clean up ui, adding accordeons

app.py

@@ -431,82 +431,90 @@ with gr.Blocks(analytics_enabled=False) as demo:
                 <small>by <a target="_blank" href="https://nickyreinert.de/">Nicky Reinert</a> | 
                 home base: https://huggingface.co/spaces/n42/pictero
                 </small>''')
-    gr.Markdown("### Device specific settings")
+    gr.Markdown("### Device")
+    gr.Markdown("(you may add a custom device address at any time)")
     with gr.Row():
-        in_devices = gr.Dropdown(label="Device:", value=config.value["device"], choices=devices, filterable=True, multiselect=False, allow_custom_value=True, info="(you may add a custom device address at any time)")
-        in_cpu_offload = gr.Radio(label="CPU Offload:", value=config.value["cpu_offload"], choices=["True", "False"], info="This may increase performance, as it offloads computations from the GPU to the CPU. But this can also lead to slower executions and lower effectiveness. Compare running time and outputs before making sure, that this setting will help you")
-    with gr.Row():
-        in_data_type = gr.Radio(label="Data Type:", value=config.value["data_type"], choices=["bfloat16", "float16", "float32"], info="`bfloat16` is not supported on MPS devices right now; `float16` may also not be supported on all devices, Half-precision weights, will save GPU memory, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16")
-        in_allow_tensorfloat32 = gr.Radio(label="Allow TensorFloat32:", value=config.value["allow_tensorfloat32"], choices=["True", "False"], info="is not supported on MPS devices right now; use TensorFloat-32 is faster, but results in slightly less accurate computations, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16 ")
-        in_variant = gr.Radio(label="Variant:", value=config.value["variant"], choices=["fp16", None], info="Use half-precision weights will save GPU memory, not all models support that, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16 ")
-        in_attention_slicing = gr.Radio(label="Attention slicing:", value=config.value["attention_slicing"], choices=["True", "False"], info="Attention operation will be cutted into multiple steps, see https://huggingface.co/docs/diffusers/optimization/mps")
-        
-    gr.Markdown("### Model specific settings")
-    with gr.Row():
-        in_models = gr.Dropdown(choices=list(models.keys()), label="Model")
-        out_model_description = gr.Textbox(value="", label="Description")
+        in_devices = gr.Dropdown(label="Device:", value=config.value["device"], choices=devices, filterable=True, multiselect=False, allow_custom_value=True, info="")
+        gr.Column("")
+        gr.Column("")
+    with gr.Accordion("Device specific settings", open=False):
+        with gr.Row():
+            in_cpu_offload = gr.Radio(label="CPU Offload:", value=config.value["cpu_offload"], choices=["True", "False"], info="This may increase performance, as it offloads computations from the GPU to the CPU. But this can also lead to slower executions and lower effectiveness. Compare running time and outputs before making sure, that this setting will help you")
+            in_data_type = gr.Radio(label="Data Type:", value=config.value["data_type"], choices=["bfloat16", "float16", "float32"], info="`bfloat16` is not supported on MPS devices right now; `float16` may also not be supported on all devices, Half-precision weights, will save GPU memory, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16")
+            in_allow_tensorfloat32 = gr.Radio(label="Allow TensorFloat32:", value=config.value["allow_tensorfloat32"], choices=["True", "False"], info="is not supported on MPS devices right now; use TensorFloat-32 is faster, but results in slightly less accurate computations, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16 ")
+        with gr.Row():
+            in_variant = gr.Radio(label="Variant:", value=config.value["variant"], choices=["fp16", None], info="Use half-precision weights will save GPU memory, not all models support that, see https://huggingface.co/docs/diffusers/main/en/optimization/fp16 ")
+            in_attention_slicing = gr.Radio(label="Attention slicing:", value=config.value["attention_slicing"], choices=["True", "False"], info="Attention operation will be cutted into multiple steps, see https://huggingface.co/docs/diffusers/optimization/mps")
+            gr.Column("")
+
+    gr.Markdown("### Model")
     with gr.Row():
         with gr.Column(scale=1):
+            in_models = gr.Dropdown(choices=list(models.keys()), label="Model")
+        with gr.Column(scale=2):
+            out_model_description = gr.Textbox(value="", label="Description")
+    with gr.Accordion("Model specific settings", open=False):
+        with gr.Row():
             in_trigger_token = gr.Textbox(value=config.value["trigger_token"], label="Trigger Token", info="will be added to your prompt to `activate` a fine tuned model")
             in_model_refiner = gr.Dropdown(value=config.value["refiner"], choices=['none'] + refiners, label="Refiner", allow_custom_value=True, multiselect=False)
-        with gr.Column(scale=1):
+            gr.Column("")
+        with gr.Row():
             in_use_safetensors = gr.Radio(label="Use safe tensors:", choices=["True", "False"], interactive=False)
             in_safety_checker = gr.Radio(label="Enable safety checker:", value=config.value["safety_checker"], choices=["True", "False"])
             in_requires_safety_checker = gr.Radio(label="Requires safety checker:", value=config.value["requires_safety_checker"], choices=["True", "False"])
 
     gr.Markdown("### Scheduler")
+    gr.Markdown("Schedulers employ various strategies for noise control, the scheduler controls parameter adaption between each inference step, depending on the right scheduler for your model, it may only take 10 or 20 steps to achieve very good results, see https://huggingface.co/docs/diffusers/using-diffusers/loading#schedulers")
     with gr.Row():
-        in_schedulers = gr.Dropdown(value="", choices=list(schedulers.keys()), allow_custom_value=True, label="Scheduler/Solver", info="schedulers employ various strategies for noise control, the scheduler controls parameter adaption between each inference step, depending on the right scheduler for your model, it may only take 10 or 20 steps to achieve very good results, see https://huggingface.co/docs/diffusers/using-diffusers/loading#schedulers" )
-        out_scheduler_description = gr.Textbox(value="", label="Description")
-    
-    # gr.Markdown("### Adapters")
-    # with gr.Row():
-    #     gr.Markdown('Choose an adapter.')
-        
-    gr.Markdown("### Auto Encoder")
-    with gr.Row():
-        gr.Markdown("**VAE** stands for Variational Auto Encoders. An 'autoencoder' is an artificial neural network that is able to encode input data and decode to output data to bascially recreate the input. The VAE whereas adds a couple of additional layers of complexity to create new and unique output.")
-    with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=1):
+            in_schedulers = gr.Dropdown(value="", choices=list(schedulers.keys()), allow_custom_value=True, label="Scheduler/Solver", info="")
+        with gr.Column(scale=2):
+            out_scheduler_description = gr.Textbox(value="", label="Description")
+    
+    with gr.Accordion("Auto Encoder", open=False):
+        with gr.Row():
+            gr.Markdown("**VAE** stands for Variational Auto Encoders. An 'autoencoder' is an artificial neural network that is able to encode input data and decode to output data to bascially recreate the input. The VAE whereas adds a couple of additional layers of complexity to create new and unique output.")
+        with gr.Row():
             in_auto_encoders = gr.Dropdown(value="None", choices=list(auto_encoders.keys()), label="Auto encoder", info="leave empty to not add an auto encoder")
             out_auto_encoder_description = gr.Textbox(value="", label="Description")
-        with gr.Column():
+            gr.Column("")
+        with gr.Row():
             in_enable_vae_slicing = gr.Radio(label="Enable VAE slicing:", value=config.value["enable_vae_slicing"], choices=["True", "False"], info="decoding the batches of latents one image at a time, which may reduce memory usage, see https://huggingface.co/docs/diffusers/main/en/optimization/memory")
             in_enable_vae_tiling= gr.Radio(label="Enable VAE tiling:", value=config.value["enable_vae_tiling"], choices=["True", "False"], info="splitting the image into overlapping tiles, decoding the tiles, and then blending the outputs together to compose the final image, see https://huggingface.co/docs/diffusers/main/en/optimization/memory")
-        
-    gr.Markdown("### Adapters")
-    with gr.Row():
-        gr.Markdown('''Adapters allow you to apply finetuned weights to your base model. They come in many flavors depending on how they were trained. See see https://huggingface.co/docs/diffusers/using-diffusers/loading_adapters''')
-    with gr.Row():
-        gr.Markdown('#### Textual Inversion Adapters')
-    with gr.Row():
-        gr.Markdown('(a technique that enables a model like Stable Diffusion to learn a new concept from just a few sample images)')
-    with gr.Row():
-        in_adapters_textual_inversion = gr.Dropdown(value="", choices=list(adapters['textual_inversion'].keys()), label="Textual Inversion Adapter", info="leave empty to not use an adapter")
-        in_adapters_textual_inversion_token = gr.Textbox(value="", label="Token", info="required to activate the token, will be added to your prompt")
-        out_adapters_textual_inversion_description = gr.Textbox(value="", label="Description")
-    with gr.Row():
-        gr.Markdown('#### LoRA')
-    with gr.Row():
-        gr.Markdown('(Low-Rank-Adaption is a performant fine tuning technique)')
-    with gr.Row():
-        in_adapters_lora = gr.Dropdown(value="None", choices=list(adapters['lora'].keys()), multiselect=True, label="LoRA Adapter", info="leave empty to not use an adapter")
-        out_adapters_lora_description = gr.Textbox(value="", label="Description")
-    with gr.Row():
-        in_adapters_lora_token = gr.Textbox(value="None", label="Token(s)", info="required to activate the token, will be added to your prompt")
-        in_adapters_lora_weight = gr.Textbox(value="", label="Weight(s)/Checkpoint(s)")
-        in_adapters_lora_balancing = gr.Textbox(value={}, label="Balancing", info="provide a list of balancing weights in the order of your LoRA adapter (according to `token`s)")
-        
+            gr.Column("")
+            
+    with gr.Accordion("Adapters", open=False):
+        with gr.Row():
+            gr.Markdown('''Adapters allow you to apply finetuned weights to your base model. They come in many flavors depending on how they were trained. See see https://huggingface.co/docs/diffusers/using-diffusers/loading_adapters''')
+        with gr.Row():
+            gr.Markdown('#### Textual Inversion Adapters')
+        with gr.Row():
+            gr.Markdown('(a technique that enables a model like Stable Diffusion to learn a new concept from just a few sample images)')
+        with gr.Row():
+            in_adapters_textual_inversion = gr.Dropdown(value="", choices=list(adapters['textual_inversion'].keys()), label="Textual Inversion Adapter", info="leave empty to not use an adapter")
+            in_adapters_textual_inversion_token = gr.Textbox(value="", label="Token", info="required to activate the token, will be added to your prompt")
+            out_adapters_textual_inversion_description = gr.Textbox(value="", label="Description")
+        with gr.Row():
+            gr.Markdown('#### LoRA')
+        with gr.Row():
+            gr.Markdown('(Low-Rank-Adaption is a performant fine tuning technique)')
+        with gr.Row():
+            in_adapters_lora = gr.Dropdown(value="None", choices=list(adapters['lora'].keys()), multiselect=True, label="LoRA Adapter", info="leave empty to not use an adapter")
+            out_adapters_lora_description = gr.Textbox(value="", label="Description")
+            in_lora_scale = gr.Slider(minimum=0, maximum=1, step=0.1, label="LoRA Scale", value=config.value["lora_scale"], info="How should the LoRA model influence the result, from 0 (no influence) to 1 (full influencer)")
+        with gr.Row():
+            in_adapters_lora_token = gr.Textbox(value="None", label="Token(s)", info="required to activate the token, will be added to your prompt")
+            in_adapters_lora_weight = gr.Textbox(value="", label="Weight(s)/Checkpoint(s)")
+            in_adapters_lora_balancing = gr.Textbox(value={}, label="Balancing", info="provide a list of balancing weights in the order of your LoRA adapter (according to `token`s)")
+            
     gr.Markdown("### Inference settings")
     with gr.Row():
         in_prompt = gr.TextArea(label="Prompt", value=config.value["prompt"])
         in_negative_prompt = gr.TextArea(label="Negative prompt", value=config.value["negative_prompt"])
     with gr.Row():
+        in_guidance_scale = gr.Slider(minimum=0, maximum=100, step=0.1, label="Guidance Scale", value=config.value["guidance_scale"], info="A low guidance scale leads to a faster inference time, with the drawback that negative prompts don’t have any effect on the denoising process.")
         in_inference_steps = gr.Number(label="Inference steps", value=config.value["inference_steps"], info="Each step improves the final result but also results in higher computation")
         in_manual_seed = gr.Number(label="Manual seed", value=config.value["manual_seed"], info="Set this to -1 or leave it empty to randomly generate an image. A fixed value will result in a similar image for every run")
-    with gr.Row():
-        in_guidance_scale = gr.Slider(minimum=0, maximum=100, step=0.1, label="Guidance Scale", value=config.value["guidance_scale"], info="A low guidance scale leads to a faster inference time, with the drawback that negative prompts don’t have any effect on the denoising process.")
-        in_lora_scale = gr.Slider(minimum=0, maximum=1, step=0.1, label="LoRA Scale", value=config.value["lora_scale"], info="How should the LoRA model influence the result, from 0 (no influence) to 1 (full influencer)")
     
     gr.Markdown("### Output")
     with gr.Row():
@@ -517,13 +525,15 @@ with gr.Blocks(analytics_enabled=False) as demo:
         btn_stop_pipeline = gr.Button(value="Stop", variant="stop")
     with gr.Row():
         out_image = gr.Image()
-        out_code = gr.Code(assemble_code(config.value), label="Code")
-    with gr.Row():
-        # out_config = gr.Code(value=str(config.value), label="Current config")
-        out_config = gr.JSON(value=config.value, label="Current config")
-    with gr.Row():
-        out_config_history = gr.Markdown(dict_list_to_markdown_table(config_history.value))
-    
+
+    with gr.Accordion("Code and Configuration", open=False):
+        with gr.Row():
+            out_code = gr.Code(assemble_code(config.value), label="Code")
+            # out_config = gr.Code(value=str(config.value), label="Current config")
+            out_config = gr.JSON(value=config.value, label="Current config")
+        with gr.Row():
+            out_config_history = gr.Markdown(dict_list_to_markdown_table(config_history.value))
+
     # `SPECIAL` CHANGE LISTENERS
     in_models.change(models_change, inputs=[in_models, in_schedulers, config], outputs=[out_model_description, in_trigger_token, in_use_safetensors, in_schedulers, config, out_config, out_code], js="(model, config) => set_model_cookie(model, config)")
     in_schedulers.change(schedulers_change, inputs=[in_schedulers, config], outputs=[out_scheduler_description, config, out_config, out_code], js="(value, config) => set_cookie_2('scheduler', value, config)")