stable diffusion generative ai

Stable Diffusion

Stable Diffusion เป็น Open-source Text-to-Image AI สร้างภาพจาก Prompt รันบนเครื่องตัวเองฟรี ปรับแต่งได้ทุกอย่าง Community ใหญ่มาก Models Extensions มากมาย

รองรับ txt2img img2img inpainting ControlNet LoRA Training Custom Models สำหรับงาน Production ใช้ ComfyUI หรือ A1111 WebUI

Installation และ Setup

# === Stable Diffusion Installation ===

# 1. AUTOMATIC1111 WebUI (แนะนำสำหรับเริ่มต้น)
# git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
# cd stable-diffusion-webui

# Windows
# webui-user.bat

# Linux/Mac
# ./webui.sh

# Arguments สำหรับ GPU VRAM น้อย
# set COMMANDLINE_ARGS=--medvram --xformers
# set COMMANDLINE_ARGS=--lowvram  # สำหรับ 4GB VRAM

# 2. ComfyUI (แนะนำสำหรับ Advanced/Production)
# git clone https://github.com/comfyanonymous/ComfyUI.git
# cd ComfyUI
# pip install -r requirements.txt
# python main.py

# 3. Diffusers (Python Library)
# pip install diffusers transformers accelerate torch

from dataclasses import dataclass, field
from typing import List, Dict, Optional

@dataclass
class SDConfig:
    """Stable Diffusion Configuration"""
    model: str = "stabilityai/stable-diffusion-xl-base-1.0"
    scheduler: str = "DPM++ 2M Karras"
    steps: int = 30
    cfg_scale: float = 7.0
    width: int = 1024
    height: int = 1024
    seed: int = -1  # -1 = random
    sampler: str = "Euler a"
    clip_skip: int = 2
    vae: str = "auto"

    # Performance
    use_xformers: bool = True
    use_fp16: bool = True
    batch_size: int = 1

@dataclass
class SDPrompt:
    positive: str
    negative: str = ""
    lora: List[str] = field(default_factory=list)
    controlnet: Optional[str] = None

    def build(self) -> Dict:
        prompt = self.positive
        # Add LoRA
        for lora in self.lora:
            prompt += f" "
        return {
            "prompt": prompt,
            "negative_prompt": self.negative or
                "low quality, bad anatomy, blurry, watermark, text, "
                "deformed, ugly, duplicate, error",
        }

# Recommended Models
models = {
    "SDXL 1.0": "1024x1024, General purpose, High quality",
    "SD 1.5": "512x512, Fast, Huge ecosystem of LoRAs",
    "Juggernaut XL": "Photorealistic, Portraits",
    "DreamShaper XL": "Fantasy, Illustration",
    "RealVisXL": "Photorealistic, Product photos",
    "AnimagineXL": "Anime style",
}

# Samplers
samplers = {
    "DPM++ 2M Karras": "Best overall quality/speed balance",
    "Euler a": "Creative, good for exploration",
    "DPM++ SDE Karras": "High quality, slower",
    "DDIM": "Deterministic, good for img2img",
}

config = SDConfig()
print("Stable Diffusion Config:")
print(f"  Model: {config.model}")
print(f"  Steps: {config.steps} | CFG: {config.cfg_scale}")
print(f"  Size: {config.width}x{config.height}")
print(f"  Sampler: {config.sampler}")

print(f"\nRecommended Models:")
for name, desc in models.items():
    print(f"  {name}: {desc}")

Prompt Writing และ ControlNet

# sd_prompts.py — Stable Diffusion Prompt Engineering
from dataclasses import dataclass, field
from typing import List

@dataclass
class PromptTemplate:
    name: str
    category: str
    positive: str
    negative: str
    settings: dict = field(default_factory=dict)

class SDPromptLibrary:
    """Stable Diffusion Prompt Library"""

    def __init__(self):
        self.templates: List[PromptTemplate] = []

    def add(self, template: PromptTemplate):
        self.templates.append(template)

    def get_by_category(self, category: str):
        return [t for t in self.templates if t.category == category]

    def show_all(self):
        categories = set(t.category for t in self.templates)
        for cat in sorted(categories):
            print(f"\n  [{cat}]")
            for t in self.get_by_category(cat):
                print(f"    {t.name}:")
                print(f"      + {t.positive[:70]}...")

# Prompt Templates
library = SDPromptLibrary()

templates = [
    PromptTemplate(
        "Product Photo", "Commercial",
        "professional product photography, luxury perfume bottle on marble surface, "
        "soft studio lighting, shallow depth of field, bokeh background, "
        "8k uhd, sharp focus, commercial quality",
        "low quality, blurry, watermark, text, deformed",
        {"steps": 35, "cfg": 7.5, "size": "1024x1024"},
    ),
    PromptTemplate(
        "Portrait", "Photography",
        "professional portrait photography, beautiful woman, natural skin texture, "
        "golden hour sunlight, shallow depth of field, canon EOS R5, "
        "85mm lens f/1.4, raw photo, film grain",
        "bad anatomy, deformed face, ugly, blurry, watermark",
        {"steps": 30, "cfg": 7.0, "size": "832x1216"},
    ),
    PromptTemplate(
        "Anime Character", "Illustration",
        "masterpiece, best quality, 1girl, cyberpunk outfit, neon city background, "
        "detailed eyes, dynamic pose, night scene, rain, reflections, "
        "volumetric lighting, anime style",
        "low quality, worst quality, bad anatomy, extra fingers",
        {"steps": 25, "cfg": 8.0, "size": "832x1216"},
    ),
    PromptTemplate(
        "Architecture", "Design",
        "architectural visualization, modern minimalist house, "
        "floor-to-ceiling windows, infinity pool, tropical garden, "
        "golden hour, dramatic sky, photorealistic rendering, 8k",
        "low quality, blurry, watermark, people",
        {"steps": 40, "cfg": 7.0, "size": "1216x832"},
    ),
    PromptTemplate(
        "UI Dashboard", "Web Design",
        "modern SaaS dashboard UI design, dark mode, data visualization, "
        "clean minimal layout, glassmorphism, gradient accents, "
        "professional design, figma mockup style",
        "realistic photo, 3d render, blurry text",
        {"steps": 30, "cfg": 7.0, "size": "1216x832"},
    ),
]

for t in templates:
    library.add(t)

library.show_all()

# ControlNet Types
controlnet_types = {
    "Canny": "ตรวจจับเส้นขอบ ควบคุม Structure ของภาพ",
    "OpenPose": "ตรวจจับท่าทางร่างกาย ควบคุม Pose",
    "Depth": "ตรวจจับความลึก ควบคุม Perspective",
    "Scribble": "ใช้ร่างคร่าวๆ เป็น Guide",
    "Tile": "Upscale ภาพ เพิ่มรายละเอียด",
    "IP-Adapter": "ใช้ภาพต้นแบบ ควบคุม Style",
    "Reference": "ใช้ภาพ Reference สำหรับ Style Transfer",
}

print(f"\n\nControlNet Types:")
for cn_type, desc in controlnet_types.items():
    print(f"  {cn_type}: {desc}")

LoRA Training และ Deployment

# === LoRA Training สำหรับ Stable Diffusion ===
# pip install kohya-ss

# 1. Dataset Preparation
# dataset/
# ├── img/
# │   ├── 20_concept_name/     # 20 = repeat count
# │   │   ├── image_001.png
# │   │   ├── image_001.txt    # Caption
# │   │   ├── image_002.png
# │   │   └── image_002.txt
# ├── reg/                      # Regularization images (optional)
# │   └── 1_class/
# │       ├── reg_001.png
# │       └── reg_002.png

# 2. Training Config (kohya_ss)
training_config = {
    # Model
    "pretrained_model": "stabilityai/stable-diffusion-xl-base-1.0",
    "output_dir": "./output/my_lora",
    "output_name": "my_custom_lora",

    # LoRA
    "network_dim": 32,        # LoRA rank (8-128)
    "network_alpha": 16,      # Alpha (usually rank/2)
    "network_module": "networks.lora",

    # Training
    "learning_rate": 1e-4,
    "unet_lr": 1e-4,
    "text_encoder_lr": 5e-5,
    "lr_scheduler": "cosine_with_restarts",
    "train_batch_size": 1,
    "max_train_epochs": 20,
    "resolution": "1024,1024",  # SDXL
    "mixed_precision": "bf16",

    # Optimization
    "optimizer_type": "AdamW8bit",
    "xformers": True,
    "gradient_checkpointing": True,
    "cache_latents": True,
    "cache_latents_to_disk": True,

    # Dataset
    "train_data_dir": "./dataset/img",
    "reg_data_dir": "./dataset/reg",
}

# 3. Training Command (kohya_ss)
# accelerate launch --num_cpu_threads_per_process=2 \
#   sdxl_train_network.py \
#   --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
#   --train_data_dir="./dataset/img" \
#   --output_dir="./output" \
#   --output_name="my_lora" \
#   --network_module=networks.lora \
#   --network_dim=32 \
#   --network_alpha=16 \
#   --learning_rate=1e-4 \
#   --lr_scheduler=cosine_with_restarts \
#   --train_batch_size=1 \
#   --max_train_epochs=20 \
#   --resolution=1024,1024 \
#   --mixed_precision=bf16 \
#   --optimizer_type=AdamW8bit \
#   --xformers \
#   --gradient_checkpointing \
#   --cache_latents

# 4. Using LoRA in A1111
# วาง .safetensors ไว้ใน models/Lora/
# ใช้ใน Prompt: 

# 5. ComfyUI Workflow
# Load Checkpoint -> Load LoRA -> KSampler -> VAE Decode -> Save Image

# 6. Production API
# from diffusers import StableDiffusionXLPipeline, AutoencoderKL
# import torch
#
# pipe = StableDiffusionXLPipeline.from_pretrained(
#     "stabilityai/stable-diffusion-xl-base-1.0",
#     torch_dtype=torch.float16,
#     variant="fp16",
# ).to("cuda")
#
# pipe.load_lora_weights("./output/my_lora.safetensors")
#
# image = pipe(
#     prompt="your prompt here",
#     negative_prompt="low quality",
#     num_inference_steps=30,
#     guidance_scale=7.0,
#     width=1024, height=1024,
# ).images[0]
#
# image.save("output.png")

print("LoRA Training Config:")
for key, value in training_config.items():
    print(f"  {key}: {value}")

Best Practices

• Model Selection: SDXL สำหรับคุณภาพสูง SD 1.5 สำหรับความเร็วและ LoRA ecosystem
• Prompt Structure: Subject + Style + Lighting + Details + Quality Tags
• Negative Prompt: ใส่เสมอ ลบสิ่งไม่ต้องการ low quality, blurry, watermark
• ControlNet: ใช้ Canny/OpenPose ควบคุม Structure ของภาพ
• LoRA Training: ใช้ 20-50 ภาพคุณภาพสูง Caption ให้ละเอียด
• Batch Processing: ใช้ ComfyUI Workflow สำหรับ Production Pipeline

Stable Diffusion คืออะไร

Open-source Text-to-Image AI Stability AI สร้างภาพจาก Prompt รันเครื่องตัวเองฟรี img2img inpainting ControlNet LoRA Community ใหญ่ Models Extensions มาก

สรุป

Stable Diffusion เป็น Open-source Text-to-Image AI รันเครื่องตัวเองฟรี SDXL คุณภาพสูง SD 1.5 เร็ว LoRA ecosystem ControlNet ควบคุม Structure LoRA Training ด้วย kohya_ss ComfyUI สำหรับ Production Diffusers Library สำหรับ Python API