New layout

gradio_demo.py

@@ -83,7 +83,7 @@ def stage1_process(input_image, gamma_correction):
     print('Start stage1_process')
     if torch.cuda.device_count() == 0:
         gr.Warning('Set this space to GPU config to make it work.')
-        return None
+        return None, None
     torch.cuda.set_device(SUPIR_device)
     LQ = HWC3(input_image)
     LQ = fix_resize(LQ, 512)
@@ -98,7 +98,7 @@ def stage1_process(input_image, gamma_correction):
     LQ *= 255.0
     LQ = LQ.round().clip(0, 255).astype(np.uint8)
     print('End stage1_process')
-    return LQ
+    return LQ, gr.update(visible = True)
 
 @spaces.GPU(duration=240)
 def llave_process(input_image, temperature, top_p, qs=None):
@@ -145,7 +145,7 @@ def stage2_process(
     print('Start stage2_process')
     if torch.cuda.device_count() == 0:
         gr.Warning('Set this space to GPU config to make it work.')
-        return None, None
+        return None, None, None
     input_image = noisy_image if denoise_image is None else denoise_image
     torch.cuda.set_device(SUPIR_device)
     event_id = str(time.time_ns())
@@ -201,7 +201,7 @@ def stage2_process(
         for i, result in enumerate(results):
             Image.fromarray(result).save(f'./history/{event_id[:5]}/{event_id[5:]}/HQ_{i}.png')
     print('End stage2_process')
-    return [input_image] + results, [input_image] + results, event_id
+    return [input_image] + results, [input_image] + results, gr.update(visible = True), event_id
 
 def load_and_reset(param_setting):
     print('Start load_and_reset')
@@ -257,7 +257,7 @@ title_html = """
     <big><center>Upscale your images up to x8 freely, without account, without watermark and download it</center></big>
     <br/>
     
-    <p>SUPIR is a practicing model scaling for photo-realistic image restoration. It is still a research project under tested and is not yet a stable commercial product. LLaVa is not integrated in this demo. If you want to auto-generate the description of your image, use another <a href="https://huggingface.co/spaces/MaziyarPanahi/llava-llama-3-8b">LLaVa space</a>. The content added by SUPIR is imagination, not real-world information. The aim of SUPIR is the beauty and the illustration.
+    <p>SUPIR is a practicing model scaling for photo-realistic image restoration. It is still a research project under tested and is not yet a stable commercial product. LlaVa is not integrated in this demo. The content added by SUPIR is imagination, not real-world information. The aim of SUPIR is the beauty and the illustration.
     
     <p><center><a href="https://arxiv.org/abs/2401.13627">Paper</a> &emsp; <a href="http://supir.xpixel.group/">Project Page</a> &emsp; <a href="https://github.com/Fanghua-Yu/SUPIR/blob/master/assets/DemoGuide.png">How to play</a> &emsp; <a href="https://huggingface.co/blog/MonsterMMORPG/supir-sota-image-upscale-better-than-magnific-ai">Local Install Guide</a></center></p>
     """
@@ -283,96 +283,88 @@ with gr.Blocks(title="SUPIR") as interface:
     You can't use SUPIR directly here because this space runs on a CPU, which is not enough for SUPIR. This is a template space. Please provide feedback if you have issues.
     </big></big></big></p>
     """)
-    with gr.Row():
-        gr.HTML(title_html)
-
-    with gr.Row():
-        with gr.Column():
-            with gr.Row(equal_height=True):
-                with gr.Column():
-                    input_image = gr.Image(label="Input", show_label=True, type="numpy", height=600, elem_id="image-input")
-            prompt = gr.Textbox(label="Image description", value="", placeholder="A person, walking, in a town, Summer, photorealistic", lines=3)
-
-            with gr.Accordion("Pre-denoising", open=False):
-                gamma_correction = gr.Slider(label="Gamma Correction", minimum=0.1, maximum=2.0, value=1.0, step=0.1)
-                denoise_button = gr.Button(value="Pre-denoise")
-                denoise_image = gr.Image(label="Denoised image", show_label=True, type="numpy", height=600, elem_id="image-s1")
-                denoise_information = gr.HTML(value="If present, the denoised image will be used for the restoration instead of the input image.")
-
-            with gr.Accordion("LLaVA options", open=False, visible=False):
-                temperature = gr.Slider(label="Temperature", info = "lower=Always similar, higher=More creative", minimum=0., maximum=1.0, value=0.2, step=0.1)
-                top_p = gr.Slider(label="Top P", info = "Percent of tokens shortlisted", minimum=0., maximum=1.0, value=0.7, step=0.1)
-                qs = gr.Textbox(label="Question", info="Ask LLaVa what description you want", value="Describe the image and its style in a very detailed manner. The image is a realistic photography, not an art painting.", lines=3)
-
-            with gr.Accordion("Restoring options", open=False):
-                upscale = gr.Radio([1, 2, 3, 4, 5, 6, 7, 8], label="Upscale factor", info="Resolution x1 to x8", value=2, interactive=True)
-                a_prompt = gr.Textbox(label="Default Positive Prompt",
-                                      info="Describe what the image represents",
-                                      value='Cinematic, High Contrast, highly detailed, taken using a Canon EOS R '
-                                            'camera, hyper detailed photo - realistic maximum detail, 32k, Color '
-                                            'Grading, ultra HD, extreme meticulous detailing, skin pore detailing, '
-                                            'hyper sharpness, perfect without deformations.',
-                                      lines=3)
-                n_prompt = gr.Textbox(label="Default Negative Prompt",
-                                      info="List what the image does NOT represent",
-                                      value='painting, oil painting, illustration, drawing, art, sketch, anime, '
-                                            'cartoon, CG Style, 3D render, unreal engine, blurring, bokeh, ugly, dirty, messy, '
-                                            'worst quality, low quality, frames, watermark, signature, jpeg artifacts, '
-                                            'deformed, lowres, over-smooth',
-                                      lines=3)
-                num_samples = gr.Slider(label="Num Samples", info="Number of generated results; I discourage to increase because the process is limited to 4 min", minimum=1, maximum=4 if not args.use_image_slider else 1
-                                        , value=1, step=1)
-                edm_steps = gr.Slider(label="Steps", info="lower=faster, higher=more details", minimum=1, maximum=200, value=default_setting.edm_steps if torch.cuda.device_count() > 0 else 1, step=1)
-                with gr.Row():
-                    with gr.Column():
-                        model_select = gr.Radio(["v0-Q", "v0-F"], label="Model Selection", info="Q=Quality, F=Fidelity", value="v0-Q",
-                                                interactive=True)
-                    with gr.Column():
-                        color_fix_type = gr.Radio(["None", "AdaIn", "Wavelet"], label="Color-Fix Type", info="AdaIn=Adaptive Instance Normalization, Wavelet=For JPEG artifacts", value="Wavelet",
-                                                  interactive=True)
-                s_cfg = gr.Slider(label="Text Guidance Scale", info="lower=follow the image, higher=follow the prompt", minimum=1.0, maximum=15.0,
-                                  value=default_setting.s_cfg_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.1)
-                s_stage2 = gr.Slider(label="Restoring Guidance Strength", minimum=0., maximum=1., value=1., step=0.05)
-                s_stage1 = gr.Slider(label="Pre-denoising Guidance Strength", minimum=-1.0, maximum=6.0, value=-1.0, step=1.0)
-                s_churn = gr.Slider(label="S-Churn", minimum=0, maximum=40, value=5, step=1)
-                s_noise = gr.Slider(label="S-Noise", minimum=1.0, maximum=1.1, value=1.003, step=0.001)
-                with gr.Row():
-                    with gr.Column():
-                        linear_CFG = gr.Checkbox(label="Linear CFG", value=True)
-                        spt_linear_CFG = gr.Slider(label="CFG Start", minimum=1.0,
-                                                        maximum=9.0, value=default_setting.spt_linear_CFG_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.5)
-                    with gr.Column():
-                        linear_s_stage2 = gr.Checkbox(label="Linear Restoring Guidance", value=False)
-                        spt_linear_s_stage2 = gr.Slider(label="Guidance Start", minimum=0.,
-                                                        maximum=1., value=0., step=0.05)
-                with gr.Row():
-                    with gr.Column():
-                        diff_dtype = gr.Radio(['fp32', 'fp16', 'bf16'], label="Diffusion Data Type", value="fp16",
-                                              interactive=True)
-                    with gr.Column():
-                        ae_dtype = gr.Radio(['fp32', 'bf16'], label="Auto-Encoder Data Type", value="bf16",
-                                            interactive=True)
-                randomize_seed = gr.Checkbox(label = "\U0001F3B2 Randomize seed", value = True, info = "If checked, result is always different")
-                seed = gr.Slider(label="Seed", minimum=0, maximum=2147483647, step=1, randomize=True)
-                with gr.Group():
-                    param_setting = gr.Radio(["Quality", "Fidelity"], interactive=True, label="Presetting", value="Quality")
-                    restart_button = gr.Button(value="Apply presetting")
-                    
-
-        with gr.Column():
-            result_slider = ImageSlider(label='Output', show_label=True, elem_id="slider1")
-            result_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery1")
-            with gr.Row():
-                with gr.Column(visible=False):
-                    llave_button = gr.Button(value="Generate description by LlaVa (disabled)")
-                with gr.Column():
-                    diffusion_button = gr.Button(value="🚀 Upscale/Restore", variant = "primary", elem_id="process_button")
-
-            with gr.Accordion("Feedback", open=True, visible=False):
-                fb_score = gr.Slider(label="Feedback Score", minimum=1, maximum=5, value=3, step=1,
-                                     interactive=True)
-                fb_text = gr.Textbox(label="Feedback Text", value="", placeholder='Please enter your feedback here.')
-                submit_button = gr.Button(value="Submit Feedback")
+    gr.HTML(title_html)
+
+    input_image = gr.Image(label="Input", show_label=True, type="numpy", height=600, elem_id="image-input")
+    prompt = gr.Textbox(label="Image description for LlaVa", value="", placeholder="A person, walking, in a town, Summer, photorealistic", lines=3, visible=False)
+    upscale = gr.Radio([1, 2, 3, 4, 5, 6, 7, 8], label="Upscale factor", info="Resolution x1 to x8", value=2, interactive=True)
+    a_prompt = gr.Textbox(label="Image description",
+                          info="The prompt that describes what the image represents",
+                          value='Cinematic, High Contrast, highly detailed, taken using a Canon EOS R '
+                                'camera, hyper detailed photo - realistic maximum detail, 32k, Color '
+                                'Grading, ultra HD, extreme meticulous detailing, skin pore detailing, '
+                                'hyper sharpness, perfect without deformations.',
+                          lines=3)
+    a_prompt_hint = gr.HTML("You can use a <a href='"'https://huggingface.co/spaces/MaziyarPanahi/llava-llama-3-8b'"'>LlaVa space</a> to auto-generate the description of your image.")
+
+    with gr.Accordion("Pre-denoising (optional)", open=False):
+        gamma_correction = gr.Slider(label="Gamma Correction", minimum=0.1, maximum=2.0, value=1.0, step=0.1)
+        denoise_button = gr.Button(value="Pre-denoise")
+        denoise_image = gr.Image(label="Denoised image", show_label=True, type="numpy", height=600, elem_id="image-s1")
+        denoise_information = gr.HTML(value="If present, the denoised image will be used for the restoration instead of the input image.", visible=False)
+
+    with gr.Accordion("LLaVA options", open=False, visible=False):
+        temperature = gr.Slider(label="Temperature", info = "lower=Always similar, higher=More creative", minimum=0., maximum=1.0, value=0.2, step=0.1)
+        top_p = gr.Slider(label="Top P", info = "Percent of tokens shortlisted", minimum=0., maximum=1.0, value=0.7, step=0.1)
+        qs = gr.Textbox(label="Question", info="Ask LLaVa what description you want", value="Describe the image and its style in a very detailed manner. The image is a realistic photography, not an art painting.", lines=3)
+
+    with gr.Accordion("Advanced options", open=False):
+        n_prompt = gr.Textbox(label="Anti image description",
+                              info="Disambiguate by listing what the image does NOT represent",
+                              value='painting, oil painting, illustration, drawing, art, sketch, anime, '
+                                    'cartoon, CG Style, 3D render, unreal engine, blurring, bokeh, ugly, dirty, messy, '
+                                    'worst quality, low quality, frames, watermark, signature, jpeg artifacts, '
+                                    'deformed, lowres, over-smooth',
+                              lines=3)
+        edm_steps = gr.Slider(label="Steps", info="lower=faster, higher=more details", minimum=1, maximum=200, value=default_setting.edm_steps if torch.cuda.device_count() > 0 else 1, step=1)
+        num_samples = gr.Slider(label="Num Samples", info="Number of generated results; I discourage to increase because the process is limited to 4 min", minimum=1, maximum=4 if not args.use_image_slider else 1
+                                , value=1, step=1)
+        with gr.Row():
+            with gr.Column():
+                model_select = gr.Radio(["v0-Q", "v0-F"], label="Model Selection", info="Q=Quality, F=Fidelity", value="v0-Q",
+                                        interactive=True)
+            with gr.Column():
+                color_fix_type = gr.Radio(["None", "AdaIn", "Wavelet"], label="Color-Fix Type", info="AdaIn=Adaptive Instance Normalization, Wavelet=For JPEG artifacts", value="Wavelet",
+                                          interactive=True)
+        s_cfg = gr.Slider(label="Text Guidance Scale", info="lower=follow the image, higher=follow the prompt", minimum=1.0, maximum=15.0,
+                          value=default_setting.s_cfg_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.1)
+        s_stage2 = gr.Slider(label="Restoring Guidance Strength", minimum=0., maximum=1., value=1., step=0.05)
+        s_stage1 = gr.Slider(label="Pre-denoising Guidance Strength", minimum=-1.0, maximum=6.0, value=-1.0, step=1.0)
+        s_churn = gr.Slider(label="S-Churn", minimum=0, maximum=40, value=5, step=1)
+        s_noise = gr.Slider(label="S-Noise", minimum=1.0, maximum=1.1, value=1.003, step=0.001)
+        with gr.Row():
+            with gr.Column():
+                linear_CFG = gr.Checkbox(label="Linear CFG", value=True)
+                spt_linear_CFG = gr.Slider(label="CFG Start", minimum=1.0,
+                                                maximum=9.0, value=default_setting.spt_linear_CFG_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.5)
+            with gr.Column():
+                linear_s_stage2 = gr.Checkbox(label="Linear Restoring Guidance", value=False)
+                spt_linear_s_stage2 = gr.Slider(label="Guidance Start", minimum=0.,
+                                                maximum=1., value=0., step=0.05)
+            with gr.Column():
+                diff_dtype = gr.Radio(['fp32', 'fp16', 'bf16'], label="Diffusion Data Type", value="fp16",
+                                      interactive=True)
+            with gr.Column():
+                ae_dtype = gr.Radio(['fp32', 'bf16'], label="Auto-Encoder Data Type", value="bf16",
+                                    interactive=True)
+        randomize_seed = gr.Checkbox(label = "\U0001F3B2 Randomize seed", value = True, info = "If checked, result is always different")
+        seed = gr.Slider(label="Seed", minimum=0, maximum=2147483647, step=1, randomize=True)
+        with gr.Group():
+            param_setting = gr.Radio(["Quality", "Fidelity"], interactive=True, label="Presetting", value="Quality")
+            restart_button = gr.Button(value="Apply presetting")
+            
+
+    llave_button = gr.Button(value="Generate description by LlaVa (disabled)", visible=False)
+    diffusion_button = gr.Button(value="🚀 Upscale/Restore", variant = "primary", elem_id="process_button")
+    restore_information = gr.HTML(value="Restart the process to get another result.", visible=False)
+    result_slider = ImageSlider(label='Output', show_label=True, elem_id="slider1")
+    result_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery1")
+
+    with gr.Accordion("Feedback", open=True, visible=False):
+        fb_score = gr.Slider(label="Feedback Score", minimum=1, maximum=5, value=3, step=1,
+                             interactive=True)
+        fb_text = gr.Textbox(label="Feedback Text", value="", placeholder='Please enter your feedback here.')
+        submit_button = gr.Button(value="Submit Feedback")
     with gr.Row():
         gr.Markdown(claim_md)
         event_id = gr.Textbox(label="Event ID", value="", visible=False)
@@ -383,7 +375,8 @@ with gr.Blocks(title="SUPIR") as interface:
         input_image,
         gamma_correction
     ], outputs=[
-        denoise_image
+        denoise_image,
+        denoise_information
     ])
 
     llave_button.click(fn = check, inputs = [
@@ -434,6 +427,7 @@ with gr.Blocks(title="SUPIR") as interface:
     ], outputs = [
         result_slider,
         result_gallery,
+        restore_information,
         event_id
     ])