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 |
-| 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`）
 

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
+}

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

@@ -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
-

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/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/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

@@ -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} '

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,