feat(swapHair): webui_steps参数可外部传入(接口2调试页可调)

- gen_super_image.py: webui_img2img + build_body_v2 接收可选 steps 参数
- run_copy_cost_colorb64.py: swapHair 接口接收 webui_steps，透传给 webui_img2img
- 未传时保持环境变量 WEBUI_STEPS 默认(15)，向后兼容

README.md

@@ -296,8 +296,8 @@ curl -X POST http://127.0.0.1:32678/api/hair/train \
 | 改动 | 说明 |
 |------|------|
 | `config.json` 的 onediff compiler 路径 | 指向本机 onediff 目录 |
-| 启动加 `HF_HUB_OFFLINE=1` | 本机无法访问 huggingface.co，用本地缓存离线加载 CLIP |
+| 启动加 `HF_HUB_OFFLINE=1` | 本机无法稳定访问 huggingface.co，用本地缓存离线加载 CLIP（需预先缓存 `openai/clip-vit-large-patch14`） |
-| SD 模型 | 实际加载 `v1-5-pruned-emaonly`（config 里写的 majicmix 不存在，webui 自动 fallback） |
+| SD 模型 | `sd_model_checkpoint` 固定为 `v1-5-pruned-emaonly.safetensors`（本机无 majicmix）。仓库内权威副本：`configs/stable-diffusion-webui.config.json`；部署时拷到 `project/onediff/stable-diffusion-webui/config.json`（该目录是独立 git 仓库，无法被本仓直接跟踪） |
 
 ### 3. 换发型推理改动（`hair_service_sd/gen_super_image.py`）
 

change_hair-hair.service

@@ -0,0 +1,19 @@
+[Unit]
+Description=change_hair hair_service_sd swapHair (8801)
+After=network-online.target change_hair-webui.service change_hair-photo.service
+Wants=change_hair-webui.service change_hair-photo.service
+
+[Service]
+Type=simple
+User=ubuntu
+WorkingDirectory=/home/ubuntu/change_hair/project/hair_service_sd
+Environment=CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
+Environment=CUDA_VISIBLE_DEVICES=0
+Environment=APP_WORKER_ID=1
+ExecStart=/home/ubuntu/miniconda3/envs/my_hair/bin/python run_copy_cost_colorb64.py
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=300
+
+[Install]
+WantedBy=multi-user.target

change_hair-photo.service

@@ -0,0 +1,17 @@
+[Unit]
+Description=change_hair photo_service LoRA scheduler (32678)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=ubuntu
+WorkingDirectory=/home/ubuntu/change_hair/project/photo_service
+Environment=CUDA_VISIBLE_DEVICES=0
+Environment=APP_WORKER_ID=1
+ExecStart=/home/ubuntu/miniconda3/envs/py310/bin/python -u lora_train_service_1.py
+Restart=on-failure
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target

change_hair-webui.service

@@ -0,0 +1,20 @@
+[Unit]
+Description=change_hair SD WebUI (57860)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=ubuntu
+WorkingDirectory=/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui
+Environment=CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
+Environment=CUDA_VISIBLE_DEVICES=0
+Environment=HF_HUB_OFFLINE=1
+Environment=TRANSFORMERS_OFFLINE=1
+ExecStart=/home/ubuntu/miniconda3/envs/sdwebui/bin/python webui.py --api --listen --xformers --port 57860
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=600
+
+[Install]
+WantedBy=multi-user.target

configs/stable-diffusion-webui.config.json

@@ -0,0 +1,40 @@
+{
+    "ldsr_steps": 100,
+    "ldsr_cached": false,
+    "SCUNET_tile": 256,
+    "SCUNET_tile_overlap": 8,
+    "SWIN_tile": 192,
+    "SWIN_tile_overlap": 8,
+    "SWIN_torch_compile": false,
+    "hypertile_enable_unet": false,
+    "hypertile_enable_unet_secondpass": false,
+    "hypertile_max_depth_unet": 3,
+    "hypertile_max_tile_unet": 256,
+    "hypertile_swap_size_unet": 3,
+    "hypertile_enable_vae": false,
+    "hypertile_max_depth_vae": 3,
+    "hypertile_max_tile_vae": 128,
+    "hypertile_swap_size_vae": 3,
+    "onediff_compiler_caches_path": "/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/extensions/onediff_sd_webui_extensions/compiler_caches",
+    "onediff_compiler_backend": "oneflow",
+    "sd_model_checkpoint": "v1-5-pruned-emaonly.safetensors",
+    "sd_checkpoint_hash": "6ce0161689b3853acaa03779ec93eafe75a02f4ced659bee03f50797806fa2fa",
+    "control_net_detectedmap_dir": "detected_maps",
+    "control_net_models_path": "",
+    "control_net_modules_path": "",
+    "control_net_unit_count": 3,
+    "control_net_model_cache_size": 2,
+    "control_net_inpaint_blur_sigma": 7,
+    "control_net_no_detectmap": false,
+    "control_net_detectmap_autosaving": false,
+    "control_net_allow_script_control": false,
+    "control_net_sync_field_args": true,
+    "controlnet_show_batch_images_in_ui": false,
+    "controlnet_increment_seed_during_batch": false,
+    "controlnet_disable_openpose_edit": false,
+    "controlnet_disable_photopea_edit": false,
+    "controlnet_photopea_warning": true,
+    "controlnet_ignore_noninpaint_mask": false,
+    "controlnet_clip_detector_on_cpu": false,
+    "controlnet_control_type_dropdown": false
+}

docs/方案A-上传文件清单.md

@@ -2,7 +2,7 @@
 
 > 适用场景：从原服务器打包上传 conda 环境 + SD WebUI 代码，在本机完成换发型推理部署。
 >
-> 本机路径：`/home/ubuntu/change_hair`（GPU：RTX 5090 32G，CUDA 12.8）
+> 本机路径：`/home/xsl/change_hair`（GPU：RTX 5090 32G，CUDA 12.8）
 >
 > 原服务器参考：`szlc@192.168.101.63`
 >
@@ -48,13 +48,13 @@
 
 ### 2.1 Conda 环境 × 3
 
-从原服务器 `/home/szlc/miniconda3/envs/` 打包，上传到本机 `/home/ubuntu/miniconda3/envs/`。
+从原服务器 `/home/szlc/miniconda3/envs/` 打包，上传到本机 `/home/xsl/miniconda3/envs/`。
 
 | 环境名 | 原服务器路径 | 本机目标路径 | 用途 | 预估大小 |
 |--------|-------------|-------------|------|---------|
-| `my_hair` | `/home/szlc/miniconda3/envs/my_hair/` | `/home/ubuntu/miniconda3/envs/my_hair/` | hair_service_sd 主服务（端口 8801） | ~5–8G |
+| `my_hair` | `/home/szlc/miniconda3/envs/my_hair/` | `/home/xsl/miniconda3/envs/my_hair/` | hair_service_sd 主服务（端口 8801） | ~5–8G |
-| `sdwebui` | `/home/szlc/miniconda3/envs/sdwebui/` | `/home/ubuntu/miniconda3/envs/sdwebui/` | SD WebUI 推理（端口 57860） | ~8–12G |
+| `sdwebui` | `/home/szlc/miniconda3/envs/sdwebui/` | `/home/xsl/miniconda3/envs/sdwebui/` | SD WebUI 推理（端口 57860） | ~8–12G |
-| `py310` | `/home/szlc/miniconda3/envs/py310/` | `/home/ubuntu/miniconda3/envs/py310/` | photo_service LoRA 调度（端口 32678） | ~2–4G |
+| `py310` | `/home/szlc/miniconda3/envs/py310/` | `/home/xsl/miniconda3/envs/py310/` | photo_service LoRA 调度（端口 32678） | ~2–4G |
 
 > ⚠️ **为什么必须上传，不能从公网安装？** 这三个环境包含特定版本的 PyTorch + CUDA 扩展、历史版本的 pip 包、以及编译好的 C++/CUDA 算子，依赖关系复杂。项目中的 `env.yaml` 是旧的 Python 3.7 规格，与实际运行环境（Python 3.10）不一致，无法用于重建。
 
@@ -72,8 +72,8 @@ tar czf py310.tar.gz     py310/
 
 ```bash
 # 在本机执行（上传完成后）
-mkdir -p /home/ubuntu/miniconda3/envs
+mkdir -p /home/xsl/miniconda3/envs
-cd /home/ubuntu/miniconda3/envs
+cd /home/xsl/miniconda3/envs
 tar xzf /path/to/my_hair.tar.gz
 tar xzf /path/to/sdwebui.tar.gz
 tar xzf /path/to/py310.tar.gz
@@ -87,7 +87,7 @@ tar xzf /path/to/py310.tar.gz
 
 | 原服务器路径 | 本机目标路径 | 说明 | 预估大小 |
 |-------------|-------------|------|---------|
-| `/home/szlc/project/onediff/stable-diffusion-webui/`（排除 `models/`） | `/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/` | webui.py、modules/、extensions/、repositories/ 等 | ~2–5G |
+| `/home/szlc/project/onediff/stable-diffusion-webui/`（排除 `models/`） | `/home/xsl/change_hair/project/onediff/stable-diffusion-webui/` | webui.py、modules/、extensions/、repositories/ 等 | ~2–5G |
 
 > ⚠️ 本机当前 `project/onediff/stable-diffusion-webui/` 目录只有 `models/` 子目录（4.6G 模型权重），缺少 `webui.py`、`modules/`、`extensions/` 等所有代码文件。
 
@@ -131,7 +131,7 @@ tar czf webui_code.tar.gz \
 
 ```bash
 # 在本机执行（保留已有 models 目录）
-cd /home/ubuntu/change_hair/project/onediff
+cd /home/xsl/change_hair/project/onediff
 tar xzf /path/to/webui_code.tar.gz
 # 解压后确认 models/ 仍在且未被覆盖
 ls -lh stable-diffusion-webui/models/Stable-diffusion/
@@ -147,13 +147,13 @@ ls -lh stable-diffusion-webui/models/Stable-diffusion/
 # ===== 必须上传（方案 A，共 4 项，约 18–30G）=====
 
 # 1. Conda 环境（3 个）
-/home/szlc/miniconda3/envs/my_hair/       → /home/ubuntu/miniconda3/envs/my_hair/
+/home/szlc/miniconda3/envs/my_hair/       → /home/xsl/miniconda3/envs/my_hair/
-/home/szlc/miniconda3/envs/sdwebui/       → /home/ubuntu/miniconda3/envs/sdwebui/
+/home/szlc/miniconda3/envs/sdwebui/       → /home/xsl/miniconda3/envs/sdwebui/
-/home/szlc/miniconda3/envs/py310/         → /home/ubuntu/miniconda3/envs/py310/
+/home/szlc/miniconda3/envs/py310/         → /home/xsl/miniconda3/envs/py310/
 
 # 2. SD WebUI 代码（排除 models/）
 /home/szlc/project/onediff/stable-diffusion-webui/（排除 models/）
-    → /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/
+    → /home/xsl/change_hair/project/onediff/stable-diffusion-webui/
 ```
 
 ---
@@ -166,16 +166,16 @@ ls -lh stable-diffusion-webui/models/Stable-diffusion/
 
 | 原服务器路径 | 本机目标路径 | 预估大小 |
 |-------------|-------------|---------|
-| `/home/szlc/project/kohya_ss_home/` | `/home/ubuntu/change_hair/project/kohya_ss_home/` | ~13G |
+| `/home/szlc/project/kohya_ss_home/` | `/home/xsl/change_hair/project/kohya_ss_home/` | ~13G |
-| `/home/szlc/miniconda3/envs/kohya/` | `/home/ubuntu/miniconda3/envs/kohya/` | ~5G |
+| `/home/szlc/miniconda3/envs/kohya/` | `/home/xsl/miniconda3/envs/kohya/` | ~5G |
 
 ### 4.2 换发色功能
 
 | 原服务器路径 | 本机目标路径 |
 |-------------|-------------|
-| `/home/szlc/project/data/ref_haircolor/` | `/home/ubuntu/change_hair/project/data/ref_haircolor/` |
+| `/home/szlc/project/data/ref_haircolor/` | `/home/xsl/change_hair/project/data/ref_haircolor/` |
-| `/home/szlc/project/data/ref_color/` | `/home/ubuntu/change_hair/project/data/ref_color/` |
+| `/home/szlc/project/data/ref_color/` | `/home/xsl/change_hair/project/data/ref_color/` |
-| `/home/szlc/project/data/ref_color_imgs/` | `/home/ubuntu/change_hair/project/data/ref_color_imgs/` |
+| `/home/szlc/project/data/ref_color_imgs/` | `/home/xsl/change_hair/project/data/ref_color_imgs/` |
 | `/home/szlc/.../weights/classify_3_2499.pth` | `project/hair_service_sd/weights/classify_3_2499.pth` |
 
 ### 4.3 更多发型
@@ -207,25 +207,25 @@ ls -lh stable-diffusion-webui/models/Stable-diffusion/
 # 从原服务器直传到本机（在本机执行）
 rsync -avP --partial \
   szlc@192.168.101.63:/home/szlc/miniconda3/envs/my_hair/ \
-  /home/ubuntu/miniconda3/envs/my_hair/
+  /home/xsl/miniconda3/envs/my_hair/
 
 rsync -avP --partial \
   szlc@192.168.101.63:/home/szlc/miniconda3/envs/sdwebui/ \
-  /home/ubuntu/miniconda3/envs/sdwebui/
+  /home/xsl/miniconda3/envs/sdwebui/
 
 rsync -avP --partial \
   szlc@192.168.101.63:/home/szlc/miniconda3/envs/py310/ \
-  /home/ubuntu/miniconda3/envs/py310/
+  /home/xsl/miniconda3/envs/py310/
 
 rsync -avP --partial \
   --exclude='models/' \
   szlc@192.168.101.63:/home/szlc/project/onediff/stable-diffusion-webui/ \
-  /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/
+  /home/xsl/change_hair/project/onediff/stable-diffusion-webui/
 ```
 
-### 方式 2：先 tar 再上传到 `/home/ubuntu/data`
+### 方式 2：先 tar 再上传到 `/home/xsl/data`
 
-若通过网盘中转，先 tar 再上传到 `/home/ubuntu/data/`，上传完成后告知我解压。
+若通过网盘中转，先 tar 再上传到 `/home/xsl/data/`，上传完成后告知我解压。
 
 ---
 
@@ -235,15 +235,15 @@ rsync -avP --partial \
 
 ```bash
 # 1. Conda 环境
-ls /home/ubuntu/miniconda3/envs/my_hair/bin/python
+ls /home/xsl/miniconda3/envs/my_hair/bin/python
-ls /home/ubuntu/miniconda3/envs/sdwebui/bin/python
+ls /home/xsl/miniconda3/envs/sdwebui/bin/python
-ls /home/ubuntu/miniconda3/envs/py310/bin/python
+ls /home/xsl/miniconda3/envs/py310/bin/python
 
 # 2. WebUI 代码
-ls /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/webui.py
+ls /home/xsl/change_hair/project/onediff/stable-diffusion-webui/webui.py
 
 # 3. 模型未被覆盖
-ls -lh /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors
+ls -lh /home/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors
 ```
 
 期望结果：4 项均存在，SD 底模约 4.0G。
@@ -254,7 +254,7 @@ ls -lh /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/St
 
 你上传完毕并告知后，我会依次执行：
 
-1. **路径适配**：运行 `adapt_paths.sh`，将所有硬编码路径从 `/home/xsl/change_hair` 改为 `/home/ubuntu/change_hair`，`/home/xsl/miniconda3` 改为 `/home/ubuntu/miniconda3`
+1. **路径适配**：运行 `adapt_paths.sh`，将所有硬编码路径从 `/home/xsl/change_hair` 改为 `/home/xsl/change_hair`，`/home/xsl/miniconda3` 改为 `/home/xsl/miniconda3`
 2. **Conda 路径修复**：运行 `fix_conda_paths.py`，修复 3 个环境中二进制文件的 shebang 和内部路径前缀
 3. **安装 Miniconda 基础环境**（如尚未安装）：从 `repo.anaconda.com` 下载安装
 4. **启动服务**：`bash start_all.sh start`（依赖顺序：webui → photo_service → hair_service_sd）
@@ -283,11 +283,11 @@ POST http://127.0.0.1:8801/api/swapHair/v1
 
 **Q：能否只上传 miniconda3 整包，不传单个 env？**
 
-可以。上传 `/home/szlc/miniconda3/` 整包到 `/home/ubuntu/miniconda3/` 也行，体积更大（约 20–30G）但更省事。
+可以。上传 `/home/szlc/miniconda3/` 整包到 `/home/xsl/miniconda3/` 也行，体积更大（约 20–30G）但更省事。
 
 **Q：webui 上传时不小心覆盖了 models 怎么办？**
 
-重新从 `/home/ubuntu/data/sdwebui/models/` 或原服务器补传 models 目录即可。
+重新从 `/home/xsl/data/sdwebui/models/` 或原服务器补传 models 目录即可。
 
 **Q：本机没有 miniconda 基础安装，只传 envs 够吗？**
 
@@ -299,4 +299,4 @@ webui 启动参数可能需要去掉 `--xformers`，改用 `--opt-sdp-no-mem-att
 
 **Q：本机磁盘空间是否足够？**
 
-当前 `/home` 剩余约 30G。3 个 conda 环境约 20–25G，解压后可能吃紧。建议解压一个删一个 tar 包，或考虑清理 `/home/ubuntu/data/`（已有 16G 冗余数据，解压完成后可删除）。
+当前 `/home` 剩余约 30G。3 个 conda 环境约 20–25G，解压后可能吃紧。建议解压一个删一个 tar 包，或考虑清理 `/home/xsl/data/`（已有 16G 冗余数据，解压完成后可删除）。

fix_conda_paths.py

@@ -9,7 +9,7 @@ OLD_PREFIXES = [
     b"/home/szlc/miniconda3",
     b"/usr/local/miniconda3",
 ]
-NEW_PREFIX = b"/home/ubuntu/miniconda3"
+NEW_PREFIX = b"/home/xsl/miniconda3"
 
 def is_probably_binary(filepath):
     """通过文件扩展名和内容判断是否为二进制文件"""
@@ -61,7 +61,7 @@ def fix_env(env_dir):
 
 if __name__ == "__main__":
     for env in sys.argv[1:]:
-        env_path = f"/home/ubuntu/miniconda3/envs/{env}"
+        env_path = f"/home/xsl/miniconda3/envs/{env}"
         if os.path.isdir(env_path):
             print(f"[{env}] 修复中...")
             fix_env(env_path)

hair_grow_service/README.md

@@ -7,12 +7,12 @@
 ### 1. 确认 webui 已启动（生发依赖 SD inpainting）
 ```bash
 ss -tlnp | grep 57860   # 确认 webui 在跑
-bash /home/ubuntu/change_hair/start_all.sh status   # 看 webui 是否就绪
+bash /home/xsl/change_hair/start_all.sh status   # 看 webui 是否就绪
 ```
 
 ### 2. 启动生发服务（端口 8899）
 ```bash
-nohup /home/ubuntu/change_hair/start_hairgrow.sh > /home/ubuntu/change_hair/project/logs/hair_grow_service.log 2>&1 &
+nohup /home/xsl/change_hair/start_hairgrow.sh > /home/xsl/change_hair/project/logs/hair_grow_service.log 2>&1 &
 ```
 
 ### 3. 打开测试页面

hair_grow_service/app.py

@@ -14,7 +14,7 @@ import base64
 import traceback
 
 # 把 hair_service_sd 加入 path，使其模块可被导入
-HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
 sys.path.insert(0, HAIR_SERVICE_DIR)
 # 切换到 hair_service_sd 目录，使 common.logger 能读到 config/configure.ini（相对路径）
 os.chdir(HAIR_SERVICE_DIR)
@@ -414,7 +414,7 @@ def train_src_img(hair_id):
     """返回发型的训练原图（hair_type_images/<hair_id>.jpg/.png）。
     用于测试页展示发型真实样子，而非套在标准脸上的效果图。
     """
-    src_dir = "/home/ubuntu/change_hair/hair_type_images"
+    src_dir = "/home/xsl/change_hair/hair_type_images"
     for ext in (".jpg", ".jpeg", ".png"):
         path = os.path.join(src_dir, hair_id + ext)
         if os.path.exists(path):

project/hair_service_sd/change_color.py

@@ -70,7 +70,11 @@ def process_infer(user_img_path, target_color, color_dir, dst_path, ratio=1.0):
     # 通过webui对照片进行增强
     s4 = time.time()
     # cv2.imwrite('/home/student/Downloads/12121.jpg',crop_img)
+    try:
         enhanced_img = enhance_hair.webui_img2img(crop_img, crop_mask, prompt=prompt)
+    except Exception as e:
+        print(f"[process_infer] webui enhance skipped: {e}")
+        enhanced_img = crop_img
     print("----------------------------webui_img2img", time.time() - s4)
 
     # 写回原图

project/hair_service_sd/config/configure.ini

@@ -32,4 +32,3 @@ name=watch-time
 [errorlogger]
 level=ERROR
 name=w-error
-

project/hair_service_sd/core/hairstyle_model.py

@@ -90,6 +90,14 @@ class HairStyle_Model(object):
         self.keypoints_processor = hair_init.keypoints_processor
         self.human_keypoint = hair_init.human_keypoint
 
+        # 用户图预处理产物内存缓存（一级缓存，避免重复 GPU 推理 + 磁盘 IO）。
+        # key = 图片字节哈希 + ratio；value = 7 个产物 + landmarks_1k 的 dict。
+        # 接口2 女性多发型场景：同一张用户图连续请求，第二次起命中缓存省 ~1.6s。
+        import collections
+        self._user_prepare_cache = {}  # {cache_key: {产物dict}}
+        self._user_prepare_cache_keys = collections.deque()  # LRU 顺序
+        self._USER_CACHE_MAX = 8  # 最多缓存 8 张图（防止内存膨胀）
+
 
         # worker_id = int(os.environ.get('APP_WORKER_ID', 1))
         # rand_max = 9527
@@ -2268,6 +2276,39 @@ class HairStyle_Model(object):
                                                                                  ref_landmark_f1k2_768)
             cv2.imwrite(another_pose_hair_image_dir, another_pose_hair_image*255)
         landmark1k_dir = osp.join(userinfo_dir, 'kpt_1k.txt')
+
+        # ===== 内存缓存（一级）：key = 图片哈希 + ratio =====
+        # 命中则跳过 landmark 检测 + get_prepare_user_768_data 整条 GPU 管线（省 ~1.6s）
+        import hashlib as _hashlib
+        _img_hash = _hashlib.md5(user_rgb_8uc3_orisize.tobytes()).hexdigest()[:16]
+        _cache_key = f"{_img_hash}_r{ratio}"
+        _cached = self._user_prepare_cache.get(_cache_key)
+
+        if _cached is not None:
+            # 命中内存缓存：直接取所有产物，跳过 landmark 检测和预处理管线
+            landmarks_origin_img_1k = _cached['landmarks_1k']
+            user_bald_res_8uc3_orisize = _cached['bald_res']
+            user_baldseg_8uc3_orisize = _cached['baldseg_ori']
+            user_baldseg_8uc3_768 = _cached['baldseg_768']
+            user_bald_8uc3_768 = _cached['bald_768']
+            user_landmark_f1k2_768 = _cached['lmk_768']
+            user_hairstyle_M = _cached['M']
+            user_matting_8uc3_bald_orisize = _cached['matting_ori']
+            # 补写磁盘文件：下游代码（功能7等）仍从 userinfo_dir 读这些文件，
+            # 而 task_id 每次不同导致 userinfo_dir 不同，必须补写保证下游可用
+            os.makedirs(userinfo_dir, exist_ok=True)
+            np.savetxt(landmark1k_dir, landmarks_origin_img_1k)
+            cv2.imwrite(osp.join(userinfo_dir, 'bald_res_ori.png'), user_bald_res_8uc3_orisize)
+            cv2.imwrite(osp.join(userinfo_dir, 'bald_seg_ori.png'), user_baldseg_8uc3_orisize)
+            cv2.imwrite(osp.join(userinfo_dir, 'user_baldseg_768.png'), user_baldseg_8uc3_768)
+            cv2.imwrite(osp.join(userinfo_dir, 'bald_seg_768.png'), user_bald_8uc3_768)
+            np.savetxt(osp.join(userinfo_dir, 'landmark_f1k2_768.txt'), user_landmark_f1k2_768)
+            np.savetxt(osp.join(userinfo_dir, 'hairstyle_M.txt'), user_hairstyle_M)
+            if user_matting_8uc3_bald_orisize is not None:
+                cv2.imwrite(osp.join(userinfo_dir, 'user_orig_mask.png'), user_matting_8uc3_bald_orisize)
+            self.logger_process.info(f"内存缓存命中 key={_cache_key}，跳过用户图预处理(补写磁盘文件)")
+        else:
+            # 未命中：走原有逻辑（landmark 检测 + 预处理管线 + 磁盘缓存）
             if not osp.exists(landmark1k_dir):
                 landmarks_origin_img_1k, bounding_box, euler_info = self.get_landmark.forward_diy(user_rgb_8uc3_orisize)
                 if landmarks_origin_img_1k is None:
@@ -2311,6 +2352,24 @@ class HairStyle_Model(object):
                 user_landmark_f1k2_768 = np.loadtxt(user_landmark_f1k2_768_dir)
                 user_hairstyle_M = np.loadtxt(user_hairstyle_M_dir)
 
+            # 写入内存缓存（含 landmarks_1k 和 matting，供后续同图请求命中）
+            self._user_prepare_cache[_cache_key] = {
+                'landmarks_1k': landmarks_origin_img_1k,
+                'bald_res': user_bald_res_8uc3_orisize,
+                'baldseg_ori': user_baldseg_8uc3_orisize,
+                'baldseg_768': user_baldseg_8uc3_768,
+                'bald_768': user_bald_8uc3_768,
+                'lmk_768': user_landmark_f1k2_768,
+                'M': user_hairstyle_M,
+                'matting_ori': user_matting_8uc3_bald_orisize,
+            }
+            self._user_prepare_cache_keys.append(_cache_key)
+            # LRU 淘汰：超过上限删除最老的
+            while len(self._user_prepare_cache_keys) > self._USER_CACHE_MAX:
+                _old = self._user_prepare_cache_keys.popleft()
+                self._user_prepare_cache.pop(_old, None)
+            self.logger_process.info(f"内存缓存写入 key={_cache_key}，当前缓存 {len(self._user_prepare_cache)} 张")
+
         # show_concat = np.concatenate((user_rgb_8uc3_orisize, user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize), axis=1)
         # ratio = 1536. / max(show_concat.shape[:2])
         # show_concat = cv2.resize(show_concat, (0, 0), fx=ratio, fy=ratio)
@@ -2318,18 +2377,26 @@ class HairStyle_Model(object):
         # cv2.waitKey()
 
         user_orig_mask_path = os.path.join(userinfo_dir, "user_orig_mask.png")
-        if not os.path.exists(user_orig_mask_path):
+        if user_matting_8uc3_bald_orisize is not None and not os.path.exists(user_orig_mask_path):
             cv2.imwrite(user_orig_mask_path, user_matting_8uc3_bald_orisize)
 
-        # 换发型
+        # 换发型（主 GAN）
+        import time as _time
+        _t_gan0 = _time.perf_counter()
         hair_gene_8uc3_768 = self.generator_hair.Generator_Hair_inference_use_pref(another_pose_hair_image,
                                                                                    user_baldseg_8uc3_768,
                                                                                    user_bald_8uc3_768,
                                                                                    user_landmark_f1k2_768, gender)
+        _t_gan = _time.perf_counter() - _t_gan0
 
 
+        # 融合（第3次matte + 融合GAN）
+        _t_fuse0 = _time.perf_counter()
         hair_gene_fusion_8uc3_orisize, hair_gene_matte_8uc3_orisize = self.process_data.get_fusion_res_hairpaste(
             user_bald_res_8uc3_orisize, hair_gene_8uc3_768, user_landmark_f1k2_768, user_hairstyle_M)
+        _t_fuse = _time.perf_counter() - _t_fuse0
+        self.logger_process.info(
+            f"功能6分步计时: 主GAN={_t_gan:.3f}s 融合={_t_fuse:.3f}s (缓存={'命中' if _cached is not None else '未命中'})")
 
         gen_hair_mask_path_2 = os.path.join(userinfo_dir, "hair_mask_2.png")
         cv2.imwrite(gen_hair_mask_path_2, hair_gene_matte_8uc3_orisize)

project/hair_service_sd/gen_super_image.py

@@ -1,4 +1,5 @@
 import io
+import os
 import os.path
 import time
 
@@ -83,13 +84,13 @@ class ControlnetRequestImg2Img:
         return encoded_image
 
 
-    def build_body_v2(self, dst_width, dst_height, cfg_scale, base_img, denoising_strength=0.7):
+    def build_body_v2(self, dst_width, dst_height, cfg_scale, base_img, denoising_strength=0.7, steps=None):
         self.body = {
             "prompt": self.prompt,
             "negative_prompt": self.neg_prompt,
             "sampler_name": "DPM++ 2M Karras",
             "batch_size": 1,
-            "steps": 20,
+            "steps": int(steps) if steps is not None else int(os.environ.get("WEBUI_STEPS", "15")),
             "width": dst_width,
             "height": dst_height,
             "cfg_scale": cfg_scale,
@@ -413,7 +414,7 @@ def encode_numpy_to_base64(img):
     encoded_image = base64.b64encode(bytes).decode('utf-8')
     return encoded_image
 
-def webui_img2img(img=None, mask_img=None, in_gender=None, task_id=None, hair_id=None, lora_material_path=None, tag="", is_hr=False, denoising_strength=0.7, inference_port="57860", refiner_switch_at=0.5):
+def webui_img2img(img=None, mask_img=None, in_gender=None, task_id=None, hair_id=None, lora_material_path=None, tag="", is_hr=False, denoising_strength=0.7, inference_port="57860", refiner_switch_at=0.5, webui_steps=None):
     # url = "http://hairservice.tslead.net:57860/sdapi/v1/img2img"
 
     neg_prompt = '(nsfw:1.5), ng_deepnegative_v1_75t, (badhandv4:1.2), (worst quality:2), (low quality:2), (normal quality:2), lowres, bad anatomy, bad hands, ((monochrome)), ((grayscale)) watermark, moles, large breast, big breast, bad_pictures,easynegative'
@@ -429,7 +430,7 @@ def webui_img2img(img=None, mask_img=None, in_gender=None, task_id=None, hair_id
     control_net = ControlnetRequestImg2Img(prompt, neg_prompt, mask_img)
     # control_net.build_body_hr(dst_width=img.shape[1], dst_height=img.shape[0], cfg_scale=7, base_img=encoded_image, denoising_strength=denoising_strength)
     if not is_hr:
-        control_net.build_body_v2(dst_width=img.shape[1], dst_height=img.shape[0], cfg_scale=7, base_img=encoded_image, denoising_strength=denoising_strength)
+        control_net.build_body_v2(dst_width=img.shape[1], dst_height=img.shape[0], cfg_scale=7, base_img=encoded_image, denoising_strength=denoising_strength, steps=webui_steps)
     else:
         control_net.build_body_hr(dst_width=img.shape[1], dst_height=img.shape[0], cfg_scale=7, base_img=encoded_image, denoising_strength=denoising_strength, refiner_switch_at=refiner_switch_at)
 

project/hair_service_sd/gunicorn_config.py

@@ -1,4 +1,4 @@
-#!/home/ubuntu/miniconda3/envs/yd/bin/python
+#!/home/xsl/miniconda3/envs/yd/bin/python
 # -*- coding: utf-8 -*-
 import os
 # 监听本机的端口

project/hair_service_sd/hairline_mask.py

@@ -25,7 +25,7 @@ import cv2
 import numpy as np
 
 # MediaPipe 模型路径
-MP_MODEL = "/home/ubuntu/change_hair/project/hair_service_sd/weights/mediapipe/face_landmarker.task"
+MP_MODEL = "/home/xsl/change_hair/project/hair_service_sd/weights/mediapipe/face_landmarker.task"
 _mp_detector = None
 
 

project/hair_service_sd/hairstyle_model_infer.py

@@ -72,6 +72,8 @@ class HairStyle_Model_Infer(object):
         # haircolor_dir = '/home/data/hair/data/ref_color/3628746832766'
         face_base, hair_matting, status = self.infer_haircolor_new(user_rgb_8uc3_orisize, haircolor_dir,target_hair_color,
                                                                return_matting=True)
+        if status != 0:
+            return user_rgb_8uc3_orisize, None, status
         user_rgb_8uc3_orisize = face_base
         # 构建一个色板
         r, g, b = target_hair_color
@@ -1016,7 +1018,7 @@ class HairStyle_Model_Infer(object):
     def infer_haircolor_new(self, user_rgb_8uc3_orisize, haircolor_dir, target_hair_color, return_matting=False):
         landmarks_origin_img_1k= self.get_landmark.forward(user_rgb_8uc3_orisize)
         if landmarks_origin_img_1k is None:
-            return None, 10001
+            return None, None, 10001
         need_process = (target_hair_color[0] * 0.299 + target_hair_color[1] * 0.587 + target_hair_color[2] * 0.114) > 150
         _, user_matting_8uc1_bald_orisize = self.process_data_infer.generator_matte.matte_inference(user_rgb_8uc3_orisize,
                                                                                               landmarks_origin_img_1k)
@@ -1036,12 +1038,15 @@ class HairStyle_Model_Infer(object):
         need_process=True
         if need_process:
             haircolor_dir_tmp = os.path.join(config.get('default', "haircolorDir"), config.get('default', "baseColor_ID"))
+            if os.path.exists(haircolor_dir_tmp):
                 face_base, status1 = self.infer_haircolor_tj(user_rgb_8uc3_orisize, haircolor_dir_tmp)
                 # cv2.imwrite('/home/student/Desktop/tmp_color/need/face_base.png', face_base)
                 if status1 == 0:
                     return face_base, user_matting_8uc1_bald_orisize, 0
                 else:
                     need_process = False
+            else:
+                need_process = False
         if not need_process:
             hair_gene_color_8uc3_768, user_matting_mask_8uc3_768 = self.change_haircolor.Change_Hair_inference(
                 user_rgb_8uc3_change_color_768,

project/hair_service_sd/run_copy_cost_colorb64.py

@@ -255,7 +255,8 @@ def change_hair_colorv3():
                 jsonify({'msg': '算法解析错误', 'result': '', 'umd': '', 'state': -1}), 400)
 
     except Exception as e:
-        print(e)
+        print(f"[hairColor ERROR] {e}")
+        traceback.print_exc()
         return make_response(
             jsonify({'msg': '算法解析错误', 'result': '', 'umd': '', 'state': -1}), 400)
 
@@ -605,6 +606,8 @@ def change_hairstyle_v4():
                 else:
                     p_tag = ""
             denoising_strength = float(input_info.get('denoising_strength', 0.6))  # 接口11 可调；默认 0.6
+            webui_steps_raw = input_info.get('webui_steps', None)  # 可选：webui img2img 步数，None用服务端默认
+            webui_steps = int(webui_steps_raw) if webui_steps_raw is not None else None
             print(f"功能8：处理发型区域，耗时：{time.time() - start_time:.3f}s")
             # cv2.imwrite(f"{task_id}_mask_dilate.jpg", mask_dilate)
             # cv2.imwrite(f"{task_id}_final_img.jpg", final_img)
@@ -618,7 +621,7 @@ def change_hairstyle_v4():
             # mask_dilate = cv2.imread("/root/project/hair_service_sd/gt_mask_dilate.jpg")
             sd_result = webui_img2img(img=final_img, mask_img=mask_dilate, in_gender=in_gender, task_id=task_id,
                                       hair_id=hair_id, lora_material_path=hair_material_dir, tag=p_tag, is_hr=is_hr,
-                                      denoising_strength=denoising_strength, inference_port="57860")
+                                      denoising_strength=denoising_strength, inference_port="57860", webui_steps=webui_steps)
             print(f"功能：webui，耗时：{time.time() - start_time:.3f}s")
 
             # 功能：后处理并上传结果

project/hair_service_sd/start_services.sh

@@ -4,7 +4,7 @@
 export CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
 
 # 创建Lora目录软连接
-TARGET_LORA="/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/Lora"
+TARGET_LORA="/home/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Lora"
 if [ -e "$TARGET_LORA" ]; then
     echo "删除已存在的Lora连接或目录: $TARGET_LORA"
     rm -rf "$TARGET_LORA"
@@ -13,16 +13,16 @@ ln -s /gz-fs/Lora "$TARGET_LORA"
 echo "创建软连接成功: /gz-fs/Lora -> $TARGET_LORA"
 
 # 1. 换发算法服务
-cd /home/ubuntu/change_hair/project/hair_service_sd
+cd /home/xsl/change_hair/project/hair_service_sd
-nohup /home/ubuntu/miniconda3/envs/condiff-train-hair/bin/python run_copy_cost_colorb64.py > hair_service.log 2>&1 &
+nohup /home/xsl/miniconda3/envs/condiff-train-hair/bin/python run_copy_cost_colorb64.py > hair_service.log 2>&1 &
 
 # 2. webUI服务推理 + onediff
-cd /home/ubuntu/change_hair/project/onediff/stable-diffusion-webui
+cd /home/xsl/change_hair/project/onediff/stable-diffusion-webui
-nohup /home/ubuntu/miniconda3/envs/onediff/bin/python webui.py --api --listen --xformers --port 57860 > webui.log 2>&1 &
+nohup /home/xsl/miniconda3/envs/onediff/bin/python webui.py --api --listen --xformers --port 57860 > webui.log 2>&1 &
 nohup ./webui.sh --api --listen --disable-safe-unpickle --port 9038 > webui_sh.log 2>&1 &
 
 # 3. photo_service
-cd /home/ubuntu/change_hair/project/photo_service
+cd /home/xsl/change_hair/project/photo_service
-nohup /home/ubuntu/miniconda3/envs/py310/bin/python lora_train_service_1.py > photo_service.log 2>&1 &
+nohup /home/xsl/miniconda3/envs/py310/bin/python lora_train_service_1.py > photo_service.log 2>&1 &
 
 echo "所有服务已在后台启动"

project/hair_service_sd/utils/enhance_hair.py

@@ -43,6 +43,9 @@ def webui_img2img(img, mask, prompt='', denoising_strength=0.35):
         }
     response = requests.post(url=url, json=request_dict)
     ret_json = response.json()
+    if 'images' not in ret_json:
+        print(f"[webui_img2img ERROR] {ret_json}")
+        raise Exception(f"webui error: {ret_json.get('error', 'unknown')}")
     result = ret_json['images'][0]
     img = cv2.imdecode(np.frombuffer(base64.b64decode(result.split(",", 1)[0]), np.uint8), cv2.IMREAD_COLOR)
     return img

project/photo_service/lora_train_service_1.py

@@ -307,13 +307,13 @@ def train_thread(sq, gpu_id):
             #   3. GPU 固定 device=0(单卡)
             #   4. 去掉 tokenizer_cache_dir(改用 HF 本地缓存 + 离线模式)
             #   5. 设置 HF_HUB_OFFLINE 避免联网检查
-            kohya_python = '/home/ubuntu/miniconda3/envs/kohya/bin/python'
+            kohya_python = '/home/ubuntu/miniconda3/envs/my_hair/bin/python'
             kohya_workdir = os.path.join(kohya_ss_home_dir, 'kohya_ss')
             base_model = '/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors'
             cmd_train = (
                 f'cd {kohya_workdir} && '
                 f'HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 CUDA_VISIBLE_DEVICES={device_id} '
-                f'/home/ubuntu/miniconda3/envs/kohya/bin/accelerate launch --num_cpu_threads_per_process=2 "./train_network.py" --enable_bucket '
+                f'/home/ubuntu/miniconda3/envs/my_hair/bin/accelerate launch --num_cpu_threads_per_process=2 "./train_network.py" --enable_bucket '
                 f'--min_bucket_reso=256 --max_bucket_reso=2048 --pretrained_model_name_or_path="{base_model}" '
                 f'--train_data_dir={images_dir} --resolution="2000,2000" '
                 f'--output_dir={model_dir} '

scripts/sync_data_to_server.sh

@@ -1,10 +1,10 @@
 #!/usr/bin/env bash
-# 断点续传大文件到云服务器 /home/ubuntu/data
+# 断点续传大文件到云服务器 /home/xsl/data
 # 用法: bash scripts/sync_data_to_server.sh [all|weights|sd|hairstyles]
 set -euo pipefail
 
 REMOTE="ubuntu@117.50.213.111"
-REMOTE_BASE="/home/ubuntu/data"
+REMOTE_BASE="/home/xsl/data"
 LOCAL_BASE="/home/xsl/change_hair"
 LOG_DIR="${LOCAL_BASE}/project/logs"
 mkdir -p "$LOG_DIR"

train_hairstyle_full.py

@@ -158,6 +158,10 @@ def step3_wait_and_callback(hair_id, template_img):
     else:
         print("  ✗ 训练超时"); return False
     time.sleep(5)  # 等训练进程写完
+    # 训练时停掉了 hair 服务，回调前需要重启
+    print("  重启 change_hair-hair 服务...")
+    os.system("sudo systemctl restart change_hair-hair")
+    time.sleep(20)
     # 触发回调生成ref材质
     print("  触发回调生成ref材质...")
     r = req.post(HAIR_CALLBACK, json={"task_id":f"train_{hair_id}","hair_id":hair_id,