Update app.py

app.py

@@ -1,162 +1,205 @@
+import spaces
 import gradio as gr
-import numpy as np
-import random
-from diffusers import AuraFlowPipeline
 import torch
-import spaces
+from PIL import Image
+from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor, pipeline
+from transformers import AutoProcessor, AutoModelForCausalLM
+from diffusers import AuraFlowPipeline
+import re
+import random
+import numpy as np
 
+# Initialize models
 device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16
 
-#torch.set_float32_matmul_precision("high")
-
-#torch._inductor.config.conv_1x1_as_mm = True
-#torch._inductor.config.coordinate_descent_tuning = True
-#torch._inductor.config.epilogue_fusion = False
-#torch._inductor.config.coordinate_descent_check_all_directions = True
-
+# AuraFlow model
 pipe = AuraFlowPipeline.from_pretrained(
-	"fal/AuraFlow",
+    "fal/AuraFlow",
     torch_dtype=torch.float16
-).to("cuda")
+).to(device)
 
-#pipe.transformer.to(memory_format=torch.channels_last)
-#pipe.vae.to(memory_format=torch.channels_last)
+# VLM Captioner
+vlm_model = PaliGemmaForConditionalGeneration.from_pretrained("gokaygokay/sd3-long-captioner-v2").to(device).eval()
+vlm_processor = PaliGemmaProcessor.from_pretrained("gokaygokay/sd3-long-captioner-v2")
 
-#pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
-#pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
+# Initialize Florence model
+florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
+florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
+
+# Prompt Enhancer
+enhancer_medium = pipeline("summarization", model="gokaygokay/Lamini-fal-prompt-enchance", device=device)
+enhancer_long = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance-Long", device=device)
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 1024
 
-@spaces.GPU
-def infer(prompt, negative_prompt="", seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=5.0, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):
+# Florence caption function
+def florence_caption(image):
+    # Convert image to PIL if it's not already
+    if not isinstance(image, Image.Image):
+        image = Image.fromarray(image)
+    
+    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
+    generated_ids = florence_model.generate(
+        input_ids=inputs["input_ids"],
+        pixel_values=inputs["pixel_values"],
+        max_new_tokens=1024,
+        early_stopping=False,
+        do_sample=False,
+        num_beams=3,
+    )
+    generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
+    parsed_answer = florence_processor.post_process_generation(
+        generated_text,
+        task="<MORE_DETAILED_CAPTION>",
+        image_size=(image.width, image.height)
+    )
+    return parsed_answer["<MORE_DETAILED_CAPTION>"]
+
+# VLM Captioner function
+def create_captions_rich(image):
+    prompt = "caption en"
+    model_inputs = vlm_processor(text=prompt, images=image, return_tensors="pt").to(device)
+    input_len = model_inputs["input_ids"].shape[-1]
+
+    with torch.inference_mode():
+        generation = vlm_model.generate(**model_inputs, repetition_penalty=1.10, max_new_tokens=256, do_sample=False)
+        generation = generation[0][input_len:]
+        decoded = vlm_processor.decode(generation, skip_special_tokens=True)
+
+    return modify_caption(decoded)
+
+# Helper function for caption modification
+def modify_caption(caption: str) -> str:
+    prefix_substrings = [
+        ('captured from ', ''),
+        ('captured at ', '')
+    ]
+    pattern = '|'.join([re.escape(opening) for opening, _ in prefix_substrings])
+    replacers = {opening: replacer for opening, replacer in prefix_substrings}
+    
+    def replace_fn(match):
+        return replacers[match.group(0)]
+    
+    return re.sub(pattern, replace_fn, caption, count=1, flags=re.IGNORECASE)
+
+# Prompt Enhancer function
+def enhance_prompt(input_prompt, model_choice):
+    if model_choice == "Medium":
+        result = enhancer_medium("Enhance the description: " + input_prompt)
+        enhanced_text = result[0]['summary_text']
+        
+    else:  # Long
+        result = enhancer_long("Enhance the description: " + input_prompt)
+        enhanced_text = result[0]['summary_text']
+    
+    return enhanced_text
 
+def generate_image(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-        
-    generator = torch.Generator().manual_seed(seed)
+    
+    generator = torch.Generator(device=device).manual_seed(seed)
     
     image = pipe(
-        prompt = prompt, 
-        negative_prompt = negative_prompt,
-        width=width,
+        prompt=prompt, 
+        negative_prompt=negative_prompt,
+        guidance_scale=guidance_scale, 
+        num_inference_steps=num_inference_steps, 
+        width=width, 
         height=height,
-        guidance_scale = guidance_scale, 
-        num_inference_steps = num_inference_steps, 
-        generator = generator
-    ).images[0] 
+        generator=generator
+    ).images[0]
     
     return image, seed
 
-examples = [
-    "A photo of a lavender cat",
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-css="""
-#col-container {
-    margin: 0 auto;
-    max-width: 520px;
+@spaces.GPU(duration=200)
+def process_workflow(image, text_prompt, vlm_model_choice, use_enhancer, model_choice, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
+    if image is not None:
+        # Convert image to PIL if it's not already
+        if not isinstance(image, Image.Image):
+            image = Image.fromarray(image)
+        
+        if vlm_model_choice == "Long Captioner":
+            prompt = create_captions_rich(image)
+        else:  # Florence
+            prompt = florence_caption(image)
+    else:
+        prompt = text_prompt
+    
+    if use_enhancer:
+        prompt = enhance_prompt(prompt, model_choice)
+    
+    generated_image, used_seed = generate_image(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps)
+    
+    return generated_image, prompt, used_seed
+
+custom_css = """
+.input-group, .output-group {
+    border: 1px solid #e0e0e0;
+    border-radius: 10px;
+    padding: 20px;
+    margin-bottom: 20px;
+    background-color: #f9f9f9;
+}
+.submit-btn {
+    background-color: #2980b9 !important;
+    color: white !important;
+}
+.submit-btn:hover {
+    background-color: #3498db !important;
 }
 """
 
-if torch.cuda.is_available():
-    power_device = "GPU"
-else:
-    power_device = "CPU"
+title = """<h1 align="center">AuraFlow with VLM Captioner and Prompt Enhancer</h1>
+<p><center>
+<a href="https://huggingface.co/fal/AuraFlow" target="_blank">[AuraFlow Model]</a>
+<a href="https://huggingface.co/spaces/multimodalart/AuraFlow" target="_blank">[Original Space]</a>
+<a href="https://huggingface.co/microsoft/Florence-2-base" target="_blank">[Florence-2 Model]</a>
+<a href="https://huggingface.co/gokaygokay/sd3-long-captioner-v2" target="_blank">[Long Captioner Model]</a>
+<a href="https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance-Long" target="_blank">[Prompt Enhancer Long]</a>
+<a href="https://huggingface.co/gokaygokay/Lamini-fal-prompt-enchance" target="_blank">[Prompt Enhancer Medium]</a>
+<p align="center">Create long prompts from images or enhance your short prompts with prompt enhancer</p>
+</center></p>
+"""
 
-with gr.Blocks(css=css) as demo:
+with gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="blue", secondary_hue="gray")) as demo:
+    gr.HTML(title)
     
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(f"""
-        # AuraFlow 0.1
-        Demo of the [AuraFlow 0.1](https://huggingface.co/fal/AuraFlow) 6.8B parameters open source diffusion transformer model
-        [[blog](https://blog.fal.ai/auraflow/)] [[model](https://huggingface.co/fal/AuraFlow)] [[fal](https://fal.ai/models/fal-ai/aura-flow)]
-        """)
-        
-        with gr.Row():
+    with gr.Row():
+        with gr.Column(scale=1):
+            with gr.Group(elem_classes="input-group"):
+                input_image = gr.Image(label="Input Image (VLM Captioner)")
+                vlm_model_choice = gr.Radio(["Florence-2", "Long Captioner"], label="VLM Model", value="Florence-2")
             
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
+            with gr.Accordion("Advanced Settings", open=False):
+                text_prompt = gr.Textbox(label="Text Prompt (optional, used if no image is uploaded)")
+                use_enhancer = gr.Checkbox(label="Use Prompt Enhancer", value=False)
+                model_choice = gr.Radio(["Medium", "Long"], label="Enhancer Model", value="Long")
+                negative_prompt = gr.Textbox(label="Negative Prompt")
+                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
+                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
+                width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
+                height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
+                guidance_scale = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=10.0, step=0.1, value=5.0)
+                num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=28)
             
-            run_button = gr.Button("Run", scale=0)
+            generate_btn = gr.Button("Generate Image", elem_classes="submit-btn")
         
-        result = gr.Image(label="Result", show_label=False)
-
-        with gr.Accordion("Advanced Settings", open=False):
-            
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-            )
-            
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
-            
-            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-            
-            with gr.Row():
-                
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,
-                )
-                
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,
-                )
-            
-            with gr.Row():
-                
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=5.0,
-                )
-                
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=28,
-                )
-        
-        gr.Examples(
-            examples = examples,
-            fn = infer,
-            inputs = [prompt],
-            outputs = [result, seed],
-            cache_examples="lazy"
-        )
-
-    gr.on(
-        triggers=[run_button.click, prompt.submit, negative_prompt.submit],
-        fn = infer,
-        inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
-        outputs = [result, seed]
+        with gr.Column(scale=1):
+            with gr.Group(elem_classes="output-group"):
+                output_image = gr.Image(label="Result", elem_id="gallery", show_label=False)
+                final_prompt = gr.Textbox(label="Final Prompt Used")
+                used_seed = gr.Number(label="Seed Used")
+    
+    generate_btn.click(
+        fn=process_workflow,
+        inputs=[
+            input_image, text_prompt, vlm_model_choice, use_enhancer, model_choice,
+            negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps
+        ],
+        outputs=[output_image, final_prompt, used_seed]
     )
 
-demo.queue().launch()
+demo.launch(debug=True)