第 12 章 · 真实图片扩散加噪代码实验¶

# 真实图片扩散：按时间步逐渐加入噪声。
ddpm_photo = Image.fromarray(load_sample_image("flower.jpg")).resize((128, 128))
ddpm_image = np.asarray(ddpm_photo).astype("float32") / 255.0
sample = torch.tensor(ddpm_image).permute(2, 0, 1).unsqueeze(0) * 2 - 1
torch.manual_seed(12)
noise = torch.randn(sample.shape)
scheduler = DDPMScheduler(num_train_timesteps=1000)
timesteps = [0, 50, 150, 300, 500]
ddpm_images = []
diff_rows = []

for t in timesteps:
    timestep = torch.tensor([t], dtype=torch.long)
    noisy = scheduler.add_noise(sample, noise, timestep)
    image = ((noisy[0].permute(1, 2, 0).numpy() + 1) / 2).clip(0, 1)
    ddpm_images.append(image)
    alpha_bar = float(scheduler.alphas_cumprod[t])
    diff_rows.append({
        "时间步": t,
        "累计保留系数": alpha_bar,
        "原图权重": np.sqrt(alpha_bar),
        "噪声权重": np.sqrt(1 - alpha_bar),
        "像素标准差": float(image.std()),
    })

diff_1d_df = pd.DataFrame(diff_rows)
display(diff_1d_df.round(4))

# 绘制真实图片在不同时间步下的前向扩散效果。
fig, axes = plt.subplots(2, 3, figsize=(9.0, 6.0))
for ax, image, timestep in zip(axes.ravel(), ddpm_images, timesteps):
    ax.imshow(image)
    ax.set_title(f"时间步 {timestep}", fontweight="bold")
    ax.set_xticks([])
    ax.set_yticks([])
for ax in axes.ravel()[len(ddpm_images):]:
    ax.axis("off")
fig.suptitle("真实图片前向加噪", x=0.08, ha="left", fontsize=14, fontweight="bold", color="#0f172a")
plt.tight_layout()
plt.show()