diff --git a/README.md b/README.md
index 06f652a..28f1138 100644
--- a/README.md
+++ b/README.md
@@ -100,6 +100,14 @@ English (TODO) [google translate](https://github-com.translate.goog/liasece/sd-w
 
 在 stable-diffusion-webui 的运行命令行中可以看到训练的过程及进度。
 
+### 使用 SD2 或 SDXL 作为基础模型
+
+请正确勾选 `Base on Stable Diffusion V2` 或者 `Base on Stable Diffusion XL` ，否则会导致训练失败。
+
+### 高级 sd_script 参数追加或覆盖
+
+在 `Append or override the sd_script args.` 文本框中输入参数，务必使用 `--` 开头的参数，例如：`--lr_scheduler="constant_with_warmup" --max_grad_norm=0.0` 。插件代码中会使用 `--` 分隔符作为参数间的分隔符。
+
 ### 一次训练多种参数训练，充分利用你睡觉时的 GPU
 
 有时，一套训练配置可能并不是最优的。等待你的训练结束然后再重新开始训练，这样的效率太低了。因此，你可以一次性配置多种参数，点击一次训练，自动组合不同的参数进行训练。
diff --git a/liasece_sd_webui_train_tools/ArgsList.py b/liasece_sd_webui_train_tools/ArgsList.py
index 8455837..17bd883 100644
--- a/liasece_sd_webui_train_tools/ArgsList.py
+++ b/liasece_sd_webui_train_tools/ArgsList.py
@@ -13,6 +13,7 @@
 import gc
 import torch
 
+from liasece_sd_webui_train_tools.util import *
 import liasece_sd_webui_train_tools.PythonContextWarper as pc
 with pc.PythonContextWarper(
         to_module_path= os.path.abspath(os.path.join(os.path.dirname(__file__), "sd_scripts")), 
@@ -30,7 +31,6 @@ def __init__(self):
         self.save_json_folder: Union[str, None] = None
         self.save_json_name: Union[str, None] = None
         self.load_json_path: Union[str, None] = None
-        self.multi_run_folder: Union[str, None] = None
         self.reg_img_folder: Union[str, None] = None
         self.sample_prompts: Union[str, None] = None  # path to a txt file that has all of the sample prompts in it,
         # one per line. Only goes to 75 tokens, will cut off the rest. Just place the prompts into the txt file per line
@@ -188,6 +188,11 @@ def __init__(self):
         self.locon_dim: Union[int, None] = None  # deprecated
         self.locon_alpha: Union[int, None] = None  # deprecated
         self.locon: bool = False  # deprecated
+        self.use_sdxl: bool = False  # use the sdxl trainer
+        self.no_half_vae: bool = False  # Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
+        self.cache_text_encoder_outputs: bool = False  # Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
+        self.cache_text_encoder_outputs_to_disk: bool = False  # Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
+        self.ext_sd_script_args: str = ""  # Append or override the sd_script args. (e.g. `--lr_scheduler="constant_with_warmup" --max_grad_norm=0.0`)
 
     # Creates the dict that is used for the rest of the code, to facilitate easier json saving and loading
     def convert_args_to_dict(self):
@@ -196,32 +201,11 @@ def convert_args_to_dict(self):
     def create_args(self) -> argparse.Namespace:
         parser = Parser()
         args = self.convert_args_to_dict()
-        multi_path = args['multi_run_folder']
-        if multi_path and ensure_path(multi_path, "multi_run_folder"):
-            for file in os.listdir(multi_path):
-                if os.path.isdir(file) or file.split(".")[-1] != "json":
-                    continue
-                args = self.convert_args_to_dict()
-                args['json_load_skip_list'] = None
-                try:
-                    ensure_file_paths(args)
-                except FileNotFoundError:
-                    print("failed to find one or more folders or paths, skipping.")
-                    continue
-                if args['tag_occurrence_txt_file']:
-                    get_occurrence_of_tags(args)
-                args = parser.create_args(self.change_dict_to_internal_names(args))
-                train_network.train(args)
-                gc.collect()
-                torch.cuda.empty_cache()
-                if not os.path.exists(os.path.join(multi_path, "complete")):
-                    os.makedirs(os.path.join(multi_path, "complete"))
-                os.rename(os.path.join(multi_path, file), os.path.join(multi_path, "complete", file))
-            print("completed all training")
-            quit()
         ensure_file_paths(args)
         if args['tag_occurrence_txt_file']:
             get_occurrence_of_tags(args)
+        if self.use_sdxl:
+            self.no_half_vae = True
         args = parser.create_args(self.change_dict_to_internal_names(args))
         return args
 
@@ -283,7 +267,7 @@ def find_max_steps(args: dict) -> int:
 
 def ensure_file_paths(args: dict) -> None:
     failed_to_find = False
-    folders_to_check = ['img_folder', 'output_folder', 'save_json_folder', 'multi_run_folder',
+    folders_to_check = ['img_folder', 'output_folder', 'save_json_folder',
                         'reg_img_folder', 'log_dir']
     for folder in folders_to_check:
         if folder in args and args[folder] is not None:
@@ -360,15 +344,33 @@ def ensure_path(path, name, ext_list=None) -> bool:
 
 class Parser:
     def __init__(self) -> None:
-        self.parser = train_network.setup_parser()
+        parser = train_network.setup_parser()
+        parser.add_argument(
+            "--cache_text_encoder_outputs", action="store_true", help="cache text encoder outputs / text encoderの出力をキャッシュする"
+        )
+        parser.add_argument(
+            "--cache_text_encoder_outputs_to_disk",
+            action="store_true",
+            help="cache text encoder outputs to disk / text encoderの出力をディスクにキャッシュする",
+        )
+        self.parser = parser
 
     def create_args(self, args: dict) -> argparse.Namespace:
         remove_epochs = False
-        args_list = []
-        skip_list = ["save_json_folder", "load_json_path", "multi_run_folder", "json_load_skip_list",
+        args_list: list[str] = []
+        skip_list = ["save_json_folder", "load_json_path", "json_load_skip_list",
                      "tag_occurrence_txt_file", "sort_tag_occurrence_alphabetically", "save_json_only",
                      "warmup_lr_ratio", "optimizer_args", "locon_dim", "locon_alpha", "locon", "lyco", "network_args",
-                     "resolution", "height_resolution"]
+                     "resolution", "height_resolution", "use_sdxl", "ext_sd_script_args"]
+
+        # decode ext_sd_script_args
+        if "ext_sd_script_args" in args and args["ext_sd_script_args"]:
+            ext_sd_script_args = args["ext_sd_script_args"].split("--")
+            for arg in ext_sd_script_args:
+                if not arg:
+                    continue
+                args_list.append(f"--{arg}")
+
         for key, value in args.items():
             if not value:
                 continue
@@ -376,6 +378,15 @@ def create_args(self, args: dict) -> argparse.Namespace:
                 continue
             if key == "max_train_steps":
                 remove_epochs = True
+            # check key is in the parser
+            already_exists = False
+            for arg in args_list:
+                if arg.startswith(f"--{key}"):
+                    already_exists = True
+                    break
+            if already_exists:
+                printD(f"Skipping {key} as it already exists")
+                continue
             if isinstance(value, bool):
                 args_list.append(f"--{key}")
             else:
diff --git a/liasece_sd_webui_train_tools/config_file.py b/liasece_sd_webui_train_tools/config_file.py
index c10942e..1f8b1d9 100644
--- a/liasece_sd_webui_train_tools/config_file.py
+++ b/liasece_sd_webui_train_tools/config_file.py
@@ -35,12 +35,14 @@
     {"train_finish_generate_all_checkpoint_preview": True},
     {"train_optimizer_type": ["Lion"]},
     {"train_learning_rate": "0.0001"},
+    {"sd_script_args": ""},
     {"train_net_dim": 128},
     {"train_alpha": 64},
     {"train_clip_skip": 2},
     {"train_mixed_precision": "fp16"},
     {"train_xformers": True},
     {"train_base_on_sd_v2": False},
+    {"use_sdxl": False},
     # preview
     {"preview_include_sub_img": False},
     {"preview_txt2img_prompt": "best quality,Amazing,finely detail,extremely detailed CG unity 8k wallpaper"},
diff --git a/liasece_sd_webui_train_tools/sd_scripts b/liasece_sd_webui_train_tools/sd_scripts
index 4c6f312..e69d341 160000
--- a/liasece_sd_webui_train_tools/sd_scripts
+++ b/liasece_sd_webui_train_tools/sd_scripts
@@ -1 +1 @@
-Subproject commit 4c6f3125fce31f86aa949026030e923f3d7d826b
+Subproject commit e69d34103bfd588958cf26bf98694c06cb8976aa
diff --git a/liasece_sd_webui_train_tools/train.py b/liasece_sd_webui_train_tools/train.py
index db258d8..32cecd4 100644
--- a/liasece_sd_webui_train_tools/train.py
+++ b/liasece_sd_webui_train_tools/train.py
@@ -6,9 +6,11 @@
 import pkg_resources
 
 from liasece_sd_webui_train_tools.ArgsList import ArgStore
+from liasece_sd_webui_train_tools.util import *
 from modules import script_loading
 
 import liasece_sd_webui_train_tools.sd_scripts.train_network as train_network
+import liasece_sd_webui_train_tools.sd_scripts.sdxl_train_network as sdxl_train_network
 
 import liasece_sd_webui_train_tools.PythonContextWarper as pc
 import liasece_sd_webui_train_tools.util as util
@@ -34,4 +36,9 @@ def train(cfg: ArgStore) -> None:
             sub_module=["library", "networks"],
         ):
         # begin training
-        train_network.train(args)
+        if cfg.use_sdxl:
+            trainer = sdxl_train_network.SdxlNetworkTrainer()
+        else:
+            trainer = train_network.NetworkTrainer()
+        printD("train begin", args)
+        trainer.train(args)
diff --git a/liasece_sd_webui_train_tools/train_ui.py b/liasece_sd_webui_train_tools/train_ui.py
index f7a9c97..1ff6d7e 100644
--- a/liasece_sd_webui_train_tools/train_ui.py
+++ b/liasece_sd_webui_train_tools/train_ui.py
@@ -31,12 +31,14 @@ def on_train_begin_click(id: str, project: str, version: str,
         train_finish_generate_all_checkpoint_preview: bool,
         train_optimizer_type: list[str],
         train_learning_rate: str,
+        sd_script_args: str,
         train_net_dim: int,
         train_alpha: int,
         train_clip_skip: int,
         train_mixed_precision: str,
         train_xformers: bool,
         train_base_on_sd_v2: bool,
+        use_sdxl: bool, # use sdxl
         # preview view config
         preview_include_sub_img: bool,
         # txt2txt
@@ -61,12 +63,14 @@ def on_train_begin_click(id: str, project: str, version: str,
         "train_finish_generate_all_checkpoint_preview": train_finish_generate_all_checkpoint_preview,
         "train_optimizer_type": train_optimizer_type,
         "train_learning_rate": train_learning_rate,
+        "sd_script_args": sd_script_args,
         "train_net_dim": int(train_net_dim),
         "train_alpha": int(train_alpha),
         "train_clip_skip": int(train_clip_skip),
         "train_mixed_precision": train_mixed_precision,
         "train_xformers": train_xformers,
         "train_base_on_sd_v2": train_base_on_sd_v2,
+        "use_sdxl": use_sdxl,
     })
     save_preview_config(project, version, {
         # preview view config
@@ -119,6 +123,8 @@ def on_train_begin_click(id: str, project: str, version: str,
             cfg.mixed_precision = train_mixed_precision
             cfg.xformers = train_xformers
             cfg.v2 = train_base_on_sd_v2
+            cfg.use_sdxl = use_sdxl
+            cfg.ext_sd_script_args = sd_script_args
             # check if reg path exist
             if os.path.exists(os.path.join(processed_path, "..", "reg")):
                 cfg.reg_img_folder = os.path.abspath(os.path.join(processed_path, "..", "reg"))
diff --git a/liasece_sd_webui_train_tools/ui.py b/liasece_sd_webui_train_tools/ui.py
index 0222d2f..a72e130 100644
--- a/liasece_sd_webui_train_tools/ui.py
+++ b/liasece_sd_webui_train_tools/ui.py
@@ -130,6 +130,7 @@ def new_ui():
                             train_base_model_refresh_button = ui.ToolButton(value=ui.refresh_symbol, elem_id="train_base_model_refresh_button")
                         with gr.Row():
                             train_base_on_sd_v2 = gr.Checkbox(label="Base on Stable Diffusion V2", value=False, elem_id="train_base_on_sd_v2", interactive = True)
+                            use_sdxl = gr.Checkbox(label="Base on Stable Diffusion XL", value=False, elem_id="use_sdxl", interactive = True)
                         with gr.Row():
                             train_xformers = gr.Checkbox(label="Use xformers", value=True, elem_id="train_xformers", interactive = True)
                         with gr.Row():
@@ -139,11 +140,12 @@ def new_ui():
                     with gr.Column():
                         train_batch_size = gr.Number(value=1, label="Batch size", elem_id="train_batch_size", interactive = True)
                         train_num_epochs = gr.Number(value=40, label="Number of epochs", elem_id="train_num_epochs", interactive = True)
-                        train_learning_rate = gr.Textbox(value="0.0001", label="Learning rate", elem_id="train_learning_rate", interactive = True)
+                        train_learning_rate = gr.Textbox(value="0.0001", label="Learning rate(Multi-select e.g. 0.0001,0.0002)", elem_id="train_learning_rate", interactive = True)
+                        sd_script_args = gr.Textbox(value="", label="Append or override the sd_script args. (e.g. `--lr_scheduler=\"constant_with_warmup\" --max_grad_norm=0.0`)", elem_id="sd_script_args", interactive = True)
                     with gr.Column():
                         train_net_dim = gr.Number(value=128, label="Net dim (128 ~ 144MB)", elem_id="train_net_dim", interactive = True)
                         train_alpha = gr.Number(value=64, label="Alpha (default is half of Net dim)", elem_id="train_alpha", interactive = True)
-                        train_optimizer_type = gr.Dropdown(label="Optimizer type",value=["Lion"], choices=["Adam", "AdamW", "AdamW8bit", "Lion", "SGDNesterov", "SGDNesterov8bit", "DAdaptation", "AdaFactor"], multiselect = True, interactive = True, elem_id="train_optimizer_type")
+                        train_optimizer_type = gr.Dropdown(label="Optimizer type(Multi-select)",value=["Lion"], choices=["Adam", "AdamW", "AdamW8bit", "Lion", "SGDNesterov", "SGDNesterov8bit", "DAdaptation", "AdaFactor"], multiselect = True, interactive = True, elem_id="train_optimizer_type")
                         train_mixed_precision = gr.Dropdown(label="Mixed precision (If your graphics card supports bf16 better)",value="fp16", choices=["fp16", "bf16"], interactive = True, elem_id="train_mixed_precision")
                 with gr.Row():
                     with gr.Column(scale=2):
@@ -248,12 +250,14 @@ def train_config_inputs():
                 train_finish_generate_all_checkpoint_preview,
                 train_optimizer_type,
                 train_learning_rate,
+                sd_script_args,
                 train_net_dim,
                 train_alpha,
                 train_clip_skip,
                 train_mixed_precision,
                 train_xformers,
                 train_base_on_sd_v2,
+                use_sdxl,
             ]
         def preview_config_inputs():
             return [
diff --git a/requirements.txt b/requirements.txt
index d33444d..f146bbc 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,10 +1,10 @@
-accelerate==0.15.0
-transformers
+accelerate==0.25.0
+diffusers[torch]==0.25.0
+transformers==4.36.2
 ftfy
 albumentations
 opencv-python
 einops
-diffusers[torch]==0.10.2
 pytorch-lightning
 bitsandbytes
 tensorboard