stable diffusion colab

Stable Diffusion บน Colab

Stable Diffusion Google Colab Text-to-Image AUTOMATIC1111 ComfyUI Diffusers SDXL LoRA ControlNet Prompt GPU T4 A100 Free

Colab Setup & Pipeline

# === Stable Diffusion on Google Colab ===

# Cell 1: Install Dependencies
# !pip install diffusers transformers accelerate safetensors
# !pip install xformers  # Memory optimization

# Cell 2: Load Model
# from diffusers import StableDiffusionXLPipeline
# import torch
#
# pipe = StableDiffusionXLPipeline.from_pretrained(
#     "stabilityai/stable-diffusion-xl-base-1.0",
#     torch_dtype=torch.float16,
#     variant="fp16",
#     use_safetensors=True
# )
# pipe = pipe.to("cuda")
# pipe.enable_xformers_memory_efficient_attention()
#
# # Cell 3: Generate Image
# prompt = "masterpiece, best quality, 1girl, long hair, blue eyes"
# negative = "worst quality, low quality, blurry, deformed"
# image = pipe(prompt, negative_prompt=negative,
#              num_inference_steps=30, guidance_scale=7.5).images[0]
# image.save("output.png")

from dataclasses import dataclass

@dataclass
class ColabSetup:
    step: str
    code: str
    time: str
    note: str

steps = [
    ColabSetup("Install Dependencies",
        "pip install diffusers transformers accelerate xformers",
        "2-3 นาที",
        "ติดตั้งครั้งเดียวต่อ Session"),
    ColabSetup("Download Model",
        "StableDiffusionXLPipeline.from_pretrained()",
        "5-10 นาที (SDXL ~6.5GB)",
        "Cache ใน /root/.cache ถ้า Session เดิม"),
    ColabSetup("Load to GPU",
        "pipe.to('cuda') + enable_xformers",
        "10-30 วินาที",
        "ใช้ float16 ลด VRAM 50%"),
    ColabSetup("Generate Image",
        "pipe(prompt, steps=30, cfg=7.5)",
        "10-30 วินาที (T4) 3-10 วินาที (A100)",
        "SDXL 1024x1024 ต้อง T4+ VRAM 15GB+"),
    ColabSetup("Save & Download",
        "image.save() + files.download()",
        "ทันที",
        "หรือ Upload ไป Google Drive"),
]

print("=== Colab Setup Steps ===")
for s in steps:
    print(f"  [{s.step}]")
    print(f"    Code: {s.code}")
    print(f"    Time: {s.time}")
    print(f"    Note: {s.note}")

Prompt Engineering

# === Prompt Templates ===

@dataclass
class PromptTemplate:
    style: str
    positive: str
    negative: str
    cfg: float
    steps: int
    sampler: str

templates = [
    PromptTemplate("Anime Character",
        "masterpiece, best quality, 1girl, long silver hair, "
        "blue eyes, school uniform, cherry blossom, spring, "
        "detailed background, soft lighting, anime style",
        "worst quality, low quality, blurry, deformed, "
        "extra fingers, bad anatomy, watermark, text",
        7.5, 30, "DPM++ 2M Karras"),
    PromptTemplate("Photorealistic Portrait",
        "RAW photo, masterpiece, best quality, ultra detailed, "
        "1woman, 25yo, beautiful face, natural skin, "
        "studio lighting, shallow depth of field, 85mm lens, "
        "bokeh, photorealistic",
        "painting, drawing, anime, cartoon, deformed, "
        "airbrushed, overexposed, blurry",
        9.0, 40, "DPM++ SDE Karras"),
    PromptTemplate("Landscape Fantasy",
        "masterpiece, epic fantasy landscape, floating islands, "
        "waterfalls, magical forest, glowing crystals, "
        "dramatic sky, golden hour, volumetric lighting, "
        "matte painting style, ultra detailed",
        "low quality, blurry, text, watermark, "
        "modern buildings, cars, people",
        8.0, 35, "Euler a"),
    PromptTemplate("Product Photo",
        "professional product photography, white sneaker, "
        "clean white background, studio lighting, "
        "high resolution, commercial, minimalist, "
        "sharp focus, 8k uhd",
        "shadow, dirty, damaged, text, watermark, "
        "blurry, low quality",
        10.0, 40, "DDIM"),
]

print("=== Prompt Templates ===")
for t in templates:
    print(f"\n  [{t.style}]")
    print(f"    Positive: {t.positive[:80]}...")
    print(f"    Negative: {t.negative[:60]}...")
    print(f"    CFG: {t.cfg} | Steps: {t.steps} | Sampler: {t.sampler}")

Model & ControlNet

# === Models & Extensions ===

@dataclass
class SDModel:
    name: str
    type_: str
    size: str
    style: str
    source: str

models = [
    SDModel("SDXL Base 1.0",
        "Checkpoint (Official)",
        "6.5GB",
        "General Purpose High Quality 1024x1024",
        "Hugging Face stabilityai/stable-diffusion-xl-base-1.0"),
    SDModel("Realistic Vision V6",
        "Checkpoint (Community)",
        "2GB (SD 1.5)",
        "Photorealistic เหมือนจริงมาก",
        "CivitAI"),
    SDModel("Anything V5",
        "Checkpoint (Community)",
        "2GB (SD 1.5)",
        "Anime สวยงาม สีสดใส",
        "CivitAI"),
    SDModel("Detail Tweaker LoRA",
        "LoRA",
        "144MB",
        "เพิ่มรายละเอียด ใช้กับ Checkpoint ไหนัก็ได้",
        "CivitAI"),
    SDModel("ControlNet Canny",
        "ControlNet",
        "1.4GB",
        "ควบคุม Output ตามขอบภาพ Edge Detection",
        "Hugging Face lllyasviel/ControlNet"),
    SDModel("ControlNet OpenPose",
        "ControlNet",
        "1.4GB",
        "ควบคุม Pose ท่าทางคน",
        "Hugging Face lllyasviel/ControlNet"),
]

print("=== SD Models ===")
for m in models:
    print(f"  [{m.name}] Type: {m.type_} | Size: {m.size}")
    print(f"    Style: {m.style}")
    print(f"    Source: {m.source}")

เคล็ดลับ

• xformers: เปิด xformers ลด VRAM 30-40%
• float16: ใช้ torch_dtype=float16 เสมอ
• SDXL: ต้อง VRAM 15GB+ ใช้ Colab T4 ได้
• Prompt: เริ่มจาก Template แล้วปรับ ดีกว่าเขียนเอง
• Seed: ใช้ Seed เดิมสำหรับ Reproducible Result

Stable Diffusion คืออะไร

Open Source Text-to-Image AI Latent Diffusion SD 1.5 SDXL SD 3.0 FLUX img2img Inpainting ControlNet LoRA GPU Upscaling

สรุป

Stable Diffusion Colab Text-to-Image SDXL AUTOMATIC1111 ComfyUI Diffusers Prompt LoRA ControlNet GPU T4 Free Production