Chuyển đến nội dung chính

Bài 5: Stable Diffusion Deep Dive — Kiến trúc & Pipeline

Latent Diffusion Models: tại sao làm việc trong latent space? UNet architecture. Text conditioning với CLIP. VAE encoder/decoder. Scheduler: DDIM, Euler, DPM++. Pipeline chi tiết từ prompt đến image.

🧠 AI & ML — Bài 4 Bài 5: Stable Diffusion Deep Dive — Kiến trúc & Pipeline

Generative AI: Tạo Hình ảnh & Video với AI

Phần 2: Diffusion Models — Cách mạng Tạo ảnh

xdev.asia

Giới thiệu

Stable Diffusion là Latent Diffusion Model (LDM) — thay vì diffusion trong pixel space (512×512×3), nó hoạt động trong latent space (64×64×4) nhỏ hơn 48 lần. Đây là breakthrough cho phép chạy diffusion trên consumer GPUs.


1. Kiến trúc Stable Diffusion

┌─────────────────────────────────────────────────────────┐
│              STABLE DIFFUSION PIPELINE                   │
│                                                         │
│  "a cat in space"                                       │
│        ↓                                                │
│  ┌──────────┐     ┌─────────────────────────┐          │
│  │   CLIP   │────→│    UNet + Scheduler      │          │
│  │  Text    │     │  (Denoise in latent)     │          │
│  │ Encoder  │     │  T steps: 20-50          │          │
│  └──────────┘     └───────────┬─────────────┘          │
│                               ↓                         │
│                       ┌──────────────┐                  │
│  Noise z (64x64x4) → │  VAE Decoder  │ → Image 512x512│
│                       └──────────────┘                  │
└─────────────────────────────────────────────────────────┘

Components chi tiết

from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")

# 4 components chính:
# 1. Text Encoder (CLIP): text → embeddings
pipe.text_encoder  # CLIPTextModel

# 2. UNet: noise predictor, conditioned on text
pipe.unet  # UNet2DConditionModel

# 3. VAE: latent ↔ pixel space
pipe.vae  # AutoencoderKL

# 4. Scheduler: noise schedule algorithm
pipe.scheduler  # EulerDiscreteScheduler

2. Pipeline Step-by-Step

import torch
from diffusers import AutoencoderKL, UNet2DConditionModel, EulerDiscreteScheduler
from transformers import CLIPTextModel, CLIPTokenizer

# Step 1: Tokenize & encode text
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14")

prompt = "a cat wearing sunglasses, digital art"
tokens = tokenizer(prompt, return_tensors="pt", padding="max_length", max_length=77)
text_embeddings = text_encoder(tokens.input_ids)[0]  # [1, 77, 768]

# Step 2: Initialize random latent
latent = torch.randn(1, 4, 64, 64)  # latent space

# Step 3: Denoise loop
scheduler = EulerDiscreteScheduler(num_train_timesteps=1000)
scheduler.set_timesteps(30)  # 30 denoising steps

for t in scheduler.timesteps:
    # Predict noise conditioned on text
    noise_pred = unet(latent, t, encoder_hidden_states=text_embeddings).sample

    # Classifier-free guidance
    noise_uncond = unet(latent, t, encoder_hidden_states=uncond_embeddings).sample
    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)

    # Update latent
    latent = scheduler.step(noise_pred, t, latent).prev_sample

# Step 4: Decode latent → pixel image
image = vae.decode(latent / 0.18215).sample

3. Classifier-Free Guidance (CFG)

$$\hat{\epsilon} = \epsilon_{uncond} + s \cdot (\epsilon_{cond} - \epsilon_{uncond})$$

# guidance_scale s controls text adherence
# s = 1: no guidance (ignore prompt)
# s = 7-8: balanced (default)
# s = 15-20: strong adherence (can be over-saturated)

guidance_scale = 7.5

# During inference: run UNet twice
noise_cond = unet(latent, t, text_embeddings)    # conditioned
noise_uncond = unet(latent, t, empty_embeddings)  # unconditioned

# Interpolate
noise_pred = noise_uncond + guidance_scale * (noise_cond - noise_uncond)

4. VAE — Latent Space Compression

# Encode: 512x512x3 → 64x64x4 (compression ratio ~48x)
with torch.no_grad():
    latent = vae.encode(image).latent_dist.sample()
    latent = latent * 0.18215  # scaling factor

# Decode: 64x64x4 → 512x512x3
with torch.no_grad():
    image = vae.decode(latent / 0.18215).sample

Tại sao latent space?

  • Memory: 512×512×3 = 786K pixels → 64×64×4 = 16K values
  • Speed: UNet xử lý tensor nhỏ hơn 48x
  • Quality: VAE đã học compress thông minh

5. Schedulers — Thuật toán Denoising

from diffusers import (
    DDPMScheduler,        # Original, 1000 steps
    DDIMScheduler,        # Deterministic, 50 steps
    EulerDiscreteScheduler,       # Fast, 20-30 steps
    DPMSolverMultistepScheduler,  # DPM++, 20 steps, high quality
    UniPCMultistepScheduler,      # 10-20 steps
)

# Đổi scheduler dễ dàng
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
SchedulerStepsSpeedQualityUse case
DDPM1000Very slowReferenceTraining
DDIM50MediumGoodGeneral
Euler20-30FastGreatDefault SDXL
DPM++ 2M20-25FastExcellentRecommended
UniPC10-15Very fastGoodReal-time

6. Stable Diffusion Versions

VersionResolutionText EncoderReleased
SD 1.5512×512CLIP ViT-L/142022
SD 2.1768×768OpenCLIP ViT-H2022
SDXL1024×1024CLIP ViT-L + OpenCLIP ViT-bigG2023
SD31024×1024Triple text encoder (CLIP×2 + T5)2024
Flux1024×1024T5 XXL2024
# SDXL — recommended cho production
pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
)
pipe.to("cuda")

image = pipe(
    prompt="a majestic lion in a forest, photorealistic",
    negative_prompt="blurry, low quality, distorted",
    num_inference_steps=30,
    guidance_scale=7.5,
    width=1024,
    height=1024,
).images[0]

7. Negative Prompts & Parameters

# Negative prompt: things to avoid
negative_prompt = "blurry, low quality, deformed, ugly, bad anatomy"

# Key parameters
image = pipe(
    prompt="...",
    negative_prompt=negative_prompt,
    num_inference_steps=30,    # More = better quality, slower
    guidance_scale=7.5,        # Text adherence (7-9 optimal)
    width=1024,
    height=1024,
    seed=42,                   # Reproducibility
).images[0]

Tổng kết

ComponentVai trò
CLIP Text EncoderText → embeddings (semantic meaning)
UNetNoise predictor, conditioned on text
VAECompress pixel ↔ latent space
SchedulerAlgorithm cho denoising steps
CFGGuidance scale điều chỉnh text adherence

📌 Bài tiếp theo: Prompt Engineering cho Image Generation — kỹ thuật viết prompt hiệu quả.