2nd/10_Wiki/Topic_Programming/AI_and_ML/사후 편집 (Post-editing).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-사후-편집-post-editing	사후 편집 (Post-editing)	10_Wiki/Topics	verified	self	Post-editing AI Image Post-editing Image Refinement	none	A	0.9	applied	image-generation post-editing midjourney flux photoshop comfyui		2026-05-10	pending	language framework python ComfyUI/Photoshop-Generative-Fill

사후 편집 (Post-editing)

매 한 줄

"매 generation 은 draft, 매 final 은 post-edit 의 결과". 매 2026 production pipeline 에서 raw text-to-image output 의 직접 ship 의 X — 매 inpaint, upscale, color-grade, retouching 의 multi-stage refinement 가 standard. Midjourney/FLUX/Imagen 4 의 base generation + ComfyUI region-edit + Photoshop Generative Fill 의 hybrid workflow 가 매 commercial baseline.

매 핵심

매 post-editing 의 정의

• 매 generated image 의 결함 의 fix + intent 의 align.
• 매 stage: ① local fix (face/hand/text), ② global polish (color, contrast), ③ composition (crop, insert), ④ upscale (output res).
• 매 zero-edit ship 의 X — 매 99% 의 commercial output 이 적어도 1 단계 의 post-edit 의 거침.

매 도구 stack (2026)

• Midjourney V8 Editor: 매 inpaint + extend (uncrop) + retexture 의 in-platform.
• Photoshop 2026 Generative Fill (FLUX-2 backed): 매 industry default — layer 호환.
• ComfyUI + FLUX.1 Fill / SDXL Inpaint: 매 open-source pipeline. 매 reproducible.
• Magnific / Krea Upscale: 매 1024 → 4K 의 detail-add upscaling.
• Topaz Photo AI: 매 noise/blur 의 cleanup.
• Adobe Firefly 4: 매 commercial-safe (training-data 의 license 명확).

매 typical 결함 카테고리

1. 해부학적 오류: hand (6 fingers), feet, eye 의 asymmetry.
2. Text 의 garbled: logo, sign, caption 의 letters 의 corruption.
3. Composition 의 mismatch: edge 의 cut, perspective 의 break.
4. Style drift: face 의 character 의 inconsistency (multi-shot).
5. Lighting 의 implausible: shadow direction 의 conflict.

매 응용

1. Marketing / e-commerce visual production.
2. Concept art / pre-viz.
3. Editorial illustration.
4. Game asset (texture, character).
5. Architectural rendering의 humanization.

💻 패턴

Pattern 1 — Inpaint (ComfyUI + FLUX.1 Fill)

# ComfyUI API workflow snippet
import json, requests, base64
from PIL import Image

def inpaint_region(image_path, mask_path, prompt):
    workflow = json.load(open("flux_fill.json"))
    workflow["3"]["inputs"]["image"] = base64.b64encode(open(image_path, "rb").read()).decode()
    workflow["4"]["inputs"]["mask"]  = base64.b64encode(open(mask_path,  "rb").read()).decode()
    workflow["6"]["inputs"]["text"]  = prompt
    workflow["7"]["inputs"]["steps"] = 28
    workflow["7"]["inputs"]["cfg"]   = 3.5
    r = requests.post("http://127.0.0.1:8188/prompt", json={"prompt": workflow})
    return r.json()["prompt_id"]

# Fix 6-finger hand
inpaint_region("draft.png", "hand_mask.png",
               "anatomically correct human hand, 5 fingers, natural pose")

Pattern 2 — Hand-fix automated detection

import cv2, mediapipe as mp
mp_hands = mp.solutions.hands.Hands(static_image_mode=True, max_num_hands=4)

def detect_bad_hands(img_path):
    img = cv2.imread(img_path)
    res = mp_hands.process(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
    bad = []
    if res.multi_hand_landmarks:
        for lm in res.multi_hand_landmarks:
            # Heuristic: finger length ratio, joint angles
            if not is_anatomically_valid(lm):
                bad.append(bbox_from_landmarks(lm, img.shape))
    return bad  # → mask + inpaint

Pattern 3 — Face consistency (IP-Adapter FaceID)

from diffusers import FluxFillPipeline
import torch

pipe = FluxFillPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Fill-dev",
    torch_dtype=torch.bfloat16
).to("cuda")

# Reference face → swap into draft
from ip_adapter import IPAdapterFaceID
adapter = IPAdapterFaceID(pipe, "ip-adapter-faceid-flux.bin")

result = adapter.generate(
    image=draft, mask=face_mask,
    face_image=reference_face,
    prompt="same person, professional headshot",
    num_inference_steps=30, guidance_scale=4.0,
)

Pattern 4 — Upscale + detail (Magnific-style)

# SUPIR / FLUX-Upscale style
from diffusers import StableDiffusionUpscalePipeline
upscaler = StableDiffusionUpscalePipeline.from_pretrained(
    "stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16
).to("cuda")

hi_res = upscaler(
    prompt="ultra-detailed photograph, sharp focus, 8k",
    image=Image.open("draft_1024.png"),
    num_inference_steps=20, guidance_scale=7,
).images[0]
hi_res.save("final_4k.png")

Pattern 5 — Color grading (LUT in Pillow)

from PIL import Image
import numpy as np

def apply_lut(img: Image.Image, lut_path: str) -> Image.Image:
    lut = np.load(lut_path)  # shape (33,33,33,3)
    arr = np.asarray(img).astype(np.float32) / 255.0
    idx = (arr * 32).astype(np.int32)
    out = lut[idx[..., 0], idx[..., 1], idx[..., 2]]
    return Image.fromarray((out * 255).astype(np.uint8))

# Apply teal-orange cinematic LUT
final = apply_lut(Image.open("graded_input.png"), "teal_orange.npy")

Pattern 6 — Photoshop scripting (Generative Fill)

// Photoshop 2026 .jsx — Generative Fill via ExtendScript
var doc = app.activeDocument;
doc.selection.select([[120,80],[420,80],[420,380],[120,380]]);
var generativeFill = stringIDToTypeID("generativeFill");
var desc = new ActionDescriptor();
desc.putString(stringIDToTypeID("prompt"), "remove power lines, clean sky");
executeAction(generativeFill, desc, DialogModes.NO);
doc.saveAs(new File("/out/cleaned.psd"));

Pattern 7 — Batch QA loop

def post_edit_pipeline(draft_path):
    img = load(draft_path)
    if has_bad_hands(img):
        img = inpaint_hands(img)
    if has_garbled_text(img):
        img = inpaint_text(img, target_text="ACME Corp")
    img = color_grade(img, lut="film_emulation.npy")
    img = upscale_4x(img)
    return img

매 결정 기준

상황	Approach
빠른 fix, 1장	Photoshop Generative Fill
Reproducible, batch	ComfyUI workflow JSON
Face/character lock	IP-Adapter FaceID + inpaint
Detail-add upscale	Magnific / SUPIR
Commercial license worry	Adobe Firefly 4

기본값: 매 ComfyUI + FLUX.1 Fill 의 reproducible base, 매 final touch 만 Photoshop.

🔗 Graph

• 부모: AI 이미지 생성 (AI Image Generation) · AI 이미지 생성 및 편집 워크플로우 (AI Image Generation & Editing Workflow)
• 변형: Inpainting · Outpainting
• 응용: 상업용 브랜드 이미지 및 디자인 시스템 구축
• Adjacent: ControlNet · IP-Adapter

🤖 LLM 활용

언제: 매 prompt 의 generation, 매 mask 의 description, 매 QA 의 결함 카테고리화. 언제 X: 매 final pixel-level decision (designer 의 eye 가 필요).

❌ 안티패턴

• Re-roll forever: 매 100 generations 의 spam 보다 매 1 inpaint 가 빠름.
• Single-pass ship: 매 raw text-to-image 의 commercial use 의 X.
• Mask 의 too-tight: 매 boundary artifact. 매 feather 8-16px 의 default.
• Upscale before fix: 매 결함 의 amplification. 매 fix → upscale 의 순서.

🧪 검증 / 중복

• Verified (FLUX.1 Fill release notes 2025-11; Adobe Firefly 4 docs; ComfyUI manager wiki).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — multi-stage post-edit pipeline + 7 patterns