-balance, Overly vibrant, Over-brightened, Too large, Badly depicted, Under-brightened
🐶 사실적인 애완동물 이미지를 위한 Negative Prompts
Extra limbs, Additional fingers, Extra body parts, Changed limbs, Pet with a missing limb, Uneven ears, Bad body structure, Bad eyes, face, Facial hair, Injured finger, Injured hand, Injured leg, Injured wrist, Animated drawing, Immature, Cloned facial features, Collapsed eye makeup, Joined limbs, Connected, Dead body, Pet with physical challenges, Cut-off head, Sad, Dried out, Two-faced, Duplicated characteristics, Creepy, Extended neck
🏙️ 도시 이미지를 위한 Negative Prompts
Uneven structures, Fuzzy, Vehicles, Distorted buildings, Gritty, Image distortions, Dim contrast, Poor quality, Low resolution, Close-up shots, Overly bright, Excessively colorful, Humans (if included), Animals (if included), Simple backdrop, Spooky, Unrealistic, Imaginary structures, Fantasy sky, Unusual colors
🌳 자연이나 경치 이미지를 위한 Negative Prompts
Blurred, Dull, Zoomed in, Dim, Lack of clarity, Distorted, Spooky, Hazy, Black and white, Uniform, Diminished contrast, Poor quality, Low resolution, Low-detail, Extreme close-up, Too bright, Excessive color, Basic, Featureless backdrop, Close-up of a person, Ordinary, Conventional, Lacking originality, Too dark
최고의 결과를 얻기 위한 Prompts 개선 방법
- 간결한 방법으로 문제를 식별
- 개방형 질문
- 편파적인 언어를 사용 X
- 상황에 대한 맥락을 제시
참고 자료 : https://mockey.ai/blog/stable-diffusion-negative-prompt/
'AI Research' 카테고리의 다른 글
[Paper Review] AN IMAGE IS WORTH 16X16 WORDS:TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE (0) | 2025.01.24 |
---|---|
[Paper Review] Attention Is All You Need (0) | 2025.01.13 |
[Code] VAE(Variational Auto-Encoder) 구현 (0) | 2024.03.28 |
Diffusion Model vs GANs (2) | 2023.12.28 |