2nd/10_Wiki/Topics/AI_and_ML/Image Parameters.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-image-parameters	Image Parameters	10_Wiki/Topics	verified	self	Image Generation Parameters Diffusion Parameters Sampler Settings	none	A	0.9	applied	diffusion image-generation stable-diffusion sampler cfg seed		2026-05-10	pending	language framework python diffusers

Image Parameters

매 한 줄

"매 같은 prompt 라도 sampler / steps / CFG / seed 가 그림을 결정한다". Diffusion-based 이미지 생성의 결과는 prompt 만큼이나 sampler 종류, sampling steps, classifier-free guidance scale, seed, scheduler 같은 운영 파라미터에 의해 좌우되며, 2026 현재 SDXL/FLUX/SD3.5/Imagen 모든 모델에서 공통적으로 적용된다.

매 핵심

매 핵심 파라미터

• prompt / negative_prompt: 텍스트 조건. 길이 ≠ 품질.
• steps (sampling steps): 보통 20-50. FLUX-schnell 는 4 steps, Turbo/LCM 은 1-4.
• CFG scale (guidance_scale): 1.5 (창의) ~ 12 (충실). FLUX 는 distilled — guidance 1 권장.
• sampler / scheduler: Euler, DPM++ 2M Karras, UniPC, DDIM, LCM.
• seed: 동일 seed + 동일 파라미터 → 재현.
• width/height: SDXL 1024², SD3.5 1024², FLUX 1024² 또는 1024×1536.
• strength (img2img): 0 (원본) ~ 1 (전부 새로).
• clip_skip: 1 (기본) — anime 모델은 2 가 흔함.
• prompt weighting: (word:1.3), [word] (A1111), (word)+ (compel).
• LoRA scale: 0.6-0.9 일반, 1.0 강함.

매 sampler 비교 (2026)

• Euler / Euler a: 빠름, 단순.
• DPM++ 2M Karras: 표준 추천 — 품질/속도 균형.
• UniPC: 적은 steps 에서 강함.
• DDIM: deterministic, 재현성 좋음.
• LCM / Turbo: distilled, 1-4 steps.
• Heun, DPM++ SDE: 고품질, 느림.

매 응용

1. ComfyUI / A1111 / Forge UI 워크플로 작성.
2. SDK 기반 batch 생성 (diffusers).
3. ControlNet / IP-Adapter 결합.
4. 프로덕션 API (Replicate, Fal, RunPod).

💻 패턴

1. diffusers — 기본 SDXL 호출

from diffusers import StableDiffusionXLPipeline
import torch
pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16, variant="fp16",
).to("cuda")
img = pipe(
    prompt="a photo of a Shiba Inu, cinematic, 50mm",
    negative_prompt="lowres, watermark",
    num_inference_steps=30, guidance_scale=6.0,
    width=1024, height=1024,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
img.save("out.png")

2. FLUX (guidance-distilled)

from diffusers import FluxPipeline
pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
img = pipe(
    prompt="a tiny astronaut riding a duck, studio light",
    guidance_scale=3.5,   # FLUX-dev: 3-5
    num_inference_steps=28,
    height=1024, width=1024,
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]

3. LCM 1-4 steps (real-time)

from diffusers import AutoPipelineForText2Image, LCMScheduler
pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16).to("cuda")
img = pipe(prompt="cyberpunk cat", num_inference_steps=2, guidance_scale=0.0).images[0]

4. img2img + strength

from diffusers import StableDiffusionXLImg2ImgPipeline
pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(...).to("cuda")
img = pipe(prompt=prompt, image=init, strength=0.55, guidance_scale=7.0, num_inference_steps=40).images[0]

5. prompt weighting (compel)

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 (red:1.4) sports car, (sunset:1.2)"
conditioning, pooled = compel(prompt)
img = pipe(prompt_embeds=conditioning, pooled_prompt_embeds=pooled, num_inference_steps=30).images[0]

6. LoRA + scale

pipe.load_lora_weights("path/to/anime.safetensors", adapter_name="anime")
pipe.set_adapters(["anime"], adapter_weights=[0.7])
img = pipe(prompt="anime girl, blue hair", num_inference_steps=30, guidance_scale=6.5).images[0]

7. ControlNet (구조 보존)

from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel
cnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", controlnet=cnet, torch_dtype=torch.float16,
).to("cuda")
img = pipe(prompt="oil painting", image=canny_img,
           controlnet_conditioning_scale=0.7, num_inference_steps=30).images[0]

8. seed sweep (탐색)

import torch, itertools
for seed, cfg in itertools.product([0,1,2,3], [4.0, 6.0, 8.0]):
    g = torch.Generator("cuda").manual_seed(seed)
    img = pipe(prompt=prompt, guidance_scale=cfg, generator=g).images[0]
    img.save(f"seed{seed}_cfg{cfg}.png")

9. scheduler 교체

from diffusers import DPMSolverMultistepScheduler
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
    pipe.scheduler.config, algorithm_type="dpmsolver++", use_karras_sigmas=True,
)

10. ComfyUI API workflow (JSON 단편)

{
  "3": { "class_type": "KSampler",
    "inputs": { "seed": 42, "steps": 30, "cfg": 6.5,
                "sampler_name": "dpmpp_2m_sde", "scheduler": "karras",
                "denoise": 1.0, "model": ["4",0], "positive": ["6",0],
                "negative": ["7",0], "latent_image": ["5",0] } }
}

매 결정 기준

상황	Approach
SDXL 기본	DPM++ 2M Karras, 30 steps, CFG 6-7
FLUX-dev	Euler, 28 steps, guidance 3.5
FLUX-schnell / Turbo / LCM	Euler, 1-4 steps, guidance 0
사진 사실성	CFG 4-6, negative_prompt 충실
스타일 강조	LoRA 0.7-0.9 + CFG 6-8
구조 고정	ControlNet, scale 0.6-0.8
배치 탐색	seed sweep + grid

기본값: SDXL = DPM++ 2M Karras, 30 steps, CFG 6.5. FLUX-dev = Euler, 28 steps, guidance 3.5.

🔗 Graph

• 부모: Diffusion-Models · Generative-AI
• 변형: Prompt Weighting
• 응용: ControlNet · LoRA · IP-Adapter
• Adjacent: Stable-Diffusion · FLUX · ComfyUI

🤖 LLM 활용

언제: 파라미터 매핑 (UI ↔ API), prompt → weighted prompt 변환, sampler 선택 가이드, batch grid 코드 생성. 언제 X: 모델별 quirk (예: SD3.5 의 specific shift) — 모델 카드 직접 확인.

❌ 안티패턴

• CFG 무작정 12 이상: burn-out, oversaturated.
• steps 100+: 30-50 이후 거의 무차이, 비용만 증가.
• FLUX 에 CFG 7: distilled — guidance 1-4 가 정상.
• seed 미고정 디버깅: 재현 불가 → 파라미터 영향 분석 불가.
• Negative prompt 남용: SDXL/FLUX 는 짧게, 또는 비워도 OK.
• prompt weighting 과용: (word:2.0)+ → broken composition.

🧪 검증 / 중복

• Verified (HuggingFace diffusers docs, FLUX model card, ComfyUI wiki 2026).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — sampler/CFG/LoRA/ControlNet 코드 패턴