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
perf(swapHair): webui降步数 + 用户图预处理内存缓存
2026-07-26 16:49:21 +08:00 · 2026-07-26 16:41:19 +08:00 · 2026-07-19 15:52:51 +08:00 · 2026-07-19 00:49:16 +08:00 · 2026-07-18 19:18:56 +08:00 · 2026-07-18 19:00:15 +08:00
34 changed files with 348 additions and 166 deletions
@@ -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 |
-| SD 模型 | 实际加载 `v1-5-pruned-emaonly`（config 里写的 majicmix 不存在，webui 自动 fallback） |
+| 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 | `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`）

@@ -8,8 +8,8 @@
  3. 串行执行，记录每个发型成功/失败

 用法:
-  cd /home/xsl/change_hair/project/hair_service_sd
-  python /home/xsl/change_hair/batch_train_hairstyles.py --src /home/xsl/change_hair/hair_type_images --gender girl [--only 圆-心形] [--start-from 心形-心形]
+  cd /home/ubuntu/change_hair/project/hair_service_sd
+  python /home/ubuntu/change_hair/batch_train_hairstyles.py --src /home/ubuntu/change_hair/hair_type_images --gender girl [--only 圆-心形] [--start-from 心形-心形]

 注意：
  - 中文 hair_id 直接用文件名（去扩展名）作为 ID
@@ -24,14 +24,14 @@ import argparse
 import subprocess
 from datetime import datetime

-SRC_DEFAULT = "/home/xsl/change_hair/hair_type_images"
-WORK_DIR = "/home/xsl/change_hair/data/batch_train_inputs"
-LOG_FILE = "/home/xsl/change_hair/data/batch_train_log.txt"
-TRAIN_SCRIPT = "/home/xsl/change_hair/train_hairstyle_full.py"
-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
-LOCK_FILE = "/home/xsl/change_hair/data/batch_train.pid"
-DONE_FILE = "/home/xsl/change_hair/data/batch_train_done.txt"  # 已完成发型清单（断点续跑）
-LORA_DIR = "/home/xsl/change_hair/data/train_material"          # LoRA 输出根目录
+SRC_DEFAULT = "/home/ubuntu/change_hair/hair_type_images"
+WORK_DIR = "/home/ubuntu/change_hair/data/batch_train_inputs"
+LOG_FILE = "/home/ubuntu/change_hair/data/batch_train_log.txt"
+TRAIN_SCRIPT = "/home/ubuntu/change_hair/train_hairstyle_full.py"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
+LOCK_FILE = "/home/ubuntu/change_hair/data/batch_train.pid"
+DONE_FILE = "/home/ubuntu/change_hair/data/batch_train_done.txt"  # 已完成发型清单（断点续跑）
+LORA_DIR = "/home/ubuntu/change_hair/data/train_material"          # LoRA 输出根目录


 def acquire_lock():
@@ -125,7 +125,7 @@ def run_one(hair_id, src_img, gender, py, done_set, force=False):
    log(f"  调用: {' '.join(cmd[:2])} ... --hair-id {hair_id}")
    try:
        # 子进程输出实时写到日志文件
-        proc_log = os.path.join("/home/xsl/change_hair/data", f"subprocess_{hair_id}.log")
+        proc_log = os.path.join("/home/ubuntu/change_hair/data", f"subprocess_{hair_id}.log")
        with open(proc_log, "w", encoding="utf-8") as f:
            r = subprocess.run(
                cmd, cwd=HAIR_SERVICE_DIR,
@@ -152,7 +152,7 @@ def main():
    ap = argparse.ArgumentParser(description="批量训练发型")
    ap.add_argument("--src", default=SRC_DEFAULT, help="原图目录")
    ap.add_argument("--gender", default="girl", choices=["boy", "girl"])
-    ap.add_argument("--py", default="/home/xsl/miniconda3/envs/my_hair/bin/python",
+    ap.add_argument("--py", default="/home/ubuntu/miniconda3/envs/my_hair/bin/python",
                    help="python 解释器")
    ap.add_argument("--only", default=None, help="只训练指定 hair_id（调试用）")
    ap.add_argument("--start-from", default=None,
@@ -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
@@ -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
@@ -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
@@ -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
+}
@@ -8,7 +8,7 @@ import sys
 import cv2
 import numpy as np

-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")

@@ -87,7 +87,7 @@ def gen_hairline_mask(img_path, mask_path):

 if __name__ == "__main__":
    img_path = sys.argv[1] if len(sys.argv) > 1 else \
-        "/home/xsl/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
+        "/home/ubuntu/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
    mask_path = sys.argv[2] if len(sys.argv) > 2 else \
-        "/home/xsl/change_hair/project/logs/hairgrow_test_mask.png"
+        "/home/ubuntu/change_hair/project/logs/hairgrow_test_mask.png"
    gen_hairline_mask(img_path, mask_path)
@@ -14,15 +14,15 @@ import time
 import base64
 import requests

-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.chdir(HAIR_SERVICE_DIR)

-BOY_IMG = "/home/xsl/change_hair/images/boy.png"
-GIRL_IMG = "/home/xsl/change_hair/images/girl.png"
-HAIRSTYLE_DIR = "/home/xsl/change_hair/project/data/ref_hairstyle"
-TRAIN_DIR = "/home/xsl/change_hair/data/train_material"
-PREVIEW_DIR = "/home/xsl/change_hair/hair_grow_service/static/previews"  # 预览图存储目录
+BOY_IMG = "/home/ubuntu/change_hair/images/boy.png"
+GIRL_IMG = "/home/ubuntu/change_hair/images/girl.png"
+HAIRSTYLE_DIR = "/home/ubuntu/change_hair/project/data/ref_hairstyle"
+TRAIN_DIR = "/home/ubuntu/change_hair/data/train_material"
+PREVIEW_DIR = "/home/ubuntu/change_hair/hair_grow_service/static/previews"  # 预览图存储目录
 SWAP_API = "http://127.0.0.1:8801/api/swapHair/v1"


@@ -8,7 +8,7 @@
  hair_template_material/<hair_id>/first##<name>.pkl       1k关键点

 用法:
-  cd /home/xsl/change_hair/project/hair_service_sd
+  cd /home/ubuntu/change_hair/project/hair_service_sd
  python gen_template_material.py --hair-id new_test_001 --img /path/to/template.jpg
 """
 import os
@@ -19,7 +19,7 @@ import pickle
 import argparse

 ssl._create_default_https_context = ssl._create_unverified_context
-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 os.chdir(HAIR_SERVICE_DIR)
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -6,12 +6,12 @@
 需要在 hair_service_sd 目录下运行（依赖其模块导入）。

 用法:
-    cd /home/xsl/change_hair/project/hair_service_sd
-    /home/xsl/miniconda3/envs/my_hair/bin/python /home/xsl/change_hair/hair_grow_cli.py \
+    cd /home/ubuntu/change_hair/project/hair_service_sd
+    /home/ubuntu/miniconda3/envs/my_hair/bin/python /home/ubuntu/change_hair/hair_grow_cli.py \
        --img test.jpg --mask mask.png --strength 0.5 -o result.jpg

    # 批量跑三档强度对比
-    /home/xsl/miniconda3/envs/my_hair/bin/python /home/xsl/change_hair/hair_grow_cli.py \
+    /home/ubuntu/miniconda3/envs/my_hair/bin/python /home/ubuntu/change_hair/hair_grow_cli.py \
        --img test.jpg --mask mask.png --compare
 """
 import os
@@ -20,7 +20,7 @@ import cv2
 import argparse

 # 把 hair_service_sd 加入路径，使其模块可被导入
-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 sys.path.insert(0, HAIR_SERVICE_DIR)

 # 设置离线模式（本机无法访问 huggingface.co）
@@ -6,9 +6,9 @@
 原始图 → 人脸检测+1k关键点 → Generator_Matte头发抠图 → 白底合成 + 居中裁剪 → 打标签

 用法:
-    cd /home/xsl/change_hair/project/hair_service_sd
-    /home/xsl/miniconda3/envs/my_hair/bin/python /home/xsl/change_hair/prepare_train_data.py \
-        --input /home/xsl/change_hair/train_images \
+    cd /home/ubuntu/change_hair/project/hair_service_sd
+    /home/ubuntu/miniconda3/envs/my_hair/bin/python /home/ubuntu/change_hair/prepare_train_data.py \
+        --input /home/ubuntu/change_hair/train_images \
        --hair-id new_hairstyle_001 \
        --gender boy
 """
@@ -22,7 +22,7 @@ import argparse
 ssl._create_default_https_context = ssl._create_unverified_context

 # 必须在 hair_service_sd 目录运行（依赖其模块）
-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 os.chdir(HAIR_SERVICE_DIR)
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -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)
-    enhanced_img = enhance_hair.webui_img2img(crop_img, crop_mask, prompt=prompt)
+    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
@@ -1,21 +1,21 @@
 [default]
 modelDir = weights
-hairstyleDir = /home/xsl/change_hair/project/data/ref_hairstyle
-haircolorDir= /home/xsl/change_hair/project/data/ref_haircolor
-userDir=/home/xsl/change_hair/project/data/userImage
-tmp_dir=/home/xsl/change_hair/project/data/tmp
-res_dir=/home/xsl/change_hair/project/data/res_dir
-userInfo_dir=/home/xsl/change_hair/project/data/user_info
+hairstyleDir = /home/ubuntu/change_hair/project/data/ref_hairstyle
+haircolorDir= /home/ubuntu/change_hair/project/data/ref_haircolor
+userDir=/home/ubuntu/change_hair/project/data/userImage
+tmp_dir=/home/ubuntu/change_hair/project/data/tmp
+res_dir=/home/ubuntu/change_hair/project/data/res_dir
+userInfo_dir=/home/ubuntu/change_hair/project/data/user_info
 baseColor_ID=HDR10_443322
 Port = 11023
-refer_dir = /home/xsl/change_hair/project/data/ref_online
-ref_user_dir = /home/xsl/change_hair/project/data/ref_user_imgs
-train_dir = /home/xsl/change_hair/data/train_material
-refImgDir=/home/xsl/change_hair/project/data/refImage
-hair_template_material_dir=/home/xsl/change_hair/project/data/hair_template_material
-ref_color=/home/xsl/change_hair/project/data/ref_color
-ref_color_img=/home/xsl/change_hair/project/data/ref_color_imgs
-upload_train_dir=/home/xsl/change_hair/project/data/upload_train_imgs
+refer_dir = /home/ubuntu/change_hair/project/data/ref_online
+ref_user_dir = /home/ubuntu/change_hair/project/data/ref_user_imgs
+train_dir = /home/ubuntu/change_hair/data/train_material
+refImgDir=/home/ubuntu/change_hair/project/data/refImage
+hair_template_material_dir=/home/ubuntu/change_hair/project/data/hair_template_material
+ref_color=/home/ubuntu/change_hair/project/data/ref_color
+ref_color_img=/home/ubuntu/change_hair/project/data/ref_color_imgs
+upload_train_dir=/home/ubuntu/change_hair/project/data/upload_train_imgs
 ;version=local
 version=online

@@ -23,7 +23,7 @@ version=online
 strength=1

 [logger]
-logpath = /home/xsl/change_hair/project/logs
+logpath = /home/ubuntu/change_hair/project/logs
 level=INFO

 [timelogger]
@@ -32,4 +32,3 @@ name=watch-time
 [errorlogger]
 level=ERROR
 name=w-error
-
@@ -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
@@ -2268,48 +2276,99 @@ 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')
-        if not osp.exists(landmark1k_dir):
-            landmarks_origin_img_1k, bounding_box, euler_info = self.get_landmark.forward_diy(user_rgb_8uc3_orisize)
+
+        # ===== 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:
+                    return None, 10001, None, None, None
+                np.savetxt(landmark1k_dir, landmarks_origin_img_1k)
+            else:
+                landmarks_origin_img_1k = np.loadtxt(landmark1k_dir)
+            # landmarks_origin_img_1k, _, _ = self.get_landmark.forward(user_rgb_8uc3_orisize)
            if landmarks_origin_img_1k is None:
                return None, 10001, None, None, None
-            np.savetxt(landmark1k_dir, landmarks_origin_img_1k)
-        else:
-            landmarks_origin_img_1k = np.loadtxt(landmark1k_dir)
-        # landmarks_origin_img_1k, _, _ = self.get_landmark.forward(user_rgb_8uc3_orisize)
-        if landmarks_origin_img_1k is None:
-            return None, 10001, None, None, None

-        user_bald_res_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_res_ori.png')
-        user_baldseg_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_seg_ori.png')
-        user_baldseg_8uc3_768_dir = osp.join(userinfo_dir, 'user_baldseg_768.png')
-        user_bald_8uc3_768_dir = osp.join(userinfo_dir, 'bald_seg_768.png')
-        user_landmark_f1k2_768_dir = osp.join(userinfo_dir, 'landmark_f1k2_768.txt')
-        user_hairstyle_M_dir = osp.join(userinfo_dir, 'hairstyle_M.txt')
+            user_bald_res_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_res_ori.png')
+            user_baldseg_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_seg_ori.png')
+            user_baldseg_8uc3_768_dir = osp.join(userinfo_dir, 'user_baldseg_768.png')
+            user_bald_8uc3_768_dir = osp.join(userinfo_dir, 'bald_seg_768.png')
+            user_landmark_f1k2_768_dir = osp.join(userinfo_dir, 'landmark_f1k2_768.txt')
+            user_hairstyle_M_dir = osp.join(userinfo_dir, 'hairstyle_M.txt')

-        condition_exist2 = True
-        pre_list = [user_bald_res_8uc3_orisize_dir, user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_768_dir, user_bald_8uc3_768_dir,
-                    user_landmark_f1k2_768_dir, user_hairstyle_M_dir]
-        for tmp_dir in pre_list:
-            if not osp.exists(tmp_dir):
-                condition_exist2 = False
+            condition_exist2 = True
+            pre_list = [user_bald_res_8uc3_orisize_dir, user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_768_dir, user_bald_8uc3_768_dir,
+                        user_landmark_f1k2_768_dir, user_hairstyle_M_dir]
+            for tmp_dir in pre_list:
+                if not osp.exists(tmp_dir):
+                    condition_exist2 = False

-        user_matting_8uc3_bald_orisize = None
-        if not condition_exist2:
-            user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize, user_baldseg_8uc3_768, user_bald_8uc3_768, \
-                user_landmark_f1k2_768, user_hairstyle_M, user_matting_8uc3_bald_orisize = self.process_data.get_prepare_user_768_data(user_rgb_8uc3_orisize, landmarks_origin_img_1k, ratio=ratio)
-            cv2.imwrite(user_bald_res_8uc3_orisize_dir, user_bald_res_8uc3_orisize)
-            cv2.imwrite(user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_orisize)
-            cv2.imwrite(user_baldseg_8uc3_768_dir, user_baldseg_8uc3_768)
-            cv2.imwrite(user_bald_8uc3_768_dir, user_bald_8uc3_768)
-            np.savetxt(user_landmark_f1k2_768_dir, user_landmark_f1k2_768)
-            np.savetxt(user_hairstyle_M_dir, user_hairstyle_M)
-        else:
-            user_bald_res_8uc3_orisize = cv2.imread(user_bald_res_8uc3_orisize_dir)
-            user_baldseg_8uc3_orisize = cv2.imread(user_baldseg_8uc3_orisize_dir)
-            user_baldseg_8uc3_768 = cv2.imread(user_baldseg_8uc3_768_dir)
-            user_bald_8uc3_768 = cv2.imread(user_bald_8uc3_768_dir)
-            user_landmark_f1k2_768 = np.loadtxt(user_landmark_f1k2_768_dir)
-            user_hairstyle_M = np.loadtxt(user_hairstyle_M_dir)
+            user_matting_8uc3_bald_orisize = None
+            if not condition_exist2:
+                user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize, user_baldseg_8uc3_768, user_bald_8uc3_768, \
+                    user_landmark_f1k2_768, user_hairstyle_M, user_matting_8uc3_bald_orisize = self.process_data.get_prepare_user_768_data(user_rgb_8uc3_orisize, landmarks_origin_img_1k, ratio=ratio)
+                cv2.imwrite(user_bald_res_8uc3_orisize_dir, user_bald_res_8uc3_orisize)
+                cv2.imwrite(user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_orisize)
+                cv2.imwrite(user_baldseg_8uc3_768_dir, user_baldseg_8uc3_768)
+                cv2.imwrite(user_bald_8uc3_768_dir, user_bald_8uc3_768)
+                np.savetxt(user_landmark_f1k2_768_dir, user_landmark_f1k2_768)
+                np.savetxt(user_hairstyle_M_dir, user_hairstyle_M)
+            else:
+                user_bald_res_8uc3_orisize = cv2.imread(user_bald_res_8uc3_orisize_dir)
+                user_baldseg_8uc3_orisize = cv2.imread(user_baldseg_8uc3_orisize_dir)
+                user_baldseg_8uc3_768 = cv2.imread(user_baldseg_8uc3_768_dir)
+                user_bald_8uc3_768 = cv2.imread(user_bald_8uc3_768_dir)
+                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])
@@ -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)
@@ -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)

@@ -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)
@@ -1036,10 +1038,13 @@ 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"))
-            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
+            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:
@@ -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
@@ -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
@@ -37,8 +37,8 @@ if version == "local":
    callback_url = 'http://service.aicloud.fit:7395/api/hair/trainCallBack'
 else:
    current_url = 'http://0.0.0.0:7393/'
-    kohya_ss_home_dir = '/home/xsl/change_hair/project/kohya_ss_home'
-    webui_lora_dir = '/home/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Lora'
+    kohya_ss_home_dir = '/home/ubuntu/change_hair/project/kohya_ss_home'
+    webui_lora_dir = '/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/Lora'
    inference_use_onediff = False
    callback_url = 'http://0.0.0.0:8801/api/hair/trainCallBack'
 base_webui_port = '57860'
@@ -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/xsl/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/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors'
+            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/xsl/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} '
@@ -10,16 +10,16 @@
 #   bash start_all.sh status   # 查看状态
 # =====================================================================
 set -u
-BASE="/home/xsl/change_hair"
+BASE="/home/ubuntu/change_hair"
 PROJ="$BASE/project"
 LOGDIR="$PROJ/logs"
 PIDD="$LOGDIR/pids"
 mkdir -p "$LOGDIR" "$PIDD"

 # conda 环境的 python 路径
-PY_HAIR="/home/xsl/miniconda3/envs/my_hair/bin/python"
-PY_SD="/home/xsl/miniconda3/envs/sdwebui/bin/python"
-PY_PHOTO="/home/xsl/miniconda3/envs/py310/bin/python"
+PY_HAIR="/home/ubuntu/miniconda3/envs/my_hair/bin/python"
+PY_SD="/home/ubuntu/miniconda3/envs/sdwebui/bin/python"
+PY_PHOTO="/home/ubuntu/miniconda3/envs/py310/bin/python"

 # 公共环境变量
 export CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1   # 旧版 cryptography 兼容
@@ -2,5 +2,5 @@
 export CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
 export CUDA_VISIBLE_DEVICES=0
 export APP_WORKER_ID=1
-cd /home/xsl/change_hair/project/hair_service_sd
-exec /home/xsl/miniconda3/envs/my_hair/bin/python run_copy_cost_colorb64.py
+cd /home/ubuntu/change_hair/project/hair_service_sd
+exec /home/ubuntu/miniconda3/envs/my_hair/bin/python run_copy_cost_colorb64.py
@@ -5,5 +5,5 @@ export CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
 export CUDA_VISIBLE_DEVICES=0
 export HF_HUB_OFFLINE=1
 export TRANSFORMERS_OFFLINE=1
-cd /home/xsl/change_hair/hair_grow_service
-exec /home/xsl/miniconda3/envs/my_hair/bin/python app.py
+cd /home/ubuntu/change_hair/hair_grow_service
+exec /home/ubuntu/miniconda3/envs/my_hair/bin/python app.py
@@ -1,5 +1,5 @@
 #!/bin/bash
 # 独立启动 photo_service，确保脱离会话
 export CRYPTOGRAPHY_OPENSSL_NO_LEGACY=1
-cd /home/xsl/change_hair/project/photo_service
-exec /home/xsl/miniconda3/envs/py310/bin/python -u lora_train_service_1.py
+cd /home/ubuntu/change_hair/project/photo_service
+exec /home/ubuntu/miniconda3/envs/py310/bin/python -u lora_train_service_1.py
@@ -4,7 +4,7 @@
 # 已传数据自动续传（--partial）

 set -u
-LOGDIR="/home/xsl/change_hair/project/logs"
+LOGDIR="/home/ubuntu/change_hair/project/logs"
 DONE_MARKER="$LOGDIR/sync_all.done"
 mkdir -p "$LOGDIR"

@@ -14,15 +14,15 @@ mkdir -p "$LOGDIR"
 # 任务定义：名称|源|目标
 # 优先级排序：换发色关键链路优先 → 换发型 → 训练相关 → 大数据最后
 TASKS=(
-  "photo_service|szlc@192.168.101.63:/home/szlc/project/photo_service/|/home/xsl/change_hair/project/photo_service/"
-  "project_data|szlc@192.168.101.63:/home/szlc/project/data/|/home/xsl/change_hair/project/data/"
-  "hair_service_sd|szlc@192.168.101.63:/home/szlc/project/hair_service_sd/|/home/xsl/change_hair/project/hair_service_sd/"
-  "conda_py310|szlc@192.168.101.63:/home/szlc/miniconda3/envs/py310/|/home/xsl/miniconda3/envs/py310/"
-  "conda_my_hair|szlc@192.168.101.63:/home/szlc/miniconda3/envs/my_hair/|/home/xsl/miniconda3/envs/my_hair/"
-  "conda_sdwebui|szlc@192.168.101.63:/home/szlc/miniconda3/envs/sdwebui/|/home/xsl/miniconda3/envs/sdwebui/"
-  "kohya_ss_home|szlc@192.168.101.63:/home/szlc/project/kohya_ss_home/|/home/xsl/change_hair/project/kohya_ss_home/"
-  "onediff|szlc@192.168.101.63:/home/szlc/project/onediff/|/home/xsl/change_hair/project/onediff/"
-  "train_material|szlc@192.168.101.63:/data/train_material/|/home/xsl/change_hair/data/train_material/"
+  "photo_service|szlc@192.168.101.63:/home/szlc/project/photo_service/|/home/ubuntu/change_hair/project/photo_service/"
+  "project_data|szlc@192.168.101.63:/home/szlc/project/data/|/home/ubuntu/change_hair/project/data/"
+  "hair_service_sd|szlc@192.168.101.63:/home/szlc/project/hair_service_sd/|/home/ubuntu/change_hair/project/hair_service_sd/"
+  "conda_py310|szlc@192.168.101.63:/home/szlc/miniconda3/envs/py310/|/home/ubuntu/miniconda3/envs/py310/"
+  "conda_my_hair|szlc@192.168.101.63:/home/szlc/miniconda3/envs/my_hair/|/home/ubuntu/miniconda3/envs/my_hair/"
+  "conda_sdwebui|szlc@192.168.101.63:/home/szlc/miniconda3/envs/sdwebui/|/home/ubuntu/miniconda3/envs/sdwebui/"
+  "kohya_ss_home|szlc@192.168.101.63:/home/szlc/project/kohya_ss_home/|/home/ubuntu/change_hair/project/kohya_ss_home/"
+  "onediff|szlc@192.168.101.63:/home/szlc/project/onediff/|/home/ubuntu/change_hair/project/onediff/"
+  "train_material|szlc@192.168.101.63:/data/train_material/|/home/ubuntu/change_hair/data/train_material/"
 )

 echo "========================================" | tee -a "$LOGDIR/sync_main.log"
@@ -1,8 +1,8 @@
 #!/bin/bash
 # 重新同步 conda 环境（上次被 sed 损坏，这次纯净拷贝）
 set -u
-CONDA_ENVS="/home/xsl/miniconda3/envs"
-LOGDIR="/home/xsl/change_hair/project/logs"
+CONDA_ENVS="/home/ubuntu/miniconda3/envs"
+LOGDIR="/home/ubuntu/change_hair/project/logs"
 mkdir -p "$CONDA_ENVS" "$LOGDIR"

 echo "conda 环境重传开始: $(date '+%H:%M:%S')" | tee -a "$LOGDIR/sync_conda.log"
@@ -2,8 +2,8 @@
 """端到端测试：生发走换发型工作流"""
 import base64, requests, time, sys

-img_path = "/home/xsl/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
-mask_path = "/home/xsl/change_hair/project/logs/hairgrow_test_mask.png"
+img_path = "/home/ubuntu/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
+mask_path = "/home/ubuntu/change_hair/project/logs/hairgrow_test_mask.png"

 with open(img_path, "rb") as f:
    img_b64 = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()
@@ -24,7 +24,7 @@ try:
    d = r.json()
    print(f"HTTP {r.status_code} | state={d.get('state')} | msg={d.get('msg','')} | 耗时={time.time()-t0:.1f}s")
    if d.get("state") == 0:
-        out = "/home/xsl/change_hair/project/logs/test_grow_swap_result.jpg"
+        out = "/home/ubuntu/change_hair/project/logs/test_grow_swap_result.jpg"
        with open(out, "wb") as f:
            f.write(base64.b64decode(d["result"]))
        print(f"✅ 生发成功! 结果图: {out}")
@@ -2,7 +2,7 @@
 """换发色接口测试"""
 import base64, json, requests, sys

-img_path = "/home/xsl/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
+img_path = "/home/ubuntu/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
 with open(img_path, "rb") as f:
    img_b64 = base64.b64encode(f.read()).decode()

@@ -21,7 +21,7 @@ try:
    result = d.get("result", "")
    if d.get("state") == 0 and result:
        # 保存结果图
-        out = "/home/xsl/change_hair/project/logs/test_haircolor_result.jpg"
+        out = "/home/ubuntu/change_hair/project/logs/test_haircolor_result.jpg"
        with open(out, "wb") as f:
            f.write(base64.b64decode(result))
        print(f"✅ 换发色成功! 结果图已保存: {out} ({len(result)} bytes base64)")
@@ -2,8 +2,8 @@
 """测试 /api/hairGrow/v1 接口"""
 import base64, requests, sys

-img_path = "/home/xsl/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
-mask_path = "/home/xsl/change_hair/project/logs/hairgrow_test_mask.png"
+img_path = "/home/ubuntu/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
+mask_path = "/home/ubuntu/change_hair/project/logs/hairgrow_test_mask.png"

 with open(img_path, "rb") as f:
    img_b64 = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()
@@ -26,7 +26,7 @@ try:
    print(f"HTTP {r.status_code} | msg={d.get('msg')} | state={d.get('state')} | 耗时={time.time()-t0:.1f}s")
    result = d.get("result", "")
    if d.get("state") == 0 and result:
-        out = "/home/xsl/change_hair/project/logs/test_hairgrow_api_result.jpg"
+        out = "/home/ubuntu/change_hair/project/logs/test_hairgrow_api_result.jpg"
        with open(out, "wb") as f:
            f.write(base64.b64decode(result))
        print(f"✅ 生发成功! 结果图: {out}")
@@ -2,7 +2,7 @@
 """换发型接口测试"""
 import base64, json, requests, sys, time

-img_path = "/home/xsl/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
+img_path = "/home/ubuntu/change_hair/project/data/userImage/8488902485_20250630055547.jpg"
 with open(img_path, "rb") as f:
    img_b64 = base64.b64encode(f.read()).decode()

@@ -22,7 +22,7 @@ try:
    print(f"HTTP {r.status_code} | msg={d.get('msg')} | state={d.get('state')} | 耗时={time.time()-t0:.1f}s")
    data = d.get("data", "")
    if d.get("state") == 0 and data:
-        out = "/home/xsl/change_hair/project/logs/test_swaphair_result.jpg"
+        out = "/home/ubuntu/change_hair/project/logs/test_swaphair_result.jpg"
        with open(out, "wb") as f:
            f.write(base64.b64decode(data))
        print(f"✅ 换发型成功! 结果图已保存: {out} ({len(data)} bytes base64)")
@@ -4,7 +4,7 @@ import requests, json

 # 用一个已有完整训练素材的发型
 HAIR_ID = "1905785164224868354"
-MATERIAL_DIR = f"/home/xsl/change_hair/data/train_material/{HAIR_ID}"
+MATERIAL_DIR = f"/home/ubuntu/change_hair/data/train_material/{HAIR_ID}"

 payload = {
    "task_id": f"test_train_{HAIR_ID}",
@@ -15,7 +15,7 @@ import time
 import json
 import shutil

-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 os.chdir(HAIR_SERVICE_DIR)
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -26,15 +26,15 @@ import requests as req

 # (hair_id, gender, template_img)
 HAIRSTYLES = [
-    ("chang_tuoyuan", "girl", "/home/xsl/change_hair/hair_type_images/chang_tuoyuan/chang_tuoyuan.jpg"),
-    ("chang_bolang",  "girl", "/home/xsl/change_hair/hair_type_images/chang_bolang/chang_bolang.jpg"),
-    ("chang_zhixian", "girl", "/home/xsl/change_hair/hair_type_images/chang_zhixian/chang_zhixian.jpg"),
-    ("chang_huaban",  "girl", "/home/xsl/change_hair/hair_type_images/chang_huaban/chang_huaban.jpg"),
-    ("chang_xinxing", "girl", "/home/xsl/change_hair/hair_type_images/chang_xinxing/chang_xinxing.jpg"),
+    ("chang_tuoyuan", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_tuoyuan/chang_tuoyuan.jpg"),
+    ("chang_bolang",  "girl", "/home/ubuntu/change_hair/hair_type_images/chang_bolang/chang_bolang.jpg"),
+    ("chang_zhixian", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_zhixian/chang_zhixian.jpg"),
+    ("chang_huaban",  "girl", "/home/ubuntu/change_hair/hair_type_images/chang_huaban/chang_huaban.jpg"),
+    ("chang_xinxing", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_xinxing/chang_xinxing.jpg"),
 ]

 HAIR_CALLBACK = "http://127.0.0.1:8801/api/hair/trainCallBack"
-UPLOAD_TRAIN_DIR = "/home/xsl/change_hair/project/data/upload_train_imgs"
+UPLOAD_TRAIN_DIR = "/home/ubuntu/change_hair/project/data/upload_train_imgs"

 # 导入 step4/step5 函数（会触发模型加载）
 from train_hairstyle_full import step4_template_material, step5_preview
@@ -6,8 +6,8 @@
 串接 prepare_train_data / gen_template_material / photo_service训练 / 回调 / 预览图

 用法:
-  cd /home/xsl/change_hair/project/hair_service_sd
-  python /home/xsl/change_hair/train_hairstyle_full.py --hair-id huaban1 --input /home/xsl/change_hair/huaban1 --gender girl --template-img /home/xsl/change_hair/huaban1/huaban1.jpg
+  cd /home/ubuntu/change_hair/project/hair_service_sd
+  python /home/ubuntu/change_hair/train_hairstyle_full.py --hair-id huaban1 --input /home/ubuntu/change_hair/huaban1 --gender girl --template-img /home/ubuntu/change_hair/huaban1/huaban1.jpg
 """
 import os
 import sys
@@ -21,7 +21,7 @@ import argparse
 import requests as req

 ssl._create_default_https_context = ssl._create_unverified_context
-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 os.chdir(HAIR_SERVICE_DIR)
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -48,9 +48,9 @@ RESOLUTIONS = [512, 768, 1024, 1280, 1536]
 PHOTO_TRAIN = "http://127.0.0.1:32678/api/hair/train"
 HAIR_CALLBACK = "http://127.0.0.1:8801/api/hair/trainCallBack"
 SWAP_API = "http://127.0.0.1:8801/api/swapHair/v1"
-GIRL_IMG = "/home/xsl/change_hair/images/girl.png"
-BOY_IMG = "/home/xsl/change_hair/images/boy.png"
-PREVIEW_DIR = "/home/xsl/change_hair/hair_grow_service/static/previews"
+GIRL_IMG = "/home/ubuntu/change_hair/images/girl.png"
+BOY_IMG = "/home/ubuntu/change_hair/images/boy.png"
+PREVIEW_DIR = "/home/ubuntu/change_hair/hair_grow_service/static/previews"

 _detector = _aligner = _matte = None
 def get_models():
@@ -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,
@@ -8,8 +8,8 @@
  阶段C: 启动服务后串行 step3(回调生成ref材质) + step4(模板) + step5(预览)

 用法:
-  cd /home/xsl/change_hair/project/hair_service_sd
-  python /home/xsl/change_hair/train_hairstyles_parallel.py
+  cd /home/ubuntu/change_hair/project/hair_service_sd
+  python /home/ubuntu/change_hair/train_hairstyles_parallel.py
 """
 import os
 import sys
@@ -18,7 +18,7 @@ import json
 import subprocess
 from concurrent.futures import ThreadPoolExecutor, as_completed

-HAIR_SERVICE_DIR = "/home/xsl/change_hair/project/hair_service_sd"
+HAIR_SERVICE_DIR = "/home/ubuntu/change_hair/project/hair_service_sd"
 os.chdir(HAIR_SERVICE_DIR)
 sys.path.insert(0, HAIR_SERVICE_DIR)
 os.environ.setdefault("HF_HUB_OFFLINE", "1")
@@ -29,11 +29,11 @@ os.environ.setdefault("CUDA_VISIBLE_DEVICES", "0")
 # ====== 配置：本次要训练的发型 ======
 # (hair_id, input_dir, gender, template_img)
 HAIRSTYLES = [
-    ("chang_tuoyuan", "/home/xsl/change_hair/hair_type_images/chang_tuoyuan", "girl", "/home/xsl/change_hair/hair_type_images/chang_tuoyuan/chang_tuoyuan.jpg"),
-    ("chang_bolang",  "/home/xsl/change_hair/hair_type_images/chang_bolang",  "girl", "/home/xsl/change_hair/hair_type_images/chang_bolang/chang_bolang.jpg"),
-    ("chang_zhixian", "/home/xsl/change_hair/hair_type_images/chang_zhixian", "girl", "/home/xsl/change_hair/hair_type_images/chang_zhixian/chang_zhixian.jpg"),
-    ("chang_huaban",  "/home/xsl/change_hair/hair_type_images/chang_huaban",  "girl", "/home/xsl/change_hair/hair_type_images/chang_huaban/chang_huaban.jpg"),
-    ("chang_xinxing", "/home/xsl/change_hair/hair_type_images/chang_xinxing", "girl", "/home/xsl/change_hair/hair_type_images/chang_xinxing/chang_xinxing.jpg"),
+    ("chang_tuoyuan", "/home/ubuntu/change_hair/hair_type_images/chang_tuoyuan", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_tuoyuan/chang_tuoyuan.jpg"),
+    ("chang_bolang",  "/home/ubuntu/change_hair/hair_type_images/chang_bolang",  "girl", "/home/ubuntu/change_hair/hair_type_images/chang_bolang/chang_bolang.jpg"),
+    ("chang_zhixian", "/home/ubuntu/change_hair/hair_type_images/chang_zhixian", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_zhixian/chang_zhixian.jpg"),
+    ("chang_huaban",  "/home/ubuntu/change_hair/hair_type_images/chang_huaban",  "girl", "/home/ubuntu/change_hair/hair_type_images/chang_huaban/chang_huaban.jpg"),
+    ("chang_xinxing", "/home/ubuntu/change_hair/hair_type_images/chang_xinxing", "girl", "/home/ubuntu/change_hair/hair_type_images/chang_xinxing/chang_xinxing.jpg"),
 ]
 PARALLEL = 2  # LoRA 训练并发数

@@ -137,7 +137,7 @@ def main():
    log(f"阶段B完成: LoRA 训练完成 {len(completed)}/{len(step1_ok)}")

    # 写一个完成清单文件，供阶段C脚本读取
-    done_file = "/home/xsl/change_hair/project/logs/train_batch_done.json"
+    done_file = "/home/ubuntu/change_hair/project/logs/train_batch_done.json"
    with open(done_file, "w") as f:
        json.dump({"completed": sorted(completed), "all": ids}, f)
    log(f"已写入完成清单: {done_file}")
@@ -6,7 +6,7 @@
 通过 ThreadPoolExecutor 控制并发数。

 用法:
-  python /home/xsl/change_hair/train_lora_parallel.py
+  python /home/ubuntu/change_hair/train_lora_parallel.py

 前提：step1 训练数据已准备好（data/train_material/<hid>/images/1_hairstyle/*.png）
 """
@@ -16,11 +16,11 @@ import time
 import subprocess
 from concurrent.futures import ThreadPoolExecutor, as_completed

-TRAIN_DIR = "/home/xsl/change_hair/data/train_material"
-KOHYA_WORKDIR = "/home/xsl/change_hair/project/kohya_ss_home/kohya_ss"
-BASE_MODEL = "/home/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors"
-KOHYA_ACCEL = "/home/xsl/miniconda3/envs/kohya/bin/accelerate"
-LOG_DIR = "/home/xsl/change_hair/project/logs"
+TRAIN_DIR = "/home/ubuntu/change_hair/data/train_material"
+KOHYA_WORKDIR = "/home/ubuntu/change_hair/project/kohya_ss_home/kohya_ss"
+BASE_MODEL = "/home/ubuntu/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors"
+KOHYA_ACCEL = "/home/ubuntu/miniconda3/envs/kohya/bin/accelerate"
+LOG_DIR = "/home/ubuntu/change_hair/project/logs"

 # 本次要训练的发型
 HAIRSTYLES = ["chang_tuoyuan", "chang_zhixian", "chang_huaban", "chang_xinxing"]
Author	SHA1	Message	Date
xsl	326b206d08	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	2026-07-26 16:49:21 +08:00
xsl	990fae929f	perf(swapHair): webui降步数 + 用户图预处理内存缓存优化1 - webui img2img steps 20→15 (gen_super_image.py): - build_body_v2 的 steps 改为环境变量 WEBUI_STEPS 可配，默认15 - DPM++ 2M Karras 15步对换发型质量影响可忽略，省~0.1s 优化2 - infer_hairstyle_diy_jy 用户图预处理内存缓存 (hairstyle_model.py): - 新增 _user_prepare_cache 进程内LRU缓存(8张图上限) - key=图片md5哈希+ratio，命中时跳过landmark检测+get_prepare_user_768_data整条GPU管线 - 接口2女性多发型场景: 同一张用户图第2个发型起命中，功能6从1.8s降至1.1s(省0.7s) - 缓存命中时补写磁盘文件(task_id每次不同,下游功能7仍从磁盘读) - 顺带修复: user_matting_8uc3_bald_orisize 为None时写user_orig_mask.png的crash - 加分步计时日志(主GAN/融合耗时),便于定位热点实测: 同图连续请求 swapHair 从4.28s降至3.5s(省18%)	2026-07-26 16:41:19 +08:00
xslandCursor	c5e50de40a	chore: 固化 WebUI 底模为 v1-5，并纳入仓库配置副本本机无 majicmix，将 sd_model_checkpoint 固定为 v1-5-pruned-emaonly；因 onediff/webui 是独立 git 仓无法直接跟踪，配置副本放在 configs/。 Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-19 15:52:51 +08:00
xslandCursor	0926c61bd0	fix: 适配 ubuntu 路径并增强换发色/训练流程稳定性将配置与训练脚本从 /home/xsl 切到本机 /home/ubuntu；换发色在 webui 增强失败或缺色板时降级返回，训练结束后自动重启 hair 服务再回调。 Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-19 00:49:16 +08:00
xsl	fc66534a74	feat: 系统配置文件适配 RTX 3090 添加系统服务配置文件： - change_hair-hair.service: hair_service_sd swapHair (8801) - change_hair-photo.service: photo_service LoRA scheduler (32678) - change_hair-webui.service: SD WebUI (57860)	2026-07-18 19:18:56 +08:00
xsl	6f34e8876c	feat: 适配 RTX 3090 (24GB) 环境优化硬件迁移：从 RTX 5090 (32GB) 迁移到 RTX 3090 (24GB) 主要改动： - 所有启动脚本和配置文件中的 /home/xsl/ 路径替换为 /home/ubuntu/ - 适配新的 24GB VRAM 环境	2026-07-18 19:00:15 +08:00