Files

T

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

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

6.1 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

인공지능 시각 언어 생성 (AI Visual Language Generation)

매 한 줄

"매 visual language 는 단순 style 이 아닌 systematic grammar". 2026 의 AI image gen 은 단발성 prompt 의 phase 를 지나, brand-grade visual grammar (color, composition, motif, lighting) 를 학습된 LoRA stack + style transfer + control net 으로 generate 하는 단계로 진입했다. FLUX, SDXL, Imagen 4 가 production-grade visual identity 의 backbone 이 됨.

매 핵심

매 visual language 의 component

Color palette: oklch tokens, dominant/accent ratio.
Composition rules: rule of thirds, negative space, symmetry/asymmetry.
Motif vocabulary: recurring shape, icon, texture.
Lighting model: rim/key/fill, time of day, mood.
Material/finish: matte/glossy, organic/synthetic.

매 generation stack (2026)

Base model: FLUX.1-dev / SDXL / Imagen 4.
Style LoRA: 30-100 ref images 로 finetune.
Subject LoRA: character/object identity.
ControlNet: pose, depth, edge, normal.
IP-Adapter: reference image guidance.
Regional prompting: per-region distinct style.

매 응용

Brand identity 의 marketing asset auto-gen.
Game art direction 의 concept art exploration.
Editorial illustration 의 series consistency.

💻 패턴

Style LoRA training (FLUX)

from diffusers import FluxPipeline
import torch
from peft import LoraConfig

# 1. Curate 50-100 ref images that share visual language
# 2. Caption with consistent trigger token
captions = ["<myStyle> a serene landscape, oil painting feel, ..."]

# 3. Train LoRA
lora_config = LoraConfig(
    r=32, lora_alpha=32,
    target_modules=["to_q","to_k","to_v","to_out.0"],
)
# train loop with 1500-3000 steps, lr=1e-4

Multi-LoRA stacking

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

# Stack: style + character
pipe.load_lora_weights("./styles/brand_v3.safetensors", adapter_name="style")
pipe.load_lora_weights("./chars/hero.safetensors",     adapter_name="char")
pipe.set_adapters(["style","char"], adapter_weights=[0.8, 0.9])

img = pipe(
    "<myStyle> <hero> standing on cliff at golden hour",
    num_inference_steps=28, guidance_scale=3.5
).images[0]

ControlNet + IP-Adapter

from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel

controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-depth-sdxl-1.0")
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet
).to("cuda")
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models",
                     weight_name="ip-adapter_sdxl.bin")
pipe.set_ip_adapter_scale(0.6)

img = pipe(
    prompt="cyberpunk skyline, our brand visual language",
    image=depth_map, ip_adapter_image=style_ref,
    num_inference_steps=30
).images[0]

Regional prompting (mask-based)

# StableDiffusion Forge / ComfyUI workflow
regions = [
  {"mask": top_half_mask,    "prompt": "<myStyle> dramatic sky, golden clouds"},
  {"mask": bottom_half_mask, "prompt": "<myStyle> reflective ocean, calm waves"},
]
img = regional_pipe(regions, base_prompt="<myStyle> seascape")

Visual grammar validation

# CLIP score against reference language vector
import open_clip
model, _, preprocess = open_clip.create_model_and_transforms("ViT-bigG-14")

ref_lang_vector = mean([model.encode_image(preprocess(r)) for r in ref_images])
gen_vec = model.encode_image(preprocess(generated))
similarity = cosine(ref_lang_vector, gen_vec)
assert similarity > 0.78, "style drift"

Palette enforcement post-process

import numpy as np
from sklearn.cluster import KMeans

def quantize_to_palette(img, palette_oklch):
    pixels = img.reshape(-1,3)
    palette_rgb = oklch_to_rgb(palette_oklch)
    # Snap each pixel to nearest palette color
    dists = np.linalg.norm(pixels[:,None,:] - palette_rgb[None,:,:], axis=2)
    nearest = np.argmin(dists, axis=1)
    return palette_rgb[nearest].reshape(img.shape).astype(np.uint8)

매 결정 기준

상황	Approach
brand 일관성 priority	LoRA + palette enforce
concept exploration	base model + prompt only
character + style 동시	Multi-LoRA stacking
정확한 layout	ControlNet (depth/canny)
1 ref image only	IP-Adapter
다른 style/region 의 분리	Regional prompting

기본값: FLUX.1-dev + style LoRA + palette post-process.

🔗 Graph

부모: AI Image Generation · Generative Models
변형: Style Transfer · LoRA Fine-tuning
응용: Brand Identity Generation · Concept Art Pipeline
Adjacent: ControlNet · IP-Adapter · 일관된 캐릭터 및 스타일 구축

🤖 LLM 활용

언제: visual brief 의 prompt expansion, LoRA caption batch authoring, palette extraction from refs. 언제 X: photographic accuracy 의 precise composition — manual photo shoot 가 정답.

❌ 안티패턴

Random prompt soup: each gen 마다 다른 keyword — no language emergence.
Single-image LoRA: overfit, mode collapse.
Skipping captions: trigger token 없으면 LoRA 가 always-on.
Negative prompt 만 의존: positive 의 vocabulary 정의가 우선.

🧪 검증 / 중복

Verified (Black Forest Labs FLUX docs 2025, diffusers library, IP-Adapter paper).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — visual language gen 의 LoRA + ControlNet stack.

6.1 KiB Raw Blame History