Files

T

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

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

6.2 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

부정 프롬프트와 가중치를 활용한 시각적 아티팩트(Artifact) 디버깅 및 제어

매 한 줄

"매 artifact 매 unwanted concept 의 누출". Diffusion model 의 산출물 의 extra fingers, melted face, watermark 의 매 negative prompt + token weighting 의 conditional vector 조정 의 억제. 매 2026 의 FLUX 매 negative prompt 의존도 ↓ — guidance distillation 으로 quality 자체 ↑.

매 핵심

매 메커니즘

Classifier-Free Guidance (CFG): noise = uncond + scale * (cond - uncond).
Negative prompt 의 uncond 의 대체 — 매 "이쪽 으로 가지 마" vector.
Token weighting ((token:1.3)) 매 cross-attention 의 token embedding scale.

매 흔한 artifact

Anatomy: extra fingers, deformed hands, asymmetric eyes.
Composition: cropped head, floating limbs, tangent edges.
Quality: blur, jpeg artifact, low resolution, oversaturation.
Concept leak: text/watermark, signature, logo.

매 응용

SDXL/SD3 매 negative prompt 의 default workflow 에 포함.
FLUX/SD3.5 매 prompt weighting 의존도 ↓ — guidance distilled.
Inpainting fix — masked region 만 negative prompt 의 적용.

💻 패턴

diffusers — negative prompt 기본

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16
).to("cuda")

img = pipe(
    prompt="portrait of a woman, studio light, sharp focus, 50mm lens",
    negative_prompt=(
        "deformed, extra fingers, mutated hands, asymmetric eyes, "
        "lowres, jpeg artifacts, watermark, signature, text, blurry"
    ),
    guidance_scale=7.0,
    num_inference_steps=30,
).images[0]

compel — token weighting (SDXL)

from compel import Compel, ReturnedEmbeddingsType

compel = Compel(
    tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
    text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
    requires_pooled=[False, True],
)

prompt = "a (majestic:1.3) lion, (golden mane:1.2), (cinematic:0.8)"
neg = "(cartoon:1.4), (anime:1.4), (3d render:1.2)"

embeds, pooled = compel(prompt)
neg_embeds, neg_pooled = compel(neg)

img = pipe(
    prompt_embeds=embeds, pooled_prompt_embeds=pooled,
    negative_prompt_embeds=neg_embeds, negative_pooled_prompt_embeds=neg_pooled,
).images[0]

A1111 syntax (community standard)

# 매 강조 — (token:weight)
masterpiece, (intricate details:1.2), (sharp focus:1.1)

# Negative
(worst quality:1.4), (low quality:1.4), (extra digits:1.3),
(bad hands:1.3), watermark, text

FLUX — minimal negative

# FLUX.1 [dev] 매 distilled — CFG ≈ 1, negative prompt 효과 ↓
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
img = pipe(
    prompt="cinematic portrait, golden hour, shallow DOF",
    guidance_scale=3.5,        # FLUX-specific
    num_inference_steps=28,
).images[0]
# 매 negative prompt 매 관습 의 대신 — positive prompt 의 specificity 의 의존

Region-specific negative (regional prompter)

# 매 ComfyUI / Forge — mask 별 negative
region_prompts = [
    {"mask": face_mask, "neg": "deformed, extra eye, asymmetric"},
    {"mask": hand_mask, "neg": "extra fingers, fused fingers, mutated"},
]

Artifact debug — A/B isolate

# 매 baseline (no negative)
img_base = pipe(prompt=p, negative_prompt="", seed=42).images[0]

# 매 add one term at a time
for term in ["deformed", "extra fingers", "lowres", "watermark"]:
    img = pipe(prompt=p, negative_prompt=term, seed=42).images[0]
    img.save(f"debug_{term}.png")
# 매 비교 의 어떤 term 의 효과 의 식별

매 결정 기준

상황	Approach
Hand artifact	`(bad hands:1.3), extra fingers, fused fingers`
Watermark/text	`watermark, signature, text, logo` (high weight)
Style leak (cartoon)	`(cartoon:1.4), (anime:1.4)`
FLUX/SD3.5 사용	negative prompt 의 minimal — positive 의 specificity ↑
Inpaint 의 fix	mask region 의 local negative

기본값: SDXL → standard negative bundle + weighting; FLUX → positive prompt 의 specificity.

🔗 Graph

부모: AI 이미지 생성 (AI Image Generation) · Diffusion Models
변형: Classifier-Free Guidance · Prompt Weighting
응용: AI 이미지 품질 최적화 및 디버깅 (Image Quality Optimization & Debugging) · 미드저니 및 스테이블 디퓨전의 부분 편집 기법
Adjacent: ControlNet · IP-Adapter · Regional Prompter

🤖 LLM 활용

언제: anatomy/composition artifact 추적, prompt A/B isolate, style leak 차단. 언제 X: FLUX-class distilled model — negative prompt 효과 ↓, positive specificity 의 의존.

❌ 안티패턴

Negative prompt 200-token wall: token budget 낭비, 효과 saturate.
High weight (>1.5): 매 collapse — output 의 distort.
Generic "ugly, bad": 매 의미 없음 — concrete artifact name 의 사용.
FLUX 의 SDXL-style negative: 매 non-effect — guidance distilled.
Seed 변경 의 비교: 매 무의미 — same seed 만 isolate.

🧪 검증 / 중복

Verified (Diffusers docs, compel library, AUTOMATIC1111 wiki, FLUX.1 model card 2024-2026).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — negative prompt + weighting, FLUX 차이

6.2 KiB Raw Blame History