Update app.py

app.py

@@ -1,179 +1,174 @@
-import gradio as gr
-import numpy as np
-import random
 import spaces
+import gradio as gr
 import torch
-from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderTiny, AutoencoderKL
-from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
-from huggingface_hub import hf_hub_download
-from optimum.quanto import freeze, qfloat8, quantize
-from live_preview_helpers import flux_pipe_call_that_returns_an_iterable_of_images
+from PIL import Image
+from transformers import AutoProcessor, AutoModelForCausalLM, pipeline
+from diffusers import DiffusionPipeline
+import random
+import numpy as np
 import os
+import subprocess
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 2048
-# Set up environment variables and device
-huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
-dtype = torch.bfloat16
+# Initialize models
 device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.bfloat16
 
-# Load VAE models
-taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
-good_vae = AutoencoderKL.from_pretrained(
-    "black-forest-labs/FLUX.1-dev",
-    subfolder="vae",
-    torch_dtype=dtype,
-    token=huggingface_token
-).to(device)
+huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
+
+import torch
+from optimum.quanto import QuantizedDiffusersModel
+
+from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
+from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
+
+
+class QuantizedFluxTransformer2DModel(QuantizedDiffusersModel):
+    base_class = FluxTransformer2DModel
 
-# Initialize FluxPipeline instead of DiffusionPipeline
-from pipelines import FluxPipeline
+
+transformer = QuantizedFluxTransformer2DModel.from_pretrained("Kijai/flux-fp8")
+transformer.to(device="cuda", dtype=torch.bfloat16)
 
 pipe = FluxPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-dev",
-    torch_dtype=dtype,
-    vae=taef1,
+    transformer=None,
+    torch_dtype=torch.bfloat16,
     token=huggingface_token
-).to(device)
-
-# Load and fuse LoRA BEFORE quantizing
-print('Loading and fusing LoRA, please wait...')
-lora_path = hf_hub_download("gokaygokay/Flux-Game-Assets-LoRA-v2", "game_asst.safetensors")
-pipe.load_lora_weights(lora_path)
-pipe.fuse_lora(lora_scale=0.125)
-pipe.unload_lora_weights()
-
-# Quantize the transformer
-print("Quantizing transformer")
-quantize(pipe.transformer, weights=qfloat8)
-freeze(pipe.transformer)
-
-# Quantize the T5 text encoder
-print("Quantizing T5 text encoder")
-quantize(pipe.text_encoder_2, weights=qfloat8)
-freeze(pipe.text_encoder_2)
-
-# Move quantized components to device (if not already)
-pipe.transformer.to(device)
-pipe.text_encoder_2.to(device)
-
-# Move other components to device
-pipe.text_encoder.to(device, dtype=dtype)
-torch.cuda.empty_cache()
-
-@spaces.GPU(duration=75)
-def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=3.5, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):
+)
+
+pipe.transformer = transformer
+
+# 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_long = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance-Long", device=device)
+
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 2048
+
+# Florence caption function
+@spaces.GPU
+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>"]
+
+# Prompt Enhancer function
+def enhance_prompt(input_prompt):
+    result = enhancer_long("Enhance the description: " + input_prompt)
+    enhanced_text = result[0]['summary_text']
+    return enhanced_text
+
+@spaces.GPU(duration=190)
+def process_workflow(image, text_prompt, use_enhancer, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True)):
+    if image is not None:
+        # Convert image to PIL if it's not already
+        if not isinstance(image, Image.Image):
+            image = Image.fromarray(image)
+        
+        prompt = florence_caption(image)
+        print(prompt)
+    else:
+        prompt = text_prompt
+    
+    if use_enhancer:
+        prompt = enhance_prompt(prompt)
+    
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator().manual_seed(seed)
     
-    for img in pipe.flux_pipe_call_that_returns_an_iterable_of_images(
-            prompt=prompt,
-            guidance_scale=guidance_scale,
-            num_inference_steps=num_inference_steps,
-            width=width,
-            height=height,
-            generator=generator,
-            output_type="pil",
-            good_vae=good_vae,
-        ):
-            yield img, seed
-
-examples = [
-    "wbgmsst, a cat, white background",
-    "wbgmsst, a warrior, white background",
-    "wbgmsst, an anime girl, white background",
-]
-
-css = """
-#col-container {
-    margin: 0 auto;
-    max-width: 520px;
+    generator = torch.Generator(device=device).manual_seed(seed)
+    
+    image = pipe(
+        prompt=prompt,
+        generator=generator,
+        num_inference_steps=num_inference_steps,
+        width=width,
+        height=height,
+        guidance_scale=guidance_scale
+    ).images[0]
+    
+    return image, prompt, 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;
+}
+"""
+
+title = """<h1 align="center">FLUX.1-dev with Florence-2 Captioner and Prompt Enhancer</h1>
+<p><center>
+<a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" target="_blank">[FLUX.1-dev Model]</a>
+<a href="https://huggingface.co/microsoft/Florence-2-base" target="_blank">[Florence-2 Model]</a>
+<a href="https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance-Long" target="_blank">[Prompt Enhancer Long]</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"""# FLUX.1 [dev]
-12B param rectified flow transformer guidance-distilled from [FLUX.1 [pro]](https://blackforestlabs.ai/)  
-[[non-commercial license](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)] [[blog](https://blackforestlabs.ai/announcing-black-forest-labs/)] [[model](https://huggingface.co/black-forest-labs/FLUX.1-dev)]
-        """)
-        
-        with gr.Row():
-            
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
+    with gr.Row():
+        with gr.Column(scale=1):
+            with gr.Group(elem_classes="input-group"):
+                input_image = gr.Image(label="Input Image (Florence-2 Captioner)")
             
-            run_button = gr.Button("Run", scale=0)
-        
-        result = gr.Image(label="Result", show_label=False)
-        
-        with gr.Accordion("Advanced Settings", open=False):
-            
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
+            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)
+                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=1, maximum=15, step=0.1, value=3.5)
+                num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=28)
             
-            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=1,
-                    maximum=15,
-                    step=0.1,
-                    value=3.5,
-                )
-  
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=28,
-                )
+            generate_btn = gr.Button("Generate Image", elem_classes="submit-btn")
         
-        gr.Examples(
-            examples=examples,
-            fn=infer,
-            inputs=[prompt],
-            outputs=[result, seed],
-            cache_examples="lazy"
-        )
-
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn=infer,
-        inputs=[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, use_enhancer, seed, randomize_seed,
+            width, height, guidance_scale, num_inference_steps
+        ],
+        outputs=[output_image, final_prompt, used_seed]
     )
 
-demo.launch()
+demo.launch(debug=True)