2nd/10_Wiki/Topic_Programming/AI_and_ML/Positive_Prompt.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-positive-prompt	Positive Prompt	10_Wiki/Topics	verified	self	Positive Prompts Prompt Prompt Description	none	A	0.9	applied	prompt-engineering image-generation stable-diffusion midjourney flux		2026-05-10	pending	language framework python diffusers, comfyui

Positive Prompt

매 한 줄

"매 image generation에서 desired content 를 describe — subject, style, composition, quality.". Stable Diffusion / FLUX / Midjourney 핵심 input. Negative prompt와 짝을 이루며, 2024-2025 modern model (FLUX.1, SD3, MJ v7)에서 매 natural language description이 weighted token보다 우세.

매 핵심

매 구성 요소

• Subject: "a woman, a robot, a cathedral".
• Action / pose: "running through forest", "sitting at desk".
• Style: "oil painting", "cyberpunk", "studio Ghibli".
• Composition: "wide angle", "close-up", "rule of thirds".
• Lighting: "golden hour", "rim light", "volumetric".
• Quality modifier: "highly detailed", "8k" (older models — modern은 less needed).
• Artist / reference: "in the style of Greg Rutkowski" (controversial).

매 model별 syntax

• SD 1.5 / SDXL: (token:1.3) weighted, BREAK 분리, comma list.
• FLUX.1 / SD3: 매 natural language paragraph가 best — token weighting less effective.
• Midjourney v7: --ar 16:9 --stylize 200 --chaos 20 flag, natural prompt.
• DALL-E 3 / GPT-Image: 매 conversational, descriptive paragraph.

매 modern best practice (2025)

• Natural language sentence > comma keyword stuffing.
• 매 subject specific, then style, then technical.
• Reference image (img2img, IPAdapter, FLUX Redux) 매 단어보다 강력.
• LoRA / fine-tune이 style token 대체.

매 응용

1. Concept art, illustration.
2. Marketing asset gen.
3. Product mockup, fashion.
4. Storyboard, film pre-vis.
5. Game asset (texture, character sheet).

💻 패턴

Diffusers SDXL (weighted)

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), portrait of a samurai warrior, "
          "intricate armor, cherry blossoms, golden hour, "
          "cinematic lighting, depth of field")
neg = "low quality, blurry, deformed hands, extra fingers"

img = pipe(prompt, negative_prompt=neg, num_inference_steps=30,
           guidance_scale=7.0).images[0]

FLUX.1 (natural language)

from diffusers import FluxPipeline
import torch
pipe = FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-dev',
                                     torch_dtype=torch.bfloat16).to('cuda')

prompt = ("A wide cinematic shot of a samurai standing under cherry "
          "blossoms at golden hour. He wears intricate red and black "
          "armor. Soft volumetric light filters through petals. "
          "Shallow depth of field with the warrior in sharp focus.")

img = pipe(prompt, guidance_scale=3.5, num_inference_steps=28,
           max_sequence_length=512).images[0]

Compel (advanced weighting, SD)

from compel import Compel
compel = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder)
embeds = compel("a cat++ playing piano in a (jazz bar)1.3")
img = pipe(prompt_embeds=embeds).images[0]

Midjourney v7 prompt format

/imagine prompt: a samurai under cherry blossoms, golden hour,
volumetric light, cinematic --ar 21:9 --stylize 300 --v 7

Modular template (programmatic)

def build_prompt(subject, style, light, mood):
    return (f"{subject}, {style} style, {light} lighting, "
            f"{mood} mood, highly detailed composition")

p = build_prompt("a lone astronaut on Mars",
                 "concept art", "soft sunset", "melancholic")

LoRA-augmented (style token)

pipe.load_lora_weights('artist_style.safetensors')
prompt = "<lora:artist_style:0.8> portrait of woman, watercolor"

매 결정 기준

상황	Approach
FLUX / SD3 / DALL-E 3	Natural paragraph, descriptive
SDXL / SD 1.5	Comma-separated, weighted tokens
Midjourney	Natural + flags (--ar, --stylize)
Specific style reproduction	LoRA + 짧은 prompt
Reference matching	img2img / IPAdapter > prompt
Batch programmatic	Template + parameter slot

기본값: modern model은 natural sentence, legacy SD는 weighted comma list.

🔗 Graph

• 부모: Prompt_Engineering · Diffusion_Models
• 변형: Negative_Prompt
• 응용: Stable_Diffusion · FLUX · Midjourney · DALL-E
• Adjacent: LoRA · IPAdapter · ControlNet · ComfyUI

🤖 LLM 활용

언제: image gen API wrapper, batch asset generation, prompt template system, A/B test variation. 언제 X: 매 reference image가 있으면 img2img / IPAdapter — 매 prompt만으론 매 정확 못 reproduce.

❌ 안티패턴

• Keyword spam: "8k, hyperdetailed, ultra hd, masterpiece, best quality, ..." — 매 modern model에 무의미.
• Contradictory style mix: "anime, photorealistic, oil painting" — 매 confused output.
• Overweight (token:2.0): 매 artifact, oversaturation.
• Artist names without consent: 매 ethical issue + many platforms ban.
• Same prompt for all models: 매 model별 syntax 다름 — port 필요.

🧪 검증 / 중복

• Verified (FLUX.1 model card, SDXL paper, Midjourney v7 docs, diffusers library docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — positive prompt structure + model-specific syntax (FLUX, SDXL, MJ v7)