Fooocus Outpaint 完整处理流程

1. Canvas Expansion & Mask Creation 画布扩展和遮罩创建

12:15:modules/async_worker.py
展开代码
def apply_outpaint(async_task, inpaint_image, inpaint_mask):
    if len(async_task.outpaint_selections) > 0:
        H, W, C = inpaint_image.shape
        if 'top' in async_task.outpaint_selections:
            inpaint_image = np.pad(inpaint_image, [[int(H * 0.3), 0], [0, 0], [0, 0]], mode='edge')
            inpaint_mask = np.pad(inpaint_mask, [[int(H * 0.3), 0], [0, 0]], mode='constant', constant_values=255)

• 根据用户选择的方向（top/bottom/left/right）扩展画布30%
• 使用edge模式复制边缘像素填充新区域
• 新区域的mask值设为255（完全需要重绘）

2. Content-Aware Filling 内容感知填充

158:170:modules/inpaint_worker.py
展开代码
# compute filling
if use_fill:
    self.interested_fill = fooocus_fill(self.interested_image, self.interested_mask)

在fooocus_fill函数中：

136:147:modules/inpaint_worker.py
展开代码
def fooocus_fill(image, mask):
    height, width, _ = image.shape
    mask = mask[:, :, 0]
    raw_image = image.copy()
    raw_mask = mask.copy()

    # 多尺度模糊填充算法
    for r in [11, 7, 5, 3]:
        kernel = np.ones((r, r), dtype=np.uint8)
        dilated_mask = cv2.dilate(mask, kernel, iterations=1)
        blurred_image = box_blur(image, r)
        image = np.where(dilated_mask[..., None] > 128, blurred_image, image)
        mask = dilated_mask

3. VAE Inpaint Encoding VAE修复编码

177:189:modules/core.py
展开代码
def encode_vae_inpaint(vae, pixels, mask):
    # 预处理mask
    w = mask.round()[..., None]
    pixels = pixels * (1 - w) + 0.5 * w  # 用0.5替换masked区域
    
    # VAE编码
    latent = vae.encode(pixels)
    B, C, H, W = latent.shape
    
    # 处理latent mask
    latent_mask = mask[:, None, :, :]
    latent_mask = torch.nn.functional.interpolate(latent_mask, size=(H * 8, W * 8), mode="bilinear").round()
    latent_mask = torch.nn.functional.max_pool2d(latent_mask, (8, 8)).round().to(latent)

4. Inpaint Head Injection 修复头部注入

194:218:modules/inpaint_worker.py
展开代码
def patch(self, inpaint_head_model_path, inpaint_latent, inpaint_latent_mask, model):
    # 加载inpaint head模型
    if inpaint_head_model is None:
        inpaint_head_model = InpaintHead()
        sd = torch.load(inpaint_head_model_path, map_location='cpu', weights_only=True)
        inpaint_head_model.load_state_dict(sd)
    
    # 特征拼接和处理
    feed = torch.cat([
        inpaint_latent_mask,
        model.model.process_latent_in(inpaint_latent)
    ], dim=1)
    
    inpaint_head_feature = inpaint_head_model(feed)

5. Diffusion Sampling (Add Noise + Denoise) 扩散采样（加噪+去噪）

280:342:modules/core.py
展开代码
def ksampler(model, positive, negative, latent, seed=None, steps=30, cfg=7.0, 
             sampler_name='dpmpp_2m_sde_gpu', scheduler='karras', denoise=1.0, ...):
    
    latent_image = latent["samples"]
    
    # 准备噪声（加噪过程）
    if disable_noise:
        noise = torch.zeros(latent_image.size(), ...)
    else:
        noise = ldm_patched.modules.sample.prepare_noise(latent_image, seed, batch_inds)
    
    # 扩散采样过程（去噪过程）
    samples = ldm_patched.modules.sample.sample(model,
                                                noise, steps, cfg, sampler_name, scheduler,
                                                positive, negative, latent_image,
                                                denoise=denoise, ...)

6. Post-processing 后处理

254:261:modules/inpaint_worker.py
展开代码
def post_process(self, img):
    a, b, c, d = self.interested_area
    content = resample_image(img, d - c, b - a)
    result = self.image.copy()
    result[a:b, c:d] = content
    result = self.color_correction(result)
    return result

总结流程：

1. 扩展画布 → 创建新区域和mask
2. 填充mask → 使用content-aware filling算法
3. VAE编码 → 将图像和mask编码到潜在空间
4. 加噪 → 准备噪声张量
5. 去噪 → 通过扩散采样生成新内容
6. 后处理 → 调整尺寸和颜色校正