fix: 交互逻辑重构

app.py

@@ -70,7 +70,6 @@ class Demo:
         self.generating = False
         self.weight_dtype = torch.bfloat16
         self.model_sections = []
-        self.model_sections_count = gr.State(0)
         
         model_id = 'stabilityai/stable-diffusion-xl-base-1.0'
         if torch.cuda.is_available():
@@ -94,185 +93,152 @@ class Demo:
         self.num_inference_steps = 3
         
         with gr.Blocks() as demo:
-            @gr.render(inputs=self.model_sections_count)
-            def render_model_sections(count):
-                print("Adding model section")
-                for i in range(count):
-                    # 创建新的模块布局
-                    with gr.Row(visible=True) as new_section:
-                        with gr.Column():
-                            gr.Markdown(f"### 模块 {i + 1}")
-                            model_dropdown = gr.Dropdown(
-                                label="预训练模型选择",
-                                choices=list(model_map.keys()),
-                                value="年龄调整",
-                                interactive=True,
-                            )
-                            seed_infr = gr.Number(
-                                label="种子值",
-                                value=42753,
-                            )
-                            slider_scale_infr = gr.Slider(
-                                minimum=-4,
-                                maximum=4,
-                                label="编辑强度",
-                                value=3,
-                                info="较大的值会导致更强的编辑效果",
-                            )
-                            # del_btn = gr.Button(value="删除", elem_classes="del-btn")
-                            
-                            # # 删除按钮的逻辑
-                            # def delete_section():
-                            #     if new_section in self.model_sections:
-                            #         self.model_sections.remove(new_section)
-                            #     return gr.update(visible=False)
-
-                            # del_btn.click(delete_section, outputs=new_section)
-                    # 添加新模块到 sections 列表
-                    self.model_sections.append(new_section)
-
             self.layout()
             demo.queue(max_size=5).launch(share=True, max_threads=2)
         
 
     def layout(self):
-        with gr.Row():
+        # with gr.Row():
 
-            if SPACE_ID == ORIGINAL_SPACE_ID:
+        #     if SPACE_ID == ORIGINAL_SPACE_ID:
 
-                self.warning = gr.Markdown(SHARED_UI_WARNING)
+        #         self.warning = gr.Markdown(SHARED_UI_WARNING)
           
-        with gr.Row():
+        # with gr.Row():
                 
-            with gr.Tab("测试") as inference_column:
+        with gr.Tab("测试") as inference_column:
 
-                with gr.Row():
+            with gr.Row():
 
-                    self.explain_infr = gr.Markdown(value='这是[概念滑块：用于扩散模型的LoRA适配器](https://sliders.baulab.info/)的演示。要尝试可以控制特定概念的模型，请选择一个模型并输入任何提示词，选择一个种子值，最后选择SDEdit时间步以保持结构。较高的SDEdit时间步会导致更多的结构变化。例如，如果选择“惊讶表情”模型，可以生成提示词“A picture of a person, realistic, 8k”的图像，并将滑块效果与原始模型生成的图像进行比较。我们还提供了几个其他预先微调的模型，如“修复”滑块，用于修复SDXL生成图像中的缺陷（请查看“预训练滑块”下拉菜单）。您还可以训练和运行自己的自定义滑块。请查看“训练”部分以进行自定义概念滑块训练。<b>当前推理正在运行SDXL Turbo！</b>')
+                self.explain_infr = gr.Markdown(value='这是[概念滑块：用于扩散模型的LoRA适配器](https://sliders.baulab.info/)的演示。要尝试可以控制特定概念的模型，请选择一个模型并输入任何提示词，选择一个种子值，最后选择SDEdit时间步以保持结构。较高的SDEdit时间步会导致更多的结构变化。例如，如果选择“惊讶表情”模型，可以生成提示词“A picture of a person, realistic, 8k”的图像，并将滑块效果与原始模型生成的图像进行比较。我们还提供了几个其他预先微调的模型，如“修复”滑块，用于修复SDXL生成图像中的缺陷（请查看“预训练滑块”下拉菜单）。您还可以训练和运行自己的自定义滑块。请查看“训练”部分以进行自定义概念滑块训练。<b>当前推理正在运行SDXL Turbo！</b>')
 
-                with gr.Row():
+            with gr.Row():
 
-                    self.prompt_input_infr = gr.Text(
-                        placeholder="photo of a person, with bokeh street background, realistic, 8k",
-                        label="提示词",
-                        info="生成图像的提示词",
-                        value="photo of a person, with bokeh street background, realistic, 8k"
-                    )
+                self.prompt_input_infr = gr.Text(
+                    placeholder="photo of a person, with bokeh street background, realistic, 8k",
+                    label="提示词",
+                    info="生成图像的提示词",
+                    value="photo of a person, with bokeh street background, realistic, 8k"
+                )
+            
+            for model_name in model_map.keys():
+                with gr.Row():
+                    model_checkbox = gr.Checkbox(label=model_name, value=False)
+                    seed_infr = gr.Number(label="种子值", value=42753)
+                    slider_scale_infr = gr.Slider(-4, 4, label="滑块刻度", value=3, info="较大的滑块刻度会导致更强的编辑效果")
+                    self.model_sections.append((model_checkbox, seed_infr, slider_scale_infr))
+
+            with gr.Row():
+                self.start_noise_infr = gr.Slider(
+                    600, 900, 
+                    value=750, 
+                    label="SDEdit时间步", 
+                    info="选择较小的值以保持更多结构"
+                )
+                self.model_type = gr.Dropdown(
+                    label="模型",
+                    choices=['SDXL Turbo', 'SDXL'],
+                    value='SDXL Turbo',
+                    interactive=True
+                )
+
+            with gr.Row():
+                self.infr_button = gr.Button(
+                    value="生成",
+                    interactive=True
+                )
+
+            with gr.Row():
+                self.image_orig = gr.Image(
+                    label="原始SD",
+                    interactive=False,
+                    type='pil',
+                )
                 
-                self.add_model_button = gr.Button(value="新增模型")
+                self.image_new = gr.Image(
+                    label=f"概念滑块",
+                    interactive=False,
+                    type='pil',
+                )
 
-                with gr.Row():
-                    self.start_noise_infr = gr.Slider(
-                        600, 900, 
-                        value=750, 
-                        label="SDEdit时间步", 
-                        info="选择较小的值以保持更多结构"
+        with gr.Tab("训练") as training_column:
+
+            with gr.Row():
+
+                self.explain_train= gr.Markdown(value='在这一部分，您可以为Stable Diffusion XL训练文本概念滑块。输入您希望进行编辑的目标概念（例如：人）。接下来，输入您希望编辑的属性的增强提示词（例如：控制人的年龄，输入“person, old”）。然后，输入属性的抑制提示词（例如：输入“person, young”）。然后按“训练”按钮。使用默认设置，训练一个滑块大约需要25分钟；然后您可以在上面的“测试”选项卡中尝试推理或下载权重。为了更快的训练，请复制此存储库并使用A100或更大的GPU进行训练。代码和详细信息在[github链接](https://github.com/rohitgandikota/sliders)。')
+
+            with gr.Row():
+
+                with gr.Column(scale=3):
+
+                    self.target_concept = gr.Text(
+                        placeholder="输入要进行编辑的目标概念...",
+                        label="编辑概念的提示词",
+                        info="对应于要编辑的概念的提示词（例如：“person”）",
+                        value = ''
                     )
-                    self.model_type = gr.Dropdown(
-                        label="模型",
-                        choices=['SDXL Turbo', 'SDXL'],
-                        value='SDXL Turbo',
-                        interactive=True
+                    
+                    self.positive_prompt = gr.Text(
+                        placeholder="输入编辑的增强提示词...",
+                        label="增强提示词",
+                        info="对应于要增强的概念的提示词（例如：“person, old”）",
+                        value = ''
                     )
-
-                with gr.Row():
-                    self.infr_button = gr.Button(
-                        value="生成",
+                    
+                    self.negative_prompt = gr.Text(
+                        placeholder="输入编辑的抑制提示词...",
+                        label="抑制提示词",
+                        info="对应于要抑制的概念的提示词（例如：“person, young”）",
+                        value = ''
+                    )
+                    
+                    self.attributes_input = gr.Text(
+                        placeholder="输入要保留的概念（用逗号分隔）。如果不需要，请留空...",
+                        label="要保留的概念",
+                        info="要保留/解缠的概念（例如：“male, female”）",
+                        value = ''
+                    )
+                    self.is_person = gr.Checkbox(
+                        label="人", 
+                        info="您是否在为人训练滑块？")
+
+                    self.rank = gr.Number(
+                        value=4,
+                        label="滑块等级",
+                        info='要训练的滑块等级'
+                    )
+                    choices = ['xattn', 'noxattn']
+                    self.train_method_input = gr.Dropdown(
+                        choices=choices,
+                        value='xattn',
+                        label='训练方法',
+                        info='训练方法。如果[* xattn *] - loras将仅在交叉注意层上。如果[* noxattn *]（官方实现） - 除交叉注意层外的所有层',
                         interactive=True
                     )
-
-                with gr.Row():
-                    self.image_orig = gr.Image(
-                        label="原始SD",
-                        interactive=False,
-                        type='pil',
+                    self.iterations_input = gr.Number(
+                        value=500,
+                        precision=0,
+                        label="迭代次数",
+                        info='用于训练的迭代次数 - 最大为1000'
                     )
-                    
-                    self.image_new = gr.Image(
-                        label=f"概念滑块",
-                        interactive=False,
-                        type='pil',
+
+                    self.lr_input = gr.Number(
+                        value=2e-4,
+                        label="学习率",
+                        info='用于训练的学习率'
                     )
 
-            with gr.Tab("训练") as training_column:
+                with gr.Column(scale=1):
 
-                with gr.Row():
+                    self.train_status = gr.Button(value='', variant='primary', interactive=False)
 
-                    self.explain_train= gr.Markdown(value='在这一部分，您可以为Stable Diffusion XL训练文本概念滑块。输入您希望进行编辑的目标概念（例如：人）。接下来，输入您希望编辑的属性的增强提示词（例如：控制人的年龄，输入“person, old”）。然后，输入属性的抑制提示词（例如：输入“person, young”）。然后按“训练”按钮。使用默认设置，训练一个滑块大约需要25分钟；然后您可以在上面的“测试”选项卡中尝试推理或下载权重。为了更快的训练，请复制此存储库并使用A100或更大的GPU进行训练。代码和详细信息在[github链接](https://github.com/rohitgandikota/sliders)。')
+                    self.train_button = gr.Button(
+                        value="训练",
+                    )
 
-                with gr.Row():
+                    self.download = gr.Files()
+                    self.model_dropdown = gr.Dropdown(choices=list(model_map.keys()))
 
-                    with gr.Column(scale=3):
-
-                        self.target_concept = gr.Text(
-                            placeholder="输入要进行编辑的目标概念...",
-                            label="编辑概念的提示词",
-                            info="对应于要编辑的概念的提示词（例如：“person”）",
-                            value = ''
-                        )
-                        
-                        self.positive_prompt = gr.Text(
-                            placeholder="输入编辑的增强提示词...",
-                            label="增强提示词",
-                            info="对应于要增强的概念的提示词（例如：“person, old”）",
-                            value = ''
-                        )
-                        
-                        self.negative_prompt = gr.Text(
-                            placeholder="输入编辑的抑制提示词...",
-                            label="抑制提示词",
-                            info="对应于要抑制的概念的提示词（例如：“person, young”）",
-                            value = ''
-                        )
-                        
-                        self.attributes_input = gr.Text(
-                            placeholder="输入要保留的概念（用逗号分隔）。如果不需要，请留空...",
-                            label="要保留的概念",
-                            info="要保留/解缠的概念（例如：“male, female”）",
-                            value = ''
-                        )
-                        self.is_person = gr.Checkbox(
-                            label="人", 
-                            info="您是否在为人训练滑块？")
-
-                        self.rank = gr.Number(
-                            value=4,
-                            label="滑块等级",
-                            info='要训练的滑块等级'
-                        )
-                        choices = ['xattn', 'noxattn']
-                        self.train_method_input = gr.Dropdown(
-                            choices=choices,
-                            value='xattn',
-                            label='训练方法',
-                            info='训练方法。如果[* xattn *] - loras将仅在交叉注意层上。如果[* noxattn *]（官方实现） - 除交叉注意层外的所有层',
-                            interactive=True
-                        )
-                        self.iterations_input = gr.Number(
-                            value=500,
-                            precision=0,
-                            label="迭代次数",
-                            info='用于训练的迭代次数 - 最大为1000'
-                        )
-
-                        self.lr_input = gr.Number(
-                            value=2e-4,
-                            label="学习率",
-                            info='用于训练的学习率'
-                        )
-
-                    with gr.Column(scale=1):
-
-                        self.train_status = gr.Button(value='', variant='primary', interactive=False)
-
-                        self.train_button = gr.Button(
-                            value="训练",
-                        )
-
-                        self.download = gr.Files()
-                        self.model_dropdown = gr.Dropdown(choices=list(model_map.keys()))
-
-        self.add_model_button.click(lambda x: x + 1, self.model_sections_count, self.model_sections_count)
         self.infr_button.click(self.inference, inputs=[
             self.prompt_input_infr,
             self.start_noise_infr,
@@ -296,30 +262,6 @@ class Demo:
         ],
         outputs=[self.train_button,  self.train_status, self.download, self.model_dropdown]
         )
-    
-    # def add_model_section(self):
-    #     with self.model_sections:
-    #         model_dropdown = gr.Dropdown(
-    #             label="预训练滑块",
-    #             choices=list(model_map.keys()),
-    #             value='年龄调整',
-    #             interactive=True
-    #         )
-    #         seed_infr = gr.Number(
-    #             label="种子值",
-    #             value=42753
-    #         )
-    #         slider_scale_infr = gr.Slider(
-    #             -4,
-    #             4,
-    #             label="滑块刻度",
-    #             value=3,
-    #             info="较大的滑块刻度会导致更强的编辑效果"
-    #         )
-    #         self.model_sections.add_child(model_dropdown)
-    #         self.model_sections.add_child(seed_infr)
-    #         self.model_sections.add_child(slider_scale_infr)
-    #     return gr.update(visible=True, children=self.model_sections.children)
 
     def train(self, target_concept,positive_prompt, negative_prompt, rank, iterations_input, lr_input, attributes_input, is_person, train_method_input, pbar = gr.Progress(track_tqdm=True)):
         iterations_input = min(int(iterations_input),1000)
@@ -353,10 +295,10 @@ class Demo:
 
     
     def inference(self, prompt, start_noise, model, pbar=gr.Progress(track_tqdm=True)):
-        model_sections = self.model_sections.get_children()
-        model_names = [section[0].value for section in model_sections]
-        seed_list = [section[1].value for section in model_sections]
-        scale_list = [section[2].value for section in model_sections]
+        model_sections = self.model_sections
+        model_names = [section[0].value for section in model_sections if section[0].value]
+        seed_list = [section[1].value for section in model_sections if section[0].value]
+        scale_list = [section[2].value for section in model_sections if section[0].value]
 
         if self.current_model != model:
             if model=='SDXL Turbo':

clean_deps.bat

@@ -0,0 +1,5 @@
+pip freeze > clean_requirements.txt
+
+pip uninstall -r clean_requirements.txt -y
+
+del clean_requirements.txt

install_deps.bat

@@ -1,4 +1,7 @@
 pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
 pip install -U xformers --index-url https://download.pytorch.org/whl/cu124
-pip install -r requirements.txt
+pip install 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.44.1.dev0-py3-none-win_amd64.whl'
+
+pip install -r requirements-win.txt
+
 pip install --upgrade gradio

requirements-win.txt

@@ -0,0 +1,23 @@
+dadaptation==3.1
+diffusers==0.32.2
+ipython==8.7.0
+lion_pytorch==0.1.2
+lpips==0.1.4
+matplotlib==3.6.2
+numpy==1.23.5
+opencv_python==4.5.5.64
+opencv_python_headless==4.7.0.68
+pandas==1.5.2
+Pillow==10.1.0
+prodigyopt==1.0
+pydantic==2.10.5
+PyYAML==6.0.1
+Requests==2.31.0
+safetensors==0.5.2
+tqdm==4.64.1
+transformers==4.48.1
+wandb==0.12.21
+accelerate==1.3.0
+gradio==5.12.0
+gradio_client==1.5.4
+huggingface-hub==0.27.1