2nd/10_Wiki/Topics/Computer_Science_and_Theory/Aesthetic-Value.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-aesthetic-value	Aesthetic Value	10_Wiki/Topics	verified	self	Aesthetics Beauty Theory Aesthetic Judgment	none	A	0.85	applied	philosophy aesthetics axiology design computational-aesthetics		2026-05-10	pending	language framework Python CLIP/aesthetic-predictor

Aesthetic Value

매 한 줄

"매 beauty 의 measurable — 매 subjective 의 X, 매 inter-subjective regularity 의 model.". Aesthetic value 의 philosophy (Kant, Hume) 의 root, 매 2026 의 computational aesthetics (CLIP-aesthetic, LAION predictor, FLUX-Pro reward model) 의 design / image-gen / UI optimization 의 quantified.

매 핵심

매 Theories

• Kant's "disinterested pleasure": aesthetic judgment 의 free of utility / desire.
• Hume's "delicacy of taste": trained sensibility 의 inter-subjective standard.
• Formalism (Bell, Fry): significant form — composition / line / color.
• Expressivism (Collingwood): art 의 emotion 의 expression.
• Institutional theory (Dickie): artworld 의 designation.

매 Computational Aesthetics

• LAION-Aesthetics predictor: CLIP embedding → MLP → 1-10 score.
• PickScore / HPSv2: human-preference reward model for image-gen.
• FLUX-Pro / Imagen 3 reward: aesthetic + prompt-alignment dual reward.
• A/B testing: empirical preference (UI design).
• Birkhoff's M = O / C: Order over Complexity (1933).

매 응용

1. Image generation reward (FLUX, SD3, Imagen 3 RLHF).
2. UI / design system scoring.
3. Photo curation (Apple Photos, Google Photos auto-pick).
4. Stock image ranking.

💻 패턴

Pattern 1 — LAION aesthetic score

import torch, clip
from huggingface_hub import hf_hub_download

device = "cuda" if torch.cuda.is_available() else "cpu"
clip_model, preprocess = clip.load("ViT-L/14", device=device)
mlp_path = hf_hub_download("LAION-AI/aesthetic-predictor", "sa_0_4_vit_l_14_linear.pth")
mlp = torch.nn.Linear(768, 1).to(device)
mlp.load_state_dict(torch.load(mlp_path))

def score(img_pil):
    with torch.no_grad():
        emb = clip_model.encode_image(preprocess(img_pil).unsqueeze(0).to(device))
        emb = emb / emb.norm(dim=-1, keepdim=True)
        return mlp(emb.float()).item()

Pattern 2 — PickScore reward (HF)

from transformers import AutoProcessor, AutoModel
proc = AutoProcessor.from_pretrained("yuvalkirstain/PickScore_v1")
model = AutoModel.from_pretrained("yuvalkirstain/PickScore_v1").cuda()

def pick_score(prompt, image):
    inputs = proc(text=prompt, images=image, return_tensors="pt", padding=True).to("cuda")
    with torch.no_grad():
        return model(**inputs).logits_per_image.item()

Pattern 3 — Birkhoff order/complexity

def birkhoff(order_count: int, complexity: int) -> float:
    return order_count / max(complexity, 1)

Pattern 4 — RLHF aesthetic reward (training)

# DDPO-style: gradient through diffusion sampling chain
def reward_fn(images, prompts):
    return 0.5 * laion_aesthetic(images) + 0.5 * pick_score(prompts, images)

매 결정 기준

상황	Approach
Photo curation	LAION-aesthetic
Image-gen RLHF	PickScore + HPSv2 ensemble
UI / web design	A/B test + heatmap
Art history analysis	Formalism + expert label

기본값: ensemble (LAION + PickScore + human eval).

🔗 Graph

• 부모: Axiology

🤖 LLM 활용

언제: image / design quality reward, preference-tuned generation, large-scale curation. 언제 X: pure subjective single-user use (preference learn), ethical/cultural sensitive context (model bias).

❌ 안티패턴

• Single-metric absolutism: LAION 의 over-fit (saturated colors).
• Ignoring cultural bias: training data 의 Western/Instagram bias.
• No human spot-check: reward gaming → aesthetic collapse.
• Treating subjective as objective: 매 score 의 ranking 의 X distance.

🧪 검증 / 중복

• Verified (LAION-Aesthetics paper, PickScore NeurIPS 2023, FLUX technical report).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — FULL content (CLIP-aesthetic, PickScore, RLHF)