Files

T

Antigravity Agent 504fd5fb42 [G1-Sync] Manual knowledge update

2026-05-10 22:08:15 +09:00

7.4 KiB

Raw Blame History

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

title

시리즈물 및 다중 샷 워크플로우

매 한 줄

"매 single-shot 의 generation 은 1세대, 매 multi-shot consistency 가 2세대 의 challenge". 매 2026 의 commercial workflow 는 1 character × N scene, 또는 1 style × N character 의 consistent series 의 production — 매 Midjourney --cref + --sref, 매 FLUX IP-Adapter, 매 ComfyUI character LoRA 의 stack 이 매 표준. 매 single best image 의 시대 는 끝남.

매 핵심

매 series workflow 의 dimension

Character lock: 매 same person 의 N pose/scene.
Style lock: 매 same painterly look 의 N subject.
World lock: 매 same environment lighting/mood.
Outfit lock: 매 same clothing 의 different pose.

매 도구 stack (2026)

Midjourney V8: --cref (char) + --sref (style) + --ow (outfit weight).
FLUX.1 + IP-Adapter Plus + LoRA: 매 reproducible, customizable.
ComfyUI + InstantID / PuLID: 매 face-only ID lock.
Krea Realtime: 매 fast iteration storyboard.
Runway Gen-4: 매 video extension.

매 typical pipeline

Character sheet 의 generation (front/side/back, neutral lighting).
Reference image 의 selection (best of 4–8).
Optional: LoRA 의 train (Replicate, fal, RunPod).
Multi-shot generation (cref + sref, 또는 LoRA + sref).
Post-edit (face restore, hand fix).
Color/style 의 batch grading.

매 응용

Comic/manhwa panels.
Brand campaign (model 5–10 shot).
Game character sheet (idle/run/attack/death).
Children's book illustration series.
Storyboard for video.

💻 패턴

Pattern 1 — Midjourney character + style lock

# Step 1 — Character sheet
/imagine character sheet of a young female adventurer, leather jacket, \
  short red hair, freckles, neutral expression, T-pose, white background \
  --ar 1:1 --v 7

# Step 2 — Save best as cref URL
# Step 3 — Generate scenes
/imagine the same character climbing a snowy mountain \
  --cref https://cdn.example.com/char.png --cw 100 \
  --sref 2934852919 --sw 200 --ar 16:9 --v 7

/imagine the same character drinking tea in a tavern at night \
  --cref https://cdn.example.com/char.png --cw 100 \
  --sref 2934852919 --sw 200 --ar 16:9 --v 7

Pattern 2 — Storyboard batch (Python)

from itertools import product

CHAR = "https://cdn.example.com/char.png"
STYLE = 2934852919
SCENES = [
    "climbing a snowy mountain at dawn",
    "fighting a wolf in deep forest",
    "drinking tea in candle-lit tavern",
    "riding a horse across plains under storm",
]
ARS = ["16:9"]

for scene, ar in product(SCENES, ARS):
    print(f"/imagine the same character {scene} "
          f"--cref {CHAR} --cw 100 --sref {STYLE} --sw 200 --ar {ar} --v 7")

Pattern 3 — FLUX IP-Adapter character lock

from diffusers import FluxPipeline
from ip_adapter_flux import IPAdapterFluxPlus
import torch

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev",
                                    torch_dtype=torch.bfloat16).to("cuda")
adapter = IPAdapterFluxPlus(pipe, "ip-adapter-plus_flux.bin")

ref = load_image("char_sheet.png")
scenes = ["mountain summit at sunset", "neon-lit alley, rain", "tavern, candlelight"]

for i, scene in enumerate(scenes):
    out = adapter.generate(image=ref,
                           prompt=f"the same character, {scene}, cinematic, 35mm",
                           num_inference_steps=30, guidance_scale=4.0,
                           seed=42 + i).images[0]
    out.save(f"shot_{i:02d}.png")

Pattern 4 — InstantID face lock (ComfyUI)

# ComfyUI workflow (JSON snippet)
{
  "InstantIDLoader": {"face_image": "ref.png"},
  "ControlNet":      {"model": "instantid_controlnet"},
  "KSampler":        {"steps": 28, "cfg": 5.0, "seed": 1234},
  "CLIPTextEncode":  {"text": "the same person, standing on a beach, golden hour"}
}
# Repeat with different prompt → identity stays locked.

Pattern 5 — LoRA training (Replicate)

# Train character LoRA
cog predict r8.im/ostris/flux-dev-lora-trainer \
  -i input_images=@char_dataset.zip \
  -i trigger_word=AYAKO \
  -i steps=2000 -i lora_rank=16
# → flux-lora-AYAKO.safetensors

# Use:
/imagine AYAKO walking in Tokyo neon street at night --v flux-lora-AYAKO --ar 16:9

Pattern 6 — Outfit / wardrobe consistency

# V8 outfit lock with --ow
/imagine the same character on a beach in same outfit \
  --cref char.png --cw 100 --ow 100 \
  --sref 2934852919 --sw 200 --ar 16:9 --v 8

Pattern 7 — Series QA grid

# Build a contact-sheet of 12 shots for visual QA
from PIL import Image
shots = [Image.open(f"shot_{i:02d}.png") for i in range(12)]
W, H = 1024, 576
sheet = Image.new("RGB", (W*4, H*3))
for i, im in enumerate(shots):
    sheet.paste(im.resize((W,H)), ((i%4)*W, (i//4)*H))
sheet.save("series_qa.png")

Pattern 8 — Video extension (Runway / Kling)

# After image series approved → animate
import requests
for i in range(12):
    r = requests.post("https://api.runwayml.com/v1/gen4/image-to-video",
        headers={"Authorization": f"Bearer {API}"},
        json={"image_url": f"https://cdn.example.com/shot_{i:02d}.png",
              "prompt": "subtle camera push-in, 5s, cinematic",
              "duration": 5, "model": "gen-4-turbo"})
    print(r.json()["video_url"])

매 결정 기준

상황	Approach
Quick 5-shot storyboard	Midjourney cref + sref
Photo-real, 20+ shots, character	FLUX + InstantID
100+ shot, very specific person	LoRA train + FLUX
Same outfit, multiple poses	V8 cref + ow
Stylized illustration series	sref + cref + Niji V7

기본값: 매 5–20 shots → Midjourney cref+sref. 매 100+ → LoRA.

🔗 Graph

부모: AI 이미지 생성 (AI Image Generation) · Character Consistency
변형: Character_Reference · InstantID · IP-Adapter
응용: 상업용 브랜드 이미지 및 디자인 시스템 구축 · Storyboard
Adjacent: 스타일 코드 · 사후 편집 (Post-editing) · LoRA Training

🤖 LLM 활용

언제: 매 storyboard scene list 의 generation, 매 prompt batch 의 expansion (장면 변형). 언제 X: 매 visual continuity 의 final judgment — 매 art director 의 eye 가 필요.

❌ 안티패턴

No reference sheet: 매 매 batch 마다 character drift.
Too high cw + free prompt: 매 character pose freeze (overfit).
Multiple cref 동시: 매 face blend → 매 third person.
Style 의 mid-series 변경: 매 final compilation 의 jarring.

🧪 검증 / 중복

Verified (Midjourney V7/V8 docs, FLUX IP-Adapter docs, Replicate LoRA trainer).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — multi-shot pipeline + 8 patterns

7.4 KiB Raw Blame History Unescape Escape