2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/CFG 스케일(Classifier-Free Guidance Scale).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-cfg-scale-classifier-free-guidance	CFG Scale (Classifier-Free Guidance)	10_Wiki/Topics	verified	self	CFG classifier-free guidance guidance scale prompt strength negative prompt conditioning strength	none	A	0.93	applied	diffusion stable-diffusion cfg guidance sampling prompt-engineering dpm-solver conditioning		2026-05-10	pending	language framework Python Diffusers / ComfyUI / Automatic1111

CFG Scale (Classifier-Free Guidance)

📌 한 줄 통찰

"매 prompt 의 strict 의 dial". 매 diffusion 의 generation 의 매 conditioned (prompt) ↔ 매 unconditional 의 trade-off. 매 high CFG = 매 prompt 의 strict 가, 매 over-saturation. 매 sweet spot 7-9 (SDXL) / 3.5-7 (Flux).

📖 핵심

매 formula

\epsilon_{\text{guided}} = \epsilon_{\text{uncond}} + s \cdot (\epsilon_{\text{cond}} - \epsilon_{\text{uncond}})

• 매 s = CFG scale.
• 매 s = 1 → 매 unconditional (prompt 무시).
• 매 s = 7 → 매 typical.
• 매 s > 15 → 매 over-cooked.
• 매 negative prompt = 매 conditional 의 두 번째 (with -1 coefficient).

매 effect

CFG	결과
< 1	매 random / blank
1-3	매 loose, 매 creative
5-7	매 balanced (default)
7-12	매 prompt-strict
13-20	매 over-saturated, 매 burned
> 20	매 garbage

매 modern alternative

Flux (Black Forest Labs)

• 매 distilled CFG (CFG=1 가능).
• 매 inference 의 fast (no double pass).

Negative prompt

• 매 unconditional 의 noise 의 swap.
• 매 explicit avoidance.

Dynamic CFG

• 매 step 의 따라 변동.
• 매 early high → 매 late low.

Adaptive CFG (CFG++)

• 매 adaptive scale.
• 매 over-saturation 회피.

매 sampler 와 의 interaction

• 매 DPM++ 2M Karras: 20 step + CFG 7.
• 매 DPM++ SDE: 30 step + CFG 5.
• 매 Euler ancestral: 매 stochastic.
• 매 Flux: CFG=1 + 4-step.

매 prompt quality 와 의 관계

• 매 좋은 prompt + CFG 7 = 매 best.
• 매 나쁜 prompt + CFG ↑ = 매 더 나쁘게 (confident garbage).
• 매 negative prompt 의 keyword 매 wrong → 매 오히려 push.

→ CFG ↑ ≠ 매 quality ↑. 매 prompt quality 가 base.

매 typical setup

모델	CFG	Steps	Sampler
SD 1.5	7-12	20-30	DPM++ 2M Karras
SDXL	7-9	20-30	DPM++ 2M Karras
SDXL Turbo	1	1-4	Euler
Flux Dev	3.5	20-50	Euler
Flux Schnell	1	4	Euler

💻 패턴

Diffusers (basic)

from diffusers import StableDiffusionXLPipeline
import torch

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

image = pipe(
    prompt='a cat with a hat, oil painting, vivid colors',
    negative_prompt='blurry, low quality, watermark',
    guidance_scale=7.0,  # 매 CFG
    num_inference_steps=30,
).images[0]

CFG sweep (find sweet spot)

import matplotlib.pyplot as plt

cfg_values = [1, 3, 5, 7, 9, 12, 15, 20]
fig, axes = plt.subplots(2, 4, figsize=(20, 10))

for ax, cfg in zip(axes.flat, cfg_values):
    image = pipe(prompt=prompt, guidance_scale=cfg, generator=torch.manual_seed(42)).images[0]
    ax.imshow(image)
    ax.set_title(f'CFG={cfg}')
    ax.axis('off')
plt.show()

Flux (CFG=1, distilled)

from diffusers import FluxPipeline

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

image = pipe(
    prompt='a cat with a hat',
    guidance_scale=3.5,  # Flux dev
    num_inference_steps=50,
).images[0]

# Schnell (4-step, distilled)
pipe_schnell = FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-schnell', torch_dtype=torch.bfloat16).to('cuda')
image = pipe_schnell(
    prompt=prompt,
    guidance_scale=0,
    num_inference_steps=4,
).images[0]

Dynamic CFG

def dynamic_cfg_callback(pipe, step, timestep, callback_kwargs):
    """매 early step 의 high CFG, 매 late 의 low."""
    progress = step / pipe.num_inference_steps
    cfg = 12 - 7 * progress  # 12 → 5
    callback_kwargs['guidance_scale'] = cfg
    return callback_kwargs

pipe(prompt=prompt, callback_on_step_end=dynamic_cfg_callback)

Custom CFG implementation

def classifier_free_guidance(model, x_t, t, prompt_emb, neg_prompt_emb, scale):
    # 매 batched: cond + uncond
    emb_combined = torch.cat([neg_prompt_emb, prompt_emb])
    x_t_combined = torch.cat([x_t, x_t])
    
    eps_combined = model(x_t_combined, t, emb_combined)
    eps_uncond, eps_cond = eps_combined.chunk(2)
    
    # 매 guided
    return eps_uncond + scale * (eps_cond - eps_uncond)

CFG++ (adaptive)

def cfg_pp(eps_cond, eps_uncond, scale, x_t, alpha_t):
    """CFG++ — 매 over-saturation 회피."""
    cfg_basic = eps_uncond + scale * (eps_cond - eps_uncond)
    # 매 sample-space adjustment
    delta = (cfg_basic - eps_uncond) * (1 - alpha_t)
    return eps_uncond + scale * (eps_cond - eps_uncond) - delta

Negative prompt strategy

# 매 좋은 negative prompt
negative_prompts = {
    'photorealistic': 'cartoon, anime, painting, drawing',
    'illustration': 'photo, photograph, photographic',
    'quality': 'blurry, low quality, jpeg artifacts, watermark, signature, deformed, ugly',
    'anatomy': 'extra limbs, deformed hands, missing fingers, distorted face',
}

prompt = 'a portrait of a woman'
style = 'photorealistic'
neg = negative_prompts['quality'] + ', ' + negative_prompts[style]

image = pipe(prompt=prompt, negative_prompt=neg, guidance_scale=7).images[0]

🤔 결정 기준

상황	CFG
Photorealistic	7-9
Stylized art	8-12
Creative / loose	3-5
Strict prompt	10-15
Flux Dev	3.5
Flux Schnell / SDXL Turbo	1
Burning / over-saturated	< 7

기본값: SDXL = 7, Flux = 3.5, Schnell = 1.

🔗 Graph

• 부모: Diffusion-Models · AI 이미지 생성 (AI Image Generation)
• 응용: Stable-Diffusion · Flux · ComfyUI
• Adjacent: Negative Prompt

🤖 LLM 활용

언제: 매 image generation tuning. 매 SD / Flux pipeline. 매 quality vs prompt-fidelity trade-off. 언제 X: 매 distilled model (CFG=1). 매 deterministic 매 sampler.

❌ 안티패턴

• CFG 의 high 의 mean fix: 매 over-saturation.
• Negative prompt 의 wrong word + high CFG: 매 confident garbage.
• Same CFG 의 모든 model: 매 distilled vs base 의 다름.
• Sampler 의 mismatch: 매 sampler 별 의 sweet spot.
• CFG = 1 가 prompt 무시: 매 unconditional.

🧪 검증 / 중복

• Verified (Ho & Salimans CFG paper, Flux docs, Diffusers).
• 신뢰도 A.
• Related: Stable-Diffusion · Flux · Negative Prompt · DPM-Solver · AI Image Generation.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — formula + dynamic / Flux / sweep + 매 diffusers code