2nd/10_Wiki/Topics/AI_and_ML/AI 이미지 생성 및 편집 워크플로우 (AI Image Generation & Editing Workflow).md

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

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	inferred_by
wiki-2026-0508-ai-이미지-생성-및-편집-워크플로우-ai-image-ge	AI Image Generation & Editing Workflow	10_Wiki/Topics	verified	self	AI 이미지 워크플로우 image generation workflow draft mode iterative refinement	none	B	0.85	conceptual	image-generation workflow midjourney stable-diffusion draft-mode iterative post-editing		2026-05-09	pending	Claude Opus 4.7 (manual cleanup 2026-05-09)

AI Image Generation & Editing Workflow

📌 한 줄 통찰

Single perfect prompt 의 myth → iterative loop. Draft (cheap variant) → select → refine → upscale → polish (post-edit). 매 round 의 quality ↑.

📖 핵심

매 5-stage workflow

Stage 1: Concept + ideation

• 매 reference (Pinterest, ArtStation).
• 매 mood board.
• 매 prompt sketch.

Stage 2: Draft generation (cheap)

• 매 dozen variant.
• Midjourney --draft mode (10x speed).
• Flux Schnell (fast).
• 매 4-8 candidate.

Stage 3: Select + iterate

• 매 best candidate.
• 매 prompt 의 refine.
• 매 next round.

Stage 4: Refine (full quality)

• 매 selected 의 high-quality regenerate.
• 매 final aspect ratio.

Stage 5: Post-edit

• 매 inpaint (specific fix).
• 매 outpaint (extend).
• 매 upscale (resolution).
• 매 retouch (Photoshop).

→ 매 stage 의 different speed / cost.

매 cost saving

Draft mode

• Midjourney V7 --draft: 10x faster, ~50% GPU cost.
• Flux Schnell: 4-step (vs 50).
• Latent Consistency Models (LCM).

→ 매 idea 의 cheap exploration.

Generation 의 cost

• Midjourney: $10-60 / month subscription.
• DALL-E 3: ~$0.08 / image.
• Stable Diffusion (self-host): GPU 운영 cost.
• Flux Pro (Replicate): $0.05 / image.

Compute optimization

• 매 quality preset (4 step LCM, 20 step DPM++, 50 step DDIM).
• 매 resolution (512 → 1024 → 4K).
• 매 batch size.

매 reference 의 활용

Style reference (sref)

• Midjourney --sref [URL].
• 매 brand 의 mood board.
• 매 campaign 의 visual cohesion.

Character reference (cref)

• Midjourney --cref [URL].
• 매 character 의 consistency.

Omni reference (oref)

• Midjourney V7+.
• 매 specific object identity.

IP-Adapter (Stable Diffusion)

• 매 reference image 의 style + structure.

LoRA

• 매 specific style / character 의 fine-tune.

매 quality control

Negative prompt (Stable Diffusion)

• 매 known defect 의 explicit.
• "ugly, deformed, watermark, low quality, blurry, extra fingers".

Specific defect 의 inpaint

• 매 detected defect 의 mask.
• 매 targeted prompt.

Upscale + face restore

• Real-ESRGAN (background).
• GFPGAN / CodeFormer (face).

Production workflow example

Marketing campaign

1. Mood board (brand 의 reference).
2. Draft 30 variants (Midjourney draft).
3. Select 5 (different angle / composition).
4. Full HD generate.
5. Inpaint defects.
6. Upscale 4K.
7. Photoshop final touch.

→ 30+ image / hour.

Product mockup

1. Real product photo (input).
2. Img2Img (style transfer).
3. Background outpaint (lifestyle context).
4. Inpaint shadow / reflection.
5. Upscale.

Concept art (game)

1. Quick sketch (artist).
2. ControlNet 의 line art.
3. Generate variations.
4. Select + paint over (Photoshop).

💻 Code

Iterative loop (Diffusers)

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained("model")

# Stage 1: Draft (low quality, fast)
prompts = [base_prompt + variation for variation in style_variations]
drafts = pipe(prompts, num_inference_steps=10, guidance_scale=5).images

# Stage 2: Select (manual or ML score)
best_idx = select_best(drafts)
best_prompt = prompts[best_idx]

# Stage 3: Full quality
final = pipe(best_prompt, num_inference_steps=50, guidance_scale=7.5).images[0]

# Stage 4: Post-edit (inpaint specific defect)
mask = detect_face_defect(final)
inpaint_pipe = StableDiffusionXLInpaintPipeline.from_pretrained("inpaint")
fixed = inpaint_pipe(prompt="perfect face", image=final, mask_image=mask).images[0]

# Stage 5: Upscale
from realesrgan import RealESRGANer
upscaler = RealESRGANer(scale=4, ...)
upscaled, _ = upscaler.enhance(np.array(fixed))

Batch + cost-aware

def smart_generate(prompt, target_quality='final'):
    if target_quality == 'draft':
        return pipe(prompt, num_inference_steps=10).images[0]
    elif target_quality == 'preview':
        return pipe(prompt, num_inference_steps=25).images[0]
    elif target_quality == 'final':
        img = pipe(prompt, num_inference_steps=50).images[0]
        return upscale(img)

Reference-driven (Flux + IP-Adapter)

from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev")
pipe.load_ip_adapter("flux-ip-adapter")

style_ref = Image.open("brand_mood.jpg")
result = pipe(
    prompt="product on table, professional photo",
    ip_adapter_image=style_ref,
    ip_adapter_scale=0.6,
).images[0]

🤔 결정 기준

Stage	추천
Ideation	Free + reference
Draft	Midjourney draft / Flux Schnell
Refine	Full quality
Post-edit	Inpaint + upscale
Production	Photoshop final

기본값: Draft 30 → Select 5 → Final + post-edit. 매 cost 의 80% saving + quality 의 maintain.

🔗 Graph

• 부모: AI Image Generation
• 변형: Iterative-Refinement · Draft-Mode · AI 모델 사후 편집 도구 (Post-editing Tools)
• Adjacent: Style Reference · ControlNet

🤖 LLM 활용

언제: 매 commercial creative project. 매 brand campaign. 언제 X: 매 single-shot idea (no iteration). 매 highly specific artist style (legal).

❌ 안티패턴

• Single prompt + accept: low quality.
• Full quality from start: cost 폭발.
• No reference: brand inconsistency.
• No post-edit: defect in production.
• Upscale 의 detail invent: hallucinated artifact.

🧪 검증 / 중복

• Verified.
• 신뢰도 B.
• Overlap with AI Image Generation / AI 모델 사후 편집 도구 (Post-editing Tools).

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-09	Manual cleanup — 5-stage workflow + cost + reference + code