token

app.py

@@ -1,181 +1,158 @@
 import spaces
-import json
-import subprocess
-from llama_cpp import Llama
-from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
-from llama_cpp_agent.providers import LlamaCppPythonProvider
-from llama_cpp_agent.chat_history import BasicChatHistory
-from llama_cpp_agent.chat_history.messages import Roles
 import gradio as gr
-from huggingface_hub import hf_hub_download
+import torch
+from PIL import Image
+from transformers import AutoProcessor, AutoModelForCausalLM, pipeline
+from diffusers import DiffusionPipeline
+import random
+import numpy as np
+import os
+import subprocess
 
-llm = None
-llm_model = None
+# Install flash-attn
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
-hf_hub_download(
-    repo_id="unsloth/Reflection-Llama-3.1-70B-GGUF",
-    filename="Reflection-Llama-3.1-70B.Q3_K_L.gguf",
-    local_dir = "./models"
-)
+# Initialize models
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32
 
-hf_hub_download(
-    repo_id="jhofseth/Reflection-Llama-3.1-70B-GGUF",
-    filename="Reflection-Llama-3.1-70B-IQ3_XXS.gguf",
-    local_dir = "./models"
-)
+huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
 
-hf_hub_download(
-    repo_id="bartowski/Reflection-Llama-3.1-70B-GGUF",
-    filename="Reflection-Llama-3.1-70B.imatrix",
-    local_dir = "./random"
-)
+# SD3.5 model
+pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large", torch_dtype=dtype, use_safetensors=True, variant="fp16", token=huggingface_token).to(device)
 
-def get_messages_formatter_type(model_name):
-    if "Llama" in model_name:
-        return MessagesFormatterType.LLAMA_3
-    else:
-        raise ValueError(f"Unsupported model: {model_name}")
+# 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 = 1024
+
+# Florence caption function
 @spaces.GPU
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    model,
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-    top_k,
-    repeat_penalty,
-):
-    global llm
-    global llm_model
-    
-    chat_template = get_messages_formatter_type(model)
+def florence_caption(image):
+    # Convert image to PIL if it's not already
+    if not isinstance(image, Image.Image):
+        image = Image.fromarray(image)
     
-    if llm is None or llm_model != model:
-        llm = Llama(
-            model_path=f"models/{model}",
-            flash_attn=True,
-            n_gpu_layers=81,
-            n_batch=1024,
-            n_ctx=8192,
-        )
-        llm_model = model
-    
-    provider = LlamaCppPythonProvider(llm)
-
-    agent = LlamaCppAgent(
-        provider,
-        system_prompt=f"{system_message}",
-        predefined_messages_formatter_type=chat_template,
-        debug_output=True
+    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,
     )
-    
-    settings = provider.get_provider_default_settings()
-    settings.temperature = temperature
-    settings.top_k = top_k
-    settings.top_p = top_p
-    settings.max_tokens = max_tokens
-    settings.repeat_penalty = repeat_penalty
-    settings.stream = True
+    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>"]
 
-    messages = BasicChatHistory()
+# 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
 
-    for msn in history:
-        user = {
-            'role': Roles.user,
-            'content': msn[0]
-        }
-        assistant = {
-            'role': Roles.assistant,
-            'content': msn[1]
-        }
-        messages.add_message(user)
-        messages.add_message(assistant)
+@spaces.GPU(duration=190)
+def process_workflow(image, text_prompt, use_enhancer, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, negative_prompt="", 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
     
-    stream = agent.get_chat_response(
-        message,
-        llm_sampling_settings=settings,
-        chat_history=messages,
-        returns_streaming_generator=True,
-        print_output=False
-    )
+    if use_enhancer:
+        prompt = enhance_prompt(prompt)
+    
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    
+    generator = torch.Generator(device=device).manual_seed(seed)
     
-    outputs = ""
-    for output in stream:
-        outputs += output
-        yield outputs
+    image = pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        generator=generator,
+        num_inference_steps=num_inference_steps,
+        width=width,
+        height=height,
+        guidance_scale=guidance_scale
+    ).images[0]
+    
+    return image, prompt, seed
 
-description = """<p><center>
-<a href="https://huggingface.co/mattshumer/ref_70_e3" target="_blank">[Reflection Llama 3.1 70B Correct Weights]</a>
-<a href="https://huggingface.co/mattshumer/Reflection-Llama-3.1-70B" target="_blank">[Old Repo]</a>
-<a href="https://huggingface.co/unsloth/Reflection-Llama-3.1-70B-GGUF" target="_blank">[Reflection-Llama-3.1-70B-GGUF]</a>
+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">Stable Diffusion 3.5 with Florence-2 Captioner and Prompt Enhancer</h1>
+<p><center>
+<a href="https://huggingface.co/stabilityai/stable-diffusion-3.5-large" target="_blank">[Stable Diffusion 3.5 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>
 """
 
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Dropdown([
-                "Reflection-Llama-3.1-70B.Q3_K_L.gguf",
-                "Reflection-Llama-3.1-70B-IQ3_XXS.gguf"
-            ],
-            value="Reflection-Llama-3.1-70B.Q3_K_L.gguf",
-            label="Model"
-        ),
-        gr.Textbox(value="You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.", label="System message"),
-        gr.Slider(minimum=1, maximum=8192, value=2048, step=1, label="Max tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p",
-        ),
-        gr.Slider(
-            minimum=0,
-            maximum=100,
-            value=40,
-            step=1,
-            label="Top-k",
-        ),
-        gr.Slider(
-            minimum=0.0,
-            maximum=2.0,
-            value=1.1,
-            step=0.1,
-            label="Repetition penalty",
-        ),
-    ],
-    theme=gr.themes.Soft(primary_hue="violet", secondary_hue="violet", neutral_hue="gray",font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set(
-        body_background_fill_dark="#16141c",
-        block_background_fill_dark="#16141c",
-        block_border_width="1px",
-        block_title_background_fill_dark="#1e1c26",
-        input_background_fill_dark="#292733",
-        button_secondary_background_fill_dark="#24212b",
-        border_color_accent_dark="#343140",
-        border_color_primary_dark="#343140",
-        background_fill_secondary_dark="#16141c",
-        color_accent_soft_dark="transparent",
-        code_background_fill_dark="#292733",
-    ),
-    retry_btn="Retry",
-    undo_btn="Undo",
-    clear_btn="Clear",
-    submit_btn="Send",
-    title="Reflection Llama-3.1 70B",
-    description=description,
-    chatbot=gr.Chatbot(
-        scale=1, 
-        likeable=False,
-        show_copy_button=True
+with gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="blue", secondary_hue="gray")) as demo:
+    gr.HTML(title)
+    
+    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)")
+            
+            with gr.Accordion("Advanced Settings", open=False):
+                text_prompt = gr.Textbox(label="Text Prompt (optional, used if no image is uploaded)")
+                negative_prompt = gr.Textbox(label="Negative Prompt")
+                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=512, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
+                height = gr.Slider(label="Height", minimum=512, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
+                guidance_scale = gr.Slider(label="Guidance Scale", minimum=0.0, maximum=7.5, step=0.1, value=4.5)
+                num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=40)
+            
+            generate_btn = gr.Button("Generate Image", elem_classes="submit-btn")
+        
+        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, negative_prompt
+        ],
+        outputs=[output_image, final_prompt, used_seed]
     )
-)
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch(debug=True)

requirements.txt

@@ -1,4 +1,11 @@
-huggingface_hub==0.22.2
-scikit-build-core
-https://github.com/abetlen/llama-cpp-python/releases/download/v0.2.90-cu124/llama_cpp_python-0.2.90-cp310-cp310-linux_x86_64.whl
-llama-cpp-agent>=0.2.25
+accelerate
+diffusers
+torch
+transformers
+git+https://github.com/huggingface/diffusers.git
+sentencepiece
+spaces
+xformers
+sentencepiece
+timm
+einops