---
id: wiki-2026-0508-오픈소스-기반-맞춤형-이미지-생성-워크플로우-구축
title: 오픈소스 기반 맞춤형 이미지 생성 워크플로우 구축
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [OSS Image Gen Workflow, ComfyUI Pipeline, Local Diffusion Stack]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [diffusion, comfyui, flux, lora, image-gen, oss]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: ComfyUI-FLUX
---

# 오픈소스 기반 맞춤형 이미지 생성 워크플로우 구축

## 매 한 줄
> **"매 base model + LoRA + ControlNet + upscaler = production stack"**. 2026 기준 FLUX.1 (Black Forest Labs) 가 SDXL 을 대체하며 OSS image gen 의 base, ComfyUI 가 node-based orchestration 의 표준. 8GB VRAM 부터 가능, 24GB 이상에서 batch + LoRA stacking 자유.

## 매 핵심

### 매 stack 구성
- **Base model**: FLUX.1-dev (12B), FLUX.1-schnell (4-step), SD3.5-Large (8B), SDXL (legacy).
- **Adapter**: LoRA (1-100MB) · IPAdapter (style transfer) · ControlNet (pose/depth/canny).
- **Upscaler**: 4x-UltraSharp, RealESRGAN, SUPIR (face-aware).
- **Orchestrator**: ComfyUI (node graph) · InvokeAI (UX) · A1111 (legacy webui).

### 매 hardware tier
- **8GB (RTX 3060/4060)**: schnell + GGUF Q4 · 1024² · 30s/img.
- **16GB (4080/4090)**: dev fp8 · LoRA stack · 60s/img.
- **24GB+ (4090/5090)**: dev fp16 · 2K render · ControlNet 동시 3개.
- **Apple Silicon (M3/M4 32GB+)**: MLX-FLUX · DiffusionKit.

### 매 응용
1. 브랜드 일관 캐릭터 — character LoRA + IPAdapter face.
2. 제품 사진 — ControlNet depth + product photo LoRA.
3. 컨셉 아트 — style LoRA + 4-stage upscale (1K→4K).

## 💻 패턴

### ComfyUI 기본 FLUX workflow (Python API)
```python
import json, urllib.request

workflow = {
  "3": {"class_type": "KSampler", "inputs": {
    "seed": 42, "steps": 20, "cfg": 3.5,
    "sampler_name": "euler", "scheduler": "simple",
    "model": ["10", 0], "positive": ["6", 0],
    "negative": ["7", 0], "latent_image": ["5", 0]
  }},
  "5": {"class_type": "EmptyLatentImage",
        "inputs": {"width": 1024, "height": 1024, "batch_size": 1}},
  "6": {"class_type": "CLIPTextEncode",
        "inputs": {"text": "studio portrait, soft light", "clip": ["10", 1]}},
  "10": {"class_type": "CheckpointLoaderSimple",
         "inputs": {"ckpt_name": "flux1-dev.safetensors"}}
}
urllib.request.urlopen("http://127.0.0.1:8188/prompt",
  data=json.dumps({"prompt": workflow}).encode())
```

### LoRA stacking (캐릭터 + 스타일)
```python
# ComfyUI LoraLoader 노드 직렬화
"15": {"class_type": "LoraLoader", "inputs": {
  "model": ["10", 0], "clip": ["10", 1],
  "lora_name": "char_alice_v3.safetensors",
  "strength_model": 0.85, "strength_clip": 1.0
}},
"16": {"class_type": "LoraLoader", "inputs": {
  "model": ["15", 0], "clip": ["15", 1],
  "lora_name": "watercolor_style.safetensors",
  "strength_model": 0.55, "strength_clip": 0.7
}}
```

### LoRA 학습 (kohya-ss FLUX)
```bash
accelerate launch sd-scripts/flux_train_network.py \
  --pretrained_model_name_or_path flux1-dev.safetensors \
  --train_data_dir ./dataset/alice \
  --output_dir ./out --output_name char_alice_v3 \
  --network_module networks.lora_flux \
  --network_dim 32 --network_alpha 16 \
  --resolution 1024 --train_batch_size 1 \
  --gradient_accumulation_steps 4 \
  --learning_rate 1e-4 --max_train_steps 1500 \
  --mixed_precision bf16 --save_precision bf16 \
  --cache_text_encoder_outputs --cache_latents
```

### ControlNet (pose-guided)
```python
"20": {"class_type": "ControlNetLoader",
       "inputs": {"control_net_name": "flux-controlnet-pose-v3.safetensors"}},
"21": {"class_type": "LoadImage", "inputs": {"image": "pose_ref.png"}},
"22": {"class_type": "ControlNetApply", "inputs": {
  "conditioning": ["6", 0], "control_net": ["20", 0],
  "image": ["21", 0], "strength": 0.65
}}
```

### SUPIR upscale (face-aware 4x)
```python
"30": {"class_type": "SUPIR_Upscale", "inputs": {
  "image": ["8", 0], "scale": 4,
  "model": "SUPIR-v0F.ckpt",
  "denoise": 0.35, "cfg_scale": 4.0,
  "color_fix_type": "Wavelet"
}}
```

### Batch automation (재현 가능)
```python
import itertools, hashlib
seeds   = [42, 1234, 7777]
prompts = ["forest", "city night", "studio"]
for p, s in itertools.product(prompts, seeds):
    hsh = hashlib.md5(f"{p}-{s}".encode()).hexdigest()[:8]
    run_workflow(prompt=p, seed=s, save_as=f"out_{hsh}.png")
```

### MLX-FLUX (Apple Silicon)
```bash
pip install mflux
mflux-generate --model dev \
  --prompt "isometric tiny city, paper craft" \
  --steps 20 --seed 42 --quantize 8 \
  --output out.png
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 8GB VRAM, 빠른 iteration | FLUX.1-schnell GGUF Q4. |
| 캐릭터 일관성 필요 | character LoRA + IPAdapter. |
| 제품 컴포지션 강제 | ControlNet depth/canny. |
| 4K 출력 | 1K base → SUPIR 4x. |
| Mac M3/M4 | MLX-FLUX (mflux). |

**기본값**: FLUX.1-dev fp8 · ComfyUI · 1024² · 20 steps · cfg 3.5 · LoRA 0.7-0.9.

## 🔗 Graph
- 부모: [[Diffusion-Models]] · [[Generative-AI]]
- 변형: [[FLUX]]
- 응용: [[ControlNet]] · [[IPAdapter]]
- Adjacent: [[ComfyUI]]

## 🤖 LLM 활용
**언제**: ComfyUI workflow JSON 작성, LoRA 학습 스크립트 튜닝, prompt engineering.
**언제 X**: real-time inference (< 1s) — closed API (Midjourney/DALL-E) 가 여전히 지연 면 우위.

## ❌ 안티패턴
- **LoRA strength 1.0 fixed**: overcooked 결과. 0.6-0.85 range 가 sweet spot.
- **CFG 7+ on FLUX**: artifact. FLUX 는 3-4가 정상 (SDXL 의 7과 다름).
- **No seed control**: 재현 불가능 → A/B 비교 불능.
- **VAE mismatch**: SDXL VAE 를 FLUX 에 사용 시 색 깨짐.

## 🧪 검증 / 중복
- Verified (Black Forest Labs FLUX docs, ComfyUI repo, kohya-ss flux_train).
- 신뢰도 A — 2024-2026 OSS image gen 의 표준.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — OSS image gen workflow (FLUX/ComfyUI) |