Update app.py -- first released version

app.py

@@ -1,96 +1,248 @@
 import gradio as gr
 import numpy as np
 import random
-from diffusers import DiffusionPipeline
 import torch
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-if torch.cuda.is_available():
-    torch.cuda.max_memory_allocated(device=device)
-    pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16", use_safetensors=True)
-    pipe.enable_xformers_memory_efficient_attention()
-    pipe = pipe.to(device)
-else: 
-    pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", use_safetensors=True)
-    pipe = pipe.to(device)
+import io, json
+from PIL import Image
+import os.path
+from weight_fusion import compose_concepts
+from regionally_controlable_sampling import sample_image, build_model, prepare_text
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+device = "cuda" if torch.cuda.is_available() else "cpu"
+power_device = "GPU" if torch.cuda.is_available() else "CPU"
 
-def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
+MAX_SEED = 100_000
 
+def generate(region1_concept,
+                region2_concept,
+                prompt,
+                region1_prompt,
+                region2_prompt,
+                negative_prompt,
+                region_neg_prompt,
+                seed,
+                randomize_seed,
+                sketch_adaptor_weight,
+                keypose_adaptor_weight
+                ):
+    
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-        
-    generator = torch.Generator().manual_seed(seed)
+
+    region1_concept, region2_concept = region1_concept.lower(), region2_concept.lower()
+    pretrained_model = merge(region1_concept, region2_concept)
+
+    keypose_condition = 'multi-concept/pose_data/two_apart.png'
+    region1 = '[0, 0, 512, 290]'
+    region2 = '[0, 650, 512, 910]'
+
+    region1_prompt = f'[<{region1_concept}1> <{region1_concept}2>, {region1_prompt}]'
+    region2_prompt = f'[<{region2_concept}1> <{region2_concept}2>, {region2_prompt}]'
+    prompt_rewrite=f"{region1_prompt}-*-{region_neg_prompt}-*-{region1}|{region2_prompt}-*-{region_neg_prompt}-*-{region2}"
+
+    result = infer(pretrained_model,
+                  prompt,
+                  prompt_rewrite,
+                  negative_prompt,
+                  seed,
+                  keypose_condition,
+                  keypose_adaptor_weight,
+                #   sketch_condition,
+                #   sketch_adaptor_weight,
+                  )
     
-    image = pipe(
-        prompt = prompt, 
-        negative_prompt = negative_prompt,
-        guidance_scale = guidance_scale, 
-        num_inference_steps = num_inference_steps, 
-        width = width, 
-        height = height,
-        generator = generator
-    ).images[0] 
+    return result
+
+def merge(concept1, concept2):
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    c1, c2 = sorted([concept1, concept2])
+    assert c1!=c2
+    merge_name = c1+'_'+c2
+
+    save_path = f'experiments/multi-concept/{merge_name}'
+
+    if os.path.isdir(save_path):
+        print(f'{save_path} already exists. Collecting merged weights from existing weights...')
+
+    else:
+        os.makedirs(save_path)
+        json_path = os.path.join(save_path,'merge_config.json')
+        alpha = 1.8
+        data = [
+            {
+                "lora_path": f"experiments/single-concept/{c1}/models/edlora_model-latest.pth",
+                "unet_alpha": alpha,
+                "text_encoder_alpha": alpha,
+                "concept_name": f"<{c1}1> <{c1}2>"
+            },
+            {
+                "lora_path": f"experiments/single-concept/{c2}/models/edlora_model-latest.pth",
+                "unet_alpha": alpha,
+                "text_encoder_alpha": alpha,
+                "concept_name": f"<{c2}1> <{c2}2>"
+            }
+        ]
+        with io.open(json_path,'w',encoding='utf8') as outfile:
+            json.dump(data, outfile, indent = 4, ensure_ascii=False)
+
+        compose_concepts(
+            concept_cfg=json_path,
+            optimize_textenc_iters=500,
+            optimize_unet_iters=50,
+            pretrained_model_path="nitrosocke/mo-di-diffusion",
+            save_path=save_path,
+            suffix='base',
+            device=device,
+        )
+        print(f'Merged weight for {c1}+{c2} saved in {save_path}!\n\n')
+
+    modelbase_path = os.path.join(save_path,'combined_model_base')
+    assert os.path.isdir(modelbase_path)
+
+    # save_path = 'experiments/multi-concept/elsa_moana_weight18/combined_model_base'
+    return modelbase_path
+
+def infer(pretrained_model,
+          prompt,
+          prompt_rewrite,
+          negative_prompt='',
+          seed=16141,
+          keypose_condition=None,
+          keypose_adaptor_weight=1.0,
+          sketch_condition=None,
+          sketch_adaptor_weight=0.0,
+          region_sketch_adaptor_weight='',
+          region_keypose_adaptor_weight=''
+          ):
+
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    pipe = build_model(pretrained_model, device)
+
+    if sketch_condition is not None and os.path.exists(sketch_condition):
+        sketch_condition = Image.open(sketch_condition).convert('L')
+        width_sketch, height_sketch = sketch_condition.size
+        print('use sketch condition')
+    else:
+        sketch_condition, width_sketch, height_sketch = None, 0, 0
+        print('skip sketch condition')
+
+    if keypose_condition is not None and os.path.exists(keypose_condition):
+        keypose_condition = Image.open(keypose_condition).convert('RGB')
+        width_pose, height_pose = keypose_condition.size
+        print('use pose condition')
+    else:
+        keypose_condition, width_pose, height_pose = None, 0, 0
+        print('skip pose condition')
+
+    if width_sketch != 0 and width_pose != 0:
+        assert width_sketch == width_pose and height_sketch == height_pose, 'conditions should be same size'
+    width, height = max(width_pose, width_sketch), max(height_pose, height_sketch)
+    kwargs = {
+        'sketch_condition': sketch_condition,
+        'keypose_condition': keypose_condition,
+        'height': height,
+        'width': width,
+    }
+
+    prompts = [prompt]
+    prompts_rewrite = [prompt_rewrite]
+    input_prompt = [prepare_text(p, p_w, height, width) for p, p_w in zip(prompts, prompts_rewrite)]
+    save_prompt = input_prompt[0][0]
+    print(save_prompt)
+
+    image = sample_image(
+        pipe,
+        input_prompt=input_prompt,
+        input_neg_prompt=[negative_prompt] * len(input_prompt),
+        generator=torch.Generator(device).manual_seed(seed),
+        sketch_adaptor_weight=sketch_adaptor_weight,
+        region_sketch_adaptor_weight=region_sketch_adaptor_weight,
+        keypose_adaptor_weight=keypose_adaptor_weight,
+        region_keypose_adaptor_weight=region_keypose_adaptor_weight,
+        **kwargs)
     
-    return image
+    return image[0]
 
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
+examples_context = [
+    'walking at Stanford university campus',
+    'in a castle',
+    'in the forest',
+    'in front of Eiffel tower'
 ]
 
+examples_region1 = ['wearing red hat, high resolution, best quality','bright smile, wearing pants, best quality']
+examples_region2 = ['smilling, wearing blue shirt, high resolution, best quality']
+
 css="""
 #col-container {
     margin: 0 auto;
-    max-width: 520px;
+    max-width: 600px;
 }
 """
 
-if torch.cuda.is_available():
-    power_device = "GPU"
-else:
-    power_device = "CPU"
-
 with gr.Blocks(css=css) as demo:
     
     with gr.Column(elem_id="col-container"):
         gr.Markdown(f"""
-        # Text-to-Image Gradio Template
+        # Orthogonal Adaptation
         Currently running on {power_device}.
         """)
-        
+        prompt = gr.Text(
+                label="ContextPrompt",
+                show_label=False,
+                max_lines=1,
+                placeholder="Enter your context prompt for overall image",
+                container=False,
+            )
         with gr.Row():
-
-            with gr.Column():
-                prompt = gr.Text(
-                    label="Prompt",
-                    show_label=False,
-                    max_lines=1,
-                    placeholder="Enter your prompt",
-                    container=False,
-                )
-                prompt2 = gr.Text(
-                    label="Prompt2",
-                    show_label=False,
-                    max_lines=1,
-                    placeholder="Enter your prompt for right character",
-                    container=False,
-                )
             
-            run_button = gr.Button("Run", scale=0)
+            region1_concept = gr.Dropdown(
+                ["Elsa", "Moana"],
+                label="Character 1",
+                info="Will add more characters later!"
+            )
+            region2_concept = gr.Dropdown(
+                ["Elsa", "Moana"], 
+                label="Character 2", 
+                info="Will add more characters later!"
+            )
+
+        with gr.Row():
+
+            region1_prompt = gr.Textbox(
+                label="Region1 Prompt",
+                show_label=False,
+                max_lines=2,
+                placeholder="Enter your prompt for character 1",
+                container=False,
+            )
+
+            region2_prompt = gr.Textbox(
+                label="Region2 Prompt",
+                show_label=False,
+                max_lines=2,
+                placeholder="Enter your prompt for character 2",
+                container=False,
+            )
+
+        run_button = gr.Button("Run", scale=1)
         
         result = gr.Image(label="Result", show_label=False)
 
         with gr.Accordion("Advanced Settings", open=False):
-            
+
             negative_prompt = gr.Text(
-                label="Negative prompt",
+                label="Context Negative prompt",
+                max_lines=1,
+                value = 'saturated, cropped, worst quality, low quality',
+                visible=False,
+            )
+
+            region_neg_prompt =  gr.Text(
+                label="Regional Negative prompt",
                 max_lines=1,
-                placeholder="Enter a negative prompt",
+                value = 'shirtless, nudity, saturated, cropped, worst quality, low quality',
                 visible=False,
             )
             
@@ -106,49 +258,60 @@ with gr.Blocks(css=css) as demo:
             
             with gr.Row():
                 
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=512,
+                sketch_adaptor_weight = gr.Slider(
+                        label="Sketch Adapter Weight",
+                        minimum = 0,
+                        maximum = 1,
+                        step=0.01,
+                        value=0,
                 )
                 
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=512,
+                keypose_adaptor_weight = gr.Slider(
+                        label="Keypose Adapter Weight",
+                        minimum = 0,
+                        maximum = 1,
+                        step= 0.01,
+                        value=1.0,
                 )
             
-            with gr.Row():
-                
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,
-                )
-                
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=12,
-                    step=1,
-                    value=2,
-                )
-        
+
         gr.Examples(
-            examples = examples,
+            label = 'Context Prompt example',
+            examples = examples_context,
             inputs = [prompt]
-        )
+        )  
+
+        with gr.Row():
+            gr.Examples(
+                label = 'Region1 Prompt example',
+                examples = examples_region1,
+                inputs = [region1_prompt]
+            )
+
+            gr.Examples(
+                label = 'Region2 Prompt example',
+                examples = [examples_region2],
+                inputs = [region2_prompt]
+            )
+
 
     run_button.click(
-        fn = infer,
-        inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
+        fn = generate,
+        inputs = [region1_concept,
+                  region2_concept,
+                  prompt,
+                  region1_prompt,
+                  region2_prompt,
+                  negative_prompt,
+                  region_neg_prompt,
+                  seed,
+                  randomize_seed,
+                #   sketch_condition,
+                #   keypose_condition,
+                  sketch_adaptor_weight,
+                  keypose_adaptor_weight
+                  ],
         outputs = [result]
     )
 
-demo.queue().launch()
+demo.queue().launch(share=True)