2nd/10_Wiki/Topics/AI_and_ML/일관된 캐릭터 및 스타일 구축.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-일관된-캐릭터-및-스타일-구축	일관된 캐릭터 및 스타일 구축	10_Wiki/Topics	verified	self	Consistent Character Brand Consistency Maintenance Character Sheet Style Lock	none	A	0.9	applied	ai image-generation character-consistency lora style		2026-05-10	pending	language framework python diffusers-flux

일관된 캐릭터 및 스타일 구축

매 한 줄

"매 character/style consistency 는 single shot 의 prompt 로 안 되며, multi-stack (LoRA + IP-Adapter + ControlNet + reference latents) 의 합으로만 stable 해진다". 2026 의 production character pipeline 은 character sheet → multi-view dataset → subject LoRA → generation-time stack → CLIP/face-similarity validation 의 매 5-step loop 로 운영됨. seed lock 만으로는 매 부족.

매 핵심

매 consistency 의 4 차원

• Identity: 얼굴, 체형, 비율 (face/body).
• Outfit: 의상 details, color, accessory.
• Style: rendering, palette, line/shading.
• Pose/Expression: 매 controllable variation.

매 stack (2026 best)

• Subject LoRA: 30-50 ref images, identity lock.
• Style LoRA: separate, 매 stack 가능.
• IP-Adapter Face / FaceID: face embedding.
• PuLID / Photomaker: zero-shot face injection.
• InstantID: identity + pose ControlNet.
• Reference-only ControlNet: latent reference.

매 응용

1. Webtoon / illustrated novel 의 character series.
2. Brand mascot 의 cross-channel reuse.
3. Game NPC 의 procedural variation with identity.

💻 패턴

Character sheet (training data)

data/hero/
├─ 01_front_neutral.png   "<hero> front view, neutral expression"
├─ 02_side_neutral.png    "<hero> side profile, neutral"
├─ 03_back.png            "<hero> back view"
├─ 04_3q_smile.png        "<hero> 3/4 view, smiling"
├─ 05_close_face.png      "<hero> close-up portrait"
├─ 06_full_body.png       "<hero> full body, T-pose"
├─ 07_action_run.png      "<hero> running pose"
...
30+ images, varied pose/expression, consistent outfit

Subject LoRA + caption

# Captions emphasize TRIGGER + variation, NOT outfit (so it's learned implicitly)
captions = [
    "<hero01>, front view, neutral expression",
    "<hero01>, side profile",
    "<hero01>, smiling, 3/4 view",
    "<hero01>, in forest, full body",
]
# Train LoRA rank 32, 2000 steps, lr 1e-4, FLUX.1-dev

Multi-LoRA at inference

from diffusers import FluxPipeline
import torch

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev",
                                    torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("./loras/hero01.safetensors", adapter_name="char")
pipe.load_lora_weights("./loras/brand_style.safetensors", adapter_name="style")
pipe.set_adapters(["char","style"], adapter_weights=[0.95, 0.7])

img = pipe(
    "<hero01>, drinking coffee in cafe, brand_style",
    num_inference_steps=28, guidance_scale=3.5,
    generator=torch.Generator("cuda").manual_seed(42)
).images[0]

PuLID (zero-shot face lock)

from pulid import PuLIDPipeline
pl = PuLIDPipeline.from_pretrained("ByteDance/PuLID-FLUX")

img = pl.generate(
    prompt="<hero> hiking on mountain, golden hour",
    id_image="hero_face_ref.png",
    id_weight=0.85,
    seed=42, steps=20
)
# No training, single ref → identity preserved

InstantID (face + pose)

from diffusers import StableDiffusionXLInstantIDPipeline, ControlNetModel
controlnet = ControlNetModel.from_pretrained("InstantX/InstantID")
pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet
).to("cuda")
pipe.load_ip_adapter_instantid("InstantX/InstantID")

face_emb = extract_face_embedding(ref_img)
img = pipe(
    prompt="<hero> samurai in feudal japan",
    image_embeds=face_emb,
    image=pose_kps_img,   # OpenPose keypoints
    controlnet_conditioning_scale=0.8,
    ip_adapter_scale=0.8,
).images[0]

Reference image guidance (IP-Adapter)

pipe.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models",
                     weight_name="ip-adapter_sdxl.bin")
pipe.set_ip_adapter_scale(0.5)

img = pipe(
    prompt="<hero> sci-fi armor, cyberpunk city",
    ip_adapter_image=hero_reference_img,
).images[0]

Validation: face similarity

from insightface.app import FaceAnalysis
import numpy as np

face = FaceAnalysis(name="buffalo_l"); face.prepare(ctx_id=0)
ref_emb = face.get(ref_img)[0].embedding
gen_emb = face.get(generated_img)[0].embedding

cos_sim = np.dot(ref_emb, gen_emb) / (np.linalg.norm(ref_emb)*np.linalg.norm(gen_emb))
assert cos_sim > 0.55, f"identity drift: {cos_sim:.3f}"

Outfit consistency check (CLIP)

import open_clip
model, _, prep = open_clip.create_model_and_transforms("ViT-bigG-14")
outfit_prompt = "white hoodie, black cargo pants, red sneakers"
txt_emb = model.encode_text(open_clip.tokenize([outfit_prompt]))
img_emb = model.encode_image(prep(generated).unsqueeze(0))
score = torch.cosine_similarity(txt_emb, img_emb)
# alert if score < 0.27

Generation-loop with retry

def generate_consistent(prompt, max_retry=4):
    for i in range(max_retry):
        seed = 1000 + i*7
        img = pipe(prompt, generator=torch.Generator("cuda").manual_seed(seed)).images[0]
        sim = face_sim(img, ref_img)
        if sim > 0.55: return img, sim, seed
    raise RuntimeError("identity could not be preserved")

Style transfer for series

# Step 1: generate composition with character locked
base = pipe("<hero> sitting on bench", lora=char_lora).images[0]

# Step 2: img2img with style LoRA
styled = i2i_pipe(
    prompt="<hero> sitting on bench, brand_style",
    image=base, strength=0.4,
    lora_stack=[char_lora, style_lora],
).images[0]

매 결정 기준

상황	Approach
30+ ref available	train subject LoRA
1-3 ref only	PuLID / Photomaker
identity + exact pose	InstantID
brand style 분리	separate Style LoRA
series of frames	seed lock + same LoRA stack
validation gate	InsightFace cos > 0.55

기본값: subject LoRA + style LoRA + IP-Adapter face + face-sim CI.

🔗 Graph

• 부모: AI Image Generation
• 변형: LoRA Fine-tuning · InstantID
• 응용: 인공지능 시각 언어 생성 (AI Visual Language Generation) · 오픈소스 이미지 모델 미세 조정 및 배포
• Adjacent: IP-Adapter · ControlNet

🤖 LLM 활용

언제: caption authoring for char dataset, prompt variation list, validation rubric. 언제 X: face similarity scoring — deterministic insightface 가 정답.

❌ 안티패턴

• Single ref overfit: 1 image LoRA → mode collapse.
• Mixing identities in dataset: 매 LoRA confused.
• Caption with outfit details: outfit 이 trigger 와 분리 안 됨 → 매 outfit 변경 어려움.
• No validation: drift 누적 unnoticed.

🧪 검증 / 중복

• Verified (PuLID paper 2024, InstantID Tencent 2024, IP-Adapter Tencent 2023, diffusers docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — character/style consistency multi-stack pipeline.