add vae slicing/tiling option, fix non random generator

app.py

@@ -87,6 +87,18 @@ def guidance_scale_change(guidance_scale, config):
 
     return config, str(config), assemble_code(config)
 
+def enable_vae_slicing_change(enable_vae_slicing, config):
+
+    config = set_config(config, 'enable_vae_slicing', enable_vae_slicing)
+
+    return config, str(config), assemble_code(config)
+
+def enable_vae_tiling_change(enable_vae_tiling, config):
+
+    config = set_config(config, 'enable_vae_tiling', enable_vae_tiling)
+
+    return config, str(config), assemble_code(config)
+
 def prompt_change(prompt, config):
     
     config = set_config(config, 'prompt', prompt)
@@ -172,7 +184,7 @@ def run_inference(config, config_history, progress=gr.Progress(track_tqdm=True))
         pipeline.scheduler = get_scheduler(config["scheduler"], pipeline.scheduler.config)
         
         if config["manual_seed"] < 0 or config["manual_seed"] is None or config["manual_seed"] == '':
-            generator = torch.Generator(config["device"])
+            generator = None
         else:
             generator = torch.manual_seed(int(config["manual_seed"])) 
         
@@ -247,9 +259,9 @@ with gr.Blocks(analytics_enabled=False) as demo:
         in_schedulers = gr.Dropdown(value="", choices=list(schedulers.keys()), 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("### Adapters")
+    # with gr.Row():
+    #     gr.Markdown('Choose an adapter.')
         
     gr.Markdown("### Inference settings")
     with gr.Row():
@@ -259,6 +271,11 @@ with gr.Blocks(analytics_enabled=False) as demo:
         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")
         in_guidance_scale = gr.Slider(minimum=0, maximum=1, step=0.01, 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.")
+    with gr.Row():
+        gr.Markdown("**VAE** stands for Variational Autoencoders. 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_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("### Output")
     with gr.Row():
@@ -286,6 +303,8 @@ with gr.Blocks(analytics_enabled=False) as demo:
     in_inference_steps.change(inference_steps_change, inputs=[in_inference_steps, config], outputs=[config, out_config, out_code])
     in_manual_seed.change(manual_seed_change, inputs=[in_manual_seed, config], outputs=[config, out_config, out_code])
     in_guidance_scale.change(guidance_scale_change, inputs=[in_guidance_scale, config], outputs=[config, out_config, out_code])
+    in_enable_vae_slicing.change(enable_vae_slicing_change, inputs=[in_enable_vae_slicing, config], outputs=[config, out_config, out_code])
+    in_enable_vae_tiling.change(enable_vae_tiling_change, inputs=[in_enable_vae_tiling, config], outputs=[config, out_config, out_code])
     in_prompt.change(prompt_change, inputs=[in_prompt, config], outputs=[config, out_config, out_code])
     in_trigger_token.change(trigger_token_change, inputs=[in_trigger_token, config], outputs=[config, out_config, out_code])
     in_negative_prompt.change(negative_prompt_change, inputs=[in_negative_prompt, config], outputs=[config, out_config, out_code])
@@ -306,6 +325,8 @@ with gr.Blocks(analytics_enabled=False) as demo:
             in_variant,
             in_safety_checker,
             in_requires_safety_checker,
+            in_enable_vae_slicing,
+            in_enable_vae_tiling,
             in_schedulers,
             in_prompt,
             in_trigger_token,

config.py

@@ -46,12 +46,14 @@ def get_initial_config():
         "refiner": "none",
         "safety_checker": "False",
         "requires_safety_checker": "False",
+        "enable_vae_slicing": "True",
+        "enable_vae_tiling": "True",
         "manual_seed": 42,
         "inference_steps": 10,
         "guidance_scale": 0.5,
         "prompt": 'A white rabbit',
         "trigger_token": '',
-        "negative_prompt": 'lowres, cropped, worst quality, low quality, chat bubble, chat bubbles, ugly',
+        "negative_prompt": 'lowres, cropped, worst quality, low quality',
     }
     
     return config
@@ -86,6 +88,8 @@ def get_config_from_url(initial_config, request: Request):
             return_config['variant'],
             return_config['safety_checker'],
             return_config['requires_safety_checker'],
+            return_config['enable_vae_slicing'],
+            return_config['enable_vae_tiling'],
             return_config['scheduler'],
             return_config['prompt'],
             return_config['trigger_token'],
@@ -147,7 +151,9 @@ def assemble_code(str_config):
             torch_dtype=data_type, 
             variant=variant).to(device)''')
     
-    if str(config["cpu_offload"]).lower() == 'false': code.append("pipeline.enable_model_cpu_offload()")
+    if str(config["cpu_offload"]).lower() != 'false': code.append("pipeline.enable_model_cpu_offload()")
+    if str(config["enable_vae_slicing"]).lower() != 'false': code.append("pipeline.enable_vae_slicing()")
+    if str(config["enable_vae_tiling"]).lower() != 'false': code.append("pipeline.enable_vae_tiling()")
     
     if config['refiner'].lower() != 'none':
         code.append(f'''refiner = DiffusionPipeline.from_pretrained(
@@ -159,7 +165,9 @@ def assemble_code(str_config):
                 variant=variant,
             ).to(device)''')
         
-        if str(config["use_safetensors"]).lower() == 'false': code.append("refiner.enable_model_cpu_offload()")
+        if str(config["cpu_offload"]).lower() != 'false': code.append("refiner.enable_model_cpu_offload()")
+        if str(config["enable_vae_slicing"]).lower() != 'false': code.append("refiner.enable_vae_slicing()")
+        if str(config["enable_vae_tiling"]).lower() != 'false': code.append("refiner.enable_vae_tiling()")
         
     code.append(f'pipeline.requires_safety_checker = {config["requires_safety_checker"]}')
 
@@ -170,7 +178,7 @@ def assemble_code(str_config):
     
     if config['manual_seed'] < 0 or config['manual_seed'] is None or config['manual_seed'] == '':
         code.append(f'# manual_seed = {config["manual_seed"]}')
-        code.append(f'generator = torch.Generator("{config["device"]}")')
+        code.append(f'generator = None')
     else:
         code.append(f'manual_seed = {config["manual_seed"]}')
         code.append(f'generator = torch.manual_seed(manual_seed)')