xsl/change_hair
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: 新增mask区域重绘功能 + 羽化贴回优化 + 发丝增强

Browse Source 
新增功能：
- inpaint_mask.py: mask区域重绘服务（enhance/pure_inpaint两种模式）
  严格只在mask区做SD inpainting，发丝自然化
- /inpaint 页面 + /api/inpaint 接口：画板手绘mask + 提示词编辑 + denoising可调
- enhance_hair.webui_img2img: 新增denoising_strength参数（原硬编码0.35）

羽化贴回优化（hair_swap_manual.py 步骤⑥）：
- feather_px（羽化范围）+ feather_alpha（羽化强度）拆分为两个独立参数
- 改为只羽化边缘（mask内部保持硬切，仅边缘带渐变）
- 新增黑底羽化对比图（纯黑背景凸显边缘过渡）

其他改动：
- OSS/COS密钥脱敏：改为可选import，未设环境变量时不崩溃
- manual页面：发型列表改为本次训练的8个（含hair_flow）
- enhance二次增强：重绘区改为「原图头发∪手绘mask」并外扩10%+减发际线带3cm
- 发型清单文档更新：179→180个可用发型（+hair_flow）

涉及文件：
- 新增: inpaint_mask.py, inpaint.html
- 修改: app.py, manual.html, hair_swap_manual.py, enhance_hair.py
- 修改: oss_module.py, upload_oss.py, cos_module.py（密钥脱敏）
- 文档: HAIRSTYLES_AVAILABLE.md, hairstyles_available.csv
This commit is contained in:
xsl
2026-07-08 20:38:31 +08:00
parent aa38085867
commit 215663177a
10 changed files with 751 additions and 229 deletions
Download Patch File Download Diff File Expand all files Collapse all files
project/hair_service_sd/hair_swap_manual.py
+46 -12
View File
@@ -36,7 +36,8 @@ def hair_swap_manual(origin_img, hand_mask, hair_id, hairstyle_process, landmark
denoising_strength=0.6, # 重绘强度（唯一能透传到webui的SD参数）
# ===== 贴回/融合（真正生效）=====
blend_dilate=(5, 5), # 步骤⑥ strict mask贴回时mask膨胀核
feather_px=0, # ★ strict贴回边缘羽化像素（0=无羽化硬边缘，>0=高斯模糊边缘）
feather_px=0, # ★ strict贴回边缘羽化范围（像素，边缘带宽度，0=无羽化）
feather_alpha=1.0, # ★ 羽化强度（0~1，0=硬切，1=最大羽化，控制边缘过渡柔和度）
seamless_dilate=(9, 9), # 步骤⑦ 泊松融合mask膨胀核
# ===== enhance 二次增强重绘 =====
enhance=False, # 步骤⑧ 是否对结果再做一次低强度SD重绘（让发丝更清晰）
@@ -230,10 +231,20 @@ def hair_swap_manual(origin_img, hand_mask, hair_id, hairstyle_process, landmark
result_strict_hard[mask_back == 0] = origin_img[mask_back == 0]
if feather_px and feather_px > 0:
# ★ 只羽化边缘：mask 内部保持硬切(255)，只有边缘带做渐变
# 1. 对 mask_back 做高斯模糊（得到边缘的渐变）
k = max(1, int(feather_px)) * 2 + 1 # 高斯核必须是奇数
mask_back_blur = cv2.GaussianBlur(mask_back, (k, k), 0)
mask_blend = (mask_back_blur.astype(np.float32) / 255)[..., None]
feather_info = f"，边缘羽化{feather_px}px"
mask_blur = cv2.GaussianBlur(mask_back, (k, k), 0)
# 2. 取 mask 内部区域（离边缘 > feather_px 的部分），强制为255（不羽化）
erode_kernel = np.ones((k, k), np.uint8)
mask_inner = cv2.erode(mask_back, erode_kernel)
# 3. 合成：内部硬切(255) + 边缘用模糊值
mask_blend_mask = np.maximum(mask_inner, mask_blur).astype(np.float32)
# 4. ★ feather_alpha 控制羽化强度：在「全硬切(原mask)」和「边缘渐变(mask_blend_mask)」之间插值
# alpha=0 → 完全硬切（无羽化），alpha=1 → 最大羽化
mask_blend_mask = mask_back.astype(np.float32) * (1 - feather_alpha) + mask_blend_mask * feather_alpha
mask_blend = (mask_blend_mask / 255)[..., None]
feather_info = f"，边缘羽化{feather_px}px 强度{feather_alpha:.2f}"
else:
mask_blend = hard_blend
result_strict = (sd_result_back.astype(np.float32) * mask_blend +
@@ -241,17 +252,38 @@ def hair_swap_manual(origin_img, hand_mask, hair_id, hairstyle_process, landmark
result_strict = np.clip(result_strict, 0, 255).astype(np.uint8)
result_strict[mask_back == 0] = origin_img[mask_back == 0]
# ★ 羽化对比图：左=硬边缘(feather=0) 右=当前羽化，裁剪mask边界附近放大
# ★ 羽化对比图（黑底版）：纯黑背景 + 贴回图，凸显羽化边缘过渡
# 用黑色背景而非原图，避免原图干扰，纯粹对比羽化效果
feather_compare = None
feather_compare_full = None
if feather_px and feather_px > 0:
# 计算 mask_blend（用于把贴回图叠加到黑底上）
ys, xs = np.where(mask_back > 10)
if len(ys) > 0:
x1 = max(0, xs.min()-20); x2 = min(origin_img.shape[1], xs.max()+20)
y1 = max(0, ys.min()-20); y2 = min(origin_img.shape[0], ys.max()+20)
hard_crop = result_strict_hard[y1:y2, x1:x2]
blur_crop = result_strict[y1:y2, x1:x2]
sep = np.full((hard_crop.shape[0], 3, 3), 128, dtype=np.uint8)
# 黑底版：贴回图 * mask_blend + 黑色(0) * (1-mask_blend)
# result_strict_hard 已是「贴回图+原图」混合，这里重新算黑底版
black_bg = np.zeros_like(origin_img)
# 无羽化版（黑底）
hard_on_black = (sd_result_back.astype(np.float32) * hard_blend +
black_bg.astype(np.float32) * (1 - hard_blend))
hard_on_black = np.clip(hard_on_black, 0, 255).astype(np.uint8)
hard_on_black[mask_back == 0] = 0 # mask外纯黑
# 有羽化版（黑底）
blur_on_black = (sd_result_back.astype(np.float32) * mask_blend +
black_bg.astype(np.float32) * (1 - mask_blend))
blur_on_black = np.clip(blur_on_black, 0, 255).astype(np.uint8)
blur_on_black[mask_back == 0] = 0 # mask外纯黑
# 局部放大版（裁剪mask边界附近）
x1 = max(0, xs.min()-30); x2 = min(origin_img.shape[1], xs.max()+30)
y1 = max(0, ys.min()-30); y2 = min(origin_img.shape[0], ys.max()+30)
hard_crop = hard_on_black[y1:y2, x1:x2]
blur_crop = blur_on_black[y1:y2, x1:x2]
sep = np.full((hard_crop.shape[0], 3, 3), 60, dtype=np.uint8) # 深灰分隔线
feather_compare = np.hstack([hard_crop, sep, blur_crop])
# 完整图版（不裁剪）
sep_full = np.full((origin_img.shape[0], 3, 3), 60, dtype=np.uint8)
feather_compare_full = np.hstack([hard_on_black, sep_full, blur_on_black])
result = result_strict if strict_mask else result_full
@@ -262,7 +294,9 @@ def hair_swap_manual(origin_img, hand_mask, hair_id, hairstyle_process, landmark
{"label": f"当前选用({'严格mask' if strict_mask else '整框覆盖'})", "b64": _enc(result)},
]
if feather_compare is not None:
step6_images.append({"label": f"★ 羽化对比(左=硬边 右=羽化{feather_px}px)", "b64": _enc(feather_compare)})
step6_images.append({"label": f"★ 黑底羽化对比·边缘(左=无羽化 右=羽化{feather_px}pxα{feather_alpha:.1f})", "b64": _enc(feather_compare)})
if feather_compare_full is not None:
step6_images.append({"label": f"★ 黑底羽化对比·完整(左=无羽化 右=羽化{feather_px}pxα{feather_alpha:.1f})", "b64": _enc(feather_compare_full)})
steps.append({
"title": "⑥ 贴回原图",
"desc": f"把SD结果逆warpAffine贴回用户原图。可调：strict_mask={strict_mask}, blend_dilate={bd}{feather_info}。整框覆盖=整个裁剪框覆盖原图；严格mask=只在mask区域融合，mask外保留原图",
@@ -390,7 +424,7 @@ def hair_swap_manual(origin_img, hand_mask, hair_id, hairstyle_process, landmark
"strict_mask": strict_mask, "seamless_blend": seamless_blend,
"is_hr": is_hr, "dilate_kernel": list(dilate_kernel),
"denoising_strength": denoising_strength,
"blend_dilate": list(blend_dilate), "feather_px": feather_px,
"blend_dilate": list(blend_dilate), "feather_px": feather_px, "feather_alpha": feather_alpha,
"seamless_dilate": list(seamless_dilate),
"enhance": enhance, "enhance_denoising": enhance_denoising,
"mask_area_ratio_pct": round(mask_area_ratio, 1),