Stable Diffusion Architecture — U-Net + CLIP + VAE

from diffusers import StableDiffusionPipeline, DiffusionPipeline, DPMSolverMultistepScheduler
from diffusers.utils import load_image
import torch

# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# STABLE DIFFUSION XL (SDXL) -- highest quality open model
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True,
)
pipe = pipe.to("cuda")

# Speed optimization
pipe.enable_vae_slicing()    # Decode VAE in slices to save memory
pipe.enable_vae_tiling()     # Tile VAE for very large images
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)  # Torch compile: 25% speedup

# Replace default scheduler with DPM-Solver++ for 25-step high quality generation (vs 50 for DDIM)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

images = pipe(
    prompt="A photorealistic astronaut riding a horse on Mars, cinematic lighting, 8K, detailed",
    negative_prompt="blurry, low quality, distorted, watermark",  # what NOT to generate
    num_inference_steps=25,        # 25 steps with DPM-Solver++ ~ 50 with DDIM
    guidance_scale=7.5,            # CFG scale: 1=no guidance, 15=strong (may be oversaturated)
    width=1024, height=1024,       # SDXL native resolution
    num_images_per_prompt=4,       # generate 4 images in one batch
    generator=torch.Generator("cuda").manual_seed(42),  # reproducible seed
).images

# Classifier-Free Guidance (CFG) explained:
# guidance_scale = w.  Final noise prediction:
# noise_guided = noise_unconditional + w * (noise_conditional - noise_unconditional)
# w=1.0: ignore text prompt entirely
# w=7.5: balanced text following (recommended)
# w=15+: very literal, may lose realism

# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# THREE COMPONENTS OF STABLE DIFFUSION
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

sd_components = {
    "VAE (Variational Autoencoder)": {
        "role": "Compress images from pixel space to latent space",
        "dims": "512x512x3 pixels -> 64x64x4 latents (64x compression)",
        "why": "Diffusion in pixel space: too slow (needs 512x512x3=786K operations). Latent: 64x64x4=16K -- 50x faster!",
    },
    "CLIP Text Encoder": {
        "role": "Convert text prompt to embedding vector",
        "dims": "Text tokens -> 77 x 768 embedding",
        "why": "Provides conditioning signal. U-Net receives text representation via cross-attention",
    },
    "U-Net (Noise Predictor)": {
        "role": "Predict the noise at each denoising timestep",
        "dims": "Receives: [latent, timestep, text_embedding] -> predicted noise",
        "why": "Denoises in latent space, conditioned on text via cross-attention in each block",
        "architecture": "Encoder-decoder with skip connections + timestep embedding + cross-attention",
    },
}

for component, details in sd_components.items():
    print(f"\n{component}:")
    for k, v in details.items():
        print(f"  {k}: {v}")