2nd/10_Wiki/Topic_Programming/AI_and_ML/버전 및 모델 (Versions and Models).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-버전-및-모델-versions-and-models	버전 및 모델 (Versions and Models)	10_Wiki/Topics	verified	self	Image Gen Model Versions Midjourney Versions FLUX Versions SDXL Lineage	none	A	0.9	applied	ai-image-generation model-versioning midjourney flux sdxl sora		2026-05-10	pending	language framework python diffusers

버전 및 모델 (Versions and Models)

매 한 줄

"매 model version 은 매 다른 aesthetic + capability profile". Midjourney v7, FLUX 1.2, SD 4 (Stable Diffusion), Sora 2, Imagen 4, DALL-E 4 — 매 2026 의 image-gen landscape 에서 version flag 의 매 careful selection 이 매 final output 의 quality 의 80% 결정.

매 핵심

매 2026 주요 모델 lineage

• Midjourney v7 (2026 Q1): 매 photorealistic + niji 7 (anime). --v 7, --niji 7.
• FLUX 1.2 (Black Forest Labs): open-weight, 매 prompt adherence 최강. flux-dev, flux-schnell, flux-pro.
• Stable Diffusion 4 / SD3.5-Large: 매 ComfyUI ecosystem 의 backbone. Native 16-channel VAE.
• DALL-E 4 (OpenAI): GPT-5 native multimodal, 매 conversational refinement.
• Imagen 4 (Google): text rendering 최강, 매 typography 작업 의 first choice.
• Recraft v4: vector + raster hybrid, 매 brand asset.
• Sora 2 / Veo 3 / Kling 2: video generation, 매 image 의 evolution.

매 Versioning 의 의미

• Aesthetic shift: v6 → v7 의 매 default style 이 painterly → photoreal 로 shift.
• Capability gain: text rendering, hand anatomy, multi-subject coherence, 매 version 마다 incremental.
• Param/flag breakage: 매 version 마다 supported flags 변경 (--style raw, --profile).

매 응용

1. Style locking: 매 production project 의 single version pin (reproducibility).
2. A/B comparison: 매 same prompt 의 multi-version sweep, 매 best select.
3. Hybrid pipeline: SD3.5 base → FLUX inpaint → Recraft vector trace.

💻 패턴

FLUX 1.2 via diffusers

from diffusers import FluxPipeline
import torch

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1.2-dev",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="a samurai in moonlit bamboo forest, cinematic, 35mm film grain",
    guidance_scale=3.5,
    num_inference_steps=28,
    max_sequence_length=512,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("out.png")

Midjourney v7 prompt with flags

ethereal forest spirit, glowing mushrooms, volumetric mist
--v 7 --style raw --ar 21:9 --s 250 --p personal_v3 --c 25

Version sweep harness

versions = ["flux-1.2-dev", "sd-3.5-large", "flux-pro"]
prompts = [...]
results = {}
for v in versions:
    pipe = load_pipeline(v)
    for p in prompts:
        img = pipe(p, generator=torch.Generator("cuda").manual_seed(42)).images[0]
        results[(v, p)] = img
        save(f"{v}/{slug(p)}.png", img)
make_contact_sheet(results)

Model registry (production)

@dataclass
class ModelVersion:
    name: str
    revision: str  # commit hash on HF
    vae: str
    text_encoders: list[str]
    pinned_at: datetime
    aesthetic_tags: list[str]

REGISTRY = {
    "hero-banner-v3": ModelVersion(
        name="black-forest-labs/FLUX.1.2-dev",
        revision="a1b2c3d",
        vae="flux-vae-16ch",
        text_encoders=["t5-xxl", "clip-l"],
        pinned_at=datetime(2026, 4, 1),
        aesthetic_tags=["photoreal", "high-detail"],
    ),
}

LoRA stack on top of base version

pipe.load_lora_weights("brand/logo-lora-flux12", adapter_name="brand")
pipe.load_lora_weights("style/cinematic-flux12", adapter_name="cine")
pipe.set_adapters(["brand", "cine"], adapter_weights=[0.8, 0.6])

ComfyUI workflow JSON snippet (version pin)

{
  "checkpoint": { "ckpt_name": "flux1.2-dev-fp8.safetensors" },
  "loras": [
    { "name": "filmgrain_v2.safetensors", "strength": 0.4 }
  ],
  "sampler": { "name": "euler", "scheduler": "simple", "steps": 28, "cfg": 3.5 }
}

매 결정 기준

목표	권장 모델 (2026)
Photorealism	FLUX 1.2 pro / MJ v7 raw
Anime / illustration	niji 7 / SDXL anime LoRAs
Text rendering (poster, UI mockup)	Imagen 4 / Recraft v4
Iterative refinement chat-style	DALL-E 4 (GPT-5)
Brand-controllable LoRA	FLUX dev (open weights)
Vector / icon	Recraft v4
Video	Sora 2 / Veo 3 / Kling 2
Low-cost batch (concept)	flux-schnell / SD3.5 turbo

기본값: FLUX 1.2-dev — open weights, LoRA ecosystem, prompt adherence 최강. 매 production 에서 pin specific revision.

🔗 Graph

• 부모: AI 이미지 생성 (AI Image Generation) · AI Image Generation Workflow (canonical)
• 변형: Midjourney-v7
• 응용: Style_Reference_(--sref)

🤖 LLM 활용

언제: model release notes 의 summary, version migration checklist, prompt syntax 의 version-specific 차이 체크. 언제 X: 매 actual aesthetic judgment — 매 visual A/B 가 ground truth. LLM 의 aesthetic claim 의 hallucination 빈번.

❌ 안티패턴

• No version pin: 매 production 의 reproducibility 죽음. 매 model card revision hash 필수.
• Latest = best 가정: 매 v7 이 v6 보다 specific style 에서 worse 의 사례 흔함.
• Mixing flags from different versions: 매 silent ignore, 매 debug 어려움.
• Single-model lock-in: 매 hybrid pipeline (one base, one inpaint, one upscale) 가 보통 best.

🧪 검증 / 중복

• Verified (Black Forest Labs FLUX 1.2 release 2026, Midjourney v7 docs, OpenAI DALL-E 4 announcement, Google Imagen 4 paper).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — 2026 image-gen model lineage + version pinning patterns