Create app.py

app.py

@@ -0,0 +1,472 @@
+import os
+import gradio as gr
+import json
+import logging
+import torch
+from PIL import Image
+import spaces
+from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler
+from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download
+import copy
+import random
+import time
+import re
+import math
+import numpy as np
+
+# Load LoRAs from JSON
+loras = [
+    {
+        "repo": "flymy-ai/qwen-image-realism-lora",
+        "image": "https://huggingface.co/flymy-ai/qwen-image-realism-lora/resolve/main/assets/flymy_realism.png",
+        "trigger_word": "Super Realism portrait of",
+        "trigger_position": "prepend",
+        "title": "Super Realism"
+    },
+    {
+        "repo": "valiantcat/Qwen-Image-Gufeng-LoRA",
+        "image": "https://huggingface.co/valiantcat/Qwen-Image-Gufeng-LoRA/resolve/main/result/output1.png",
+        "trigger_word": "gfwm, The image is a digital illustration",
+        "trigger_position": "prepend",
+        "title": "Gufeng Style"
+    },
+    {
+        "repo": "lichorosario/qwen-image-dottrmstr",
+        "image": "https://huggingface.co/lichorosario/qwen-image-dottrmstr/resolve/main/images/Day_of_the_Tentacle_Remastered_(PC)_08.jpg",
+        "trigger_word": "DOTTRMSTR",
+        "trigger_position": "prepend",
+        "title": "Day of the Tentacle Style"
+    }
+]
+
+# Initialize the base model
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+base_model = "Qwen/Qwen-Image"
+
+# Scheduler configuration from the Qwen-Image-Lightning repository
+scheduler_config = {
+    "base_image_seq_len": 256,
+    "base_shift": math.log(3),
+    "invert_sigmas": False,
+    "max_image_seq_len": 8192,
+    "max_shift": math.log(3),
+    "num_train_timesteps": 1000,
+    "shift": 1.0,
+    "shift_terminal": None,
+    "stochastic_sampling": False,
+    "time_shift_type": "exponential",
+    "use_beta_sigmas": False,
+    "use_dynamic_shifting": True,
+    "use_exponential_sigmas": False,
+    "use_karras_sigmas": False,
+}
+
+scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
+pipe = DiffusionPipeline.from_pretrained(
+    base_model, scheduler=scheduler, torch_dtype=dtype
+).to(device)
+
+# Store Lightning LoRA info
+lightning_lora = {
+    "repo": "lightx2v/Qwen-Image-Lightning",
+    "weight_name": "Qwen-Image-Lightning-8steps-V1.0.safetensors",
+    "loaded": False
+}
+
+MAX_SEED = np.iinfo(np.int32).max
+
+class calculateDuration:
+    def __init__(self, activity_name=""):
+        self.activity_name = activity_name
+
+    def __enter__(self):
+        self.start_time = time.time()
+        return self
+    
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.end_time = time.time()
+        self.elapsed_time = self.end_time - self.start_time
+        if self.activity_name:
+            print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
+        else:
+            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
+
+def get_image_size(aspect_ratio):
+    """Converts aspect ratio string to width, height tuple."""
+    if aspect_ratio == "1:1":
+        return 1024, 1024
+    elif aspect_ratio == "16:9":
+        return 1152, 640
+    elif aspect_ratio == "9:16":
+        return 640, 1152
+    elif aspect_ratio == "4:3":
+        return 1024, 768
+    elif aspect_ratio == "3:4":
+        return 768, 1024
+    elif aspect_ratio == "3:2":
+        return 1024, 688
+    elif aspect_ratio == "2:3":
+        return 688, 1024
+    else:
+        return 1024, 1024
+
+def update_selection(evt: gr.SelectData, aspect_ratio):
+    selected_lora = loras[evt.index]
+    new_placeholder = f"Type a prompt for {selected_lora['title']}"
+    lora_repo = selected_lora["repo"]
+    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
+    
+    # Update aspect ratio if specified in LoRA config
+    if "aspect" in selected_lora:
+        if selected_lora["aspect"] == "portrait":
+            aspect_ratio = "9:16"
+        elif selected_lora["aspect"] == "landscape":
+            aspect_ratio = "16:9"
+        else:
+            aspect_ratio = "1:1"
+    
+    return (
+        gr.update(placeholder=new_placeholder),
+        updated_text,
+        evt.index,
+        aspect_ratio,
+    )
+
+def handle_speed_mode(speed_mode):
+    """Handle the speed/quality toggle for Lightning LoRA."""
+    global lightning_lora
+    
+    if speed_mode == "Speed (8 steps)":
+        # Load Lightning LoRA if not already loaded
+        if not lightning_lora["loaded"]:
+            with calculateDuration("Loading Lightning LoRA"):
+                pipe.load_lora_weights(
+                    lightning_lora["repo"], 
+                    weight_name=lightning_lora["weight_name"],
+                    adapter_name="lightning"
+                )
+                lightning_lora["loaded"] = True
+            return gr.update(value="Lightning LoRA loaded for fast generation"), 8, 1.0
+        return gr.update(value="Lightning LoRA already loaded"), 8, 1.0
+    else:  # Quality mode
+        # Unload Lightning LoRA if loaded
+        if lightning_lora["loaded"]:
+            with calculateDuration("Unloading Lightning LoRA"):
+                pipe.unload_lora_weights()
+                lightning_lora["loaded"] = False
+            return gr.update(value="Lightning LoRA unloaded for quality generation"), 28, 3.5
+        return gr.update(value="Quality mode active"), 28, 3.5
+
+@spaces.GPU(duration=70)
+def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt=""):
+    pipe.to("cuda")
+    generator = torch.Generator(device="cuda").manual_seed(seed)
+    
+    with calculateDuration("Generating image"):
+        # Generate image
+        image = pipe(
+            prompt=prompt_mash,
+            negative_prompt=negative_prompt,
+            num_inference_steps=steps,
+            true_cfg_scale=cfg_scale,  # Use true_cfg_scale for Qwen-Image
+            width=width,
+            height=height,
+            generator=generator,
+        ).images[0]
+        
+    return image
+
+@spaces.GPU(duration=70)
+def run_lora(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, progress=gr.Progress(track_tqdm=True)):
+    if selected_index is None:
+        raise gr.Error("You must select a LoRA before proceeding.")
+    
+    selected_lora = loras[selected_index]
+    lora_path = selected_lora["repo"]
+    trigger_word = selected_lora["trigger_word"]
+    
+    # Prepare prompt with trigger word
+    if trigger_word:
+        if "trigger_position" in selected_lora:
+            if selected_lora["trigger_position"] == "prepend":
+                prompt_mash = f"{trigger_word} {prompt}"
+            else:
+                prompt_mash = f"{prompt} {trigger_word}"
+        else:
+            prompt_mash = f"{trigger_word} {prompt}"
+    else:
+        prompt_mash = prompt
+
+    # First, handle Lightning LoRA based on speed mode
+    global lightning_lora
+    if speed_mode == "Speed (8 steps)" and not lightning_lora["loaded"]:
+        with calculateDuration("Loading Lightning LoRA"):
+            pipe.load_lora_weights(
+                lightning_lora["repo"], 
+                weight_name=lightning_lora["weight_name"],
+                adapter_name="lightning"
+            )
+            lightning_lora["loaded"] = True
+    elif speed_mode == "Quality (28 steps)" and lightning_lora["loaded"]:
+        with calculateDuration("Unloading Lightning LoRA"):
+            pipe.unload_lora_weights()
+            lightning_lora["loaded"] = False
+
+    # Load the selected style LoRA
+    with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
+        weight_name = selected_lora.get("weights", None)
+        
+        # If Lightning is loaded, we need to handle multiple LoRAs
+        if lightning_lora["loaded"]:
+            pipe.load_lora_weights(
+                lora_path, 
+                weight_name=weight_name, 
+                low_cpu_mem_usage=True,
+                adapter_name="style"
+            )
+            # Set both adapters active
+            pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
+        else:
+            pipe.load_lora_weights(
+                lora_path, 
+                weight_name=weight_name, 
+                low_cpu_mem_usage=True
+            )
+                
+    # Set random seed for reproducibility
+    with calculateDuration("Randomizing seed"):
+        if randomize_seed:
+            seed = random.randint(0, MAX_SEED)
+    
+    # Get image dimensions from aspect ratio
+    width, height = get_image_size(aspect_ratio)
+    
+    # Generate the image
+    final_image = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale)
+    
+    return final_image, seed
+
+def get_huggingface_safetensors(link):
+    split_link = link.split("/")
+    if len(split_link) != 2:
+        raise Exception("Invalid Hugging Face repository link format.")
+
+    print(f"Repository attempted: {split_link}")
+    
+    # Load model card
+    model_card = ModelCard.load(link)
+    base_model = model_card.data.get("base_model")
+    print(f"Base model: {base_model}")
+
+    # Validate model type (for Qwen-Image)
+    acceptable_models = {"Qwen/Qwen-Image"}
+    
+    models_to_check = base_model if isinstance(base_model, list) else [base_model]
+    
+    if not any(model in acceptable_models for model in models_to_check):
+        raise Exception("Not a Qwen-Image LoRA!")
+        
+    # Extract image and trigger word
+    image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
+    trigger_word = model_card.data.get("instance_prompt", "")
+    image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
+
+    # Initialize Hugging Face file system
+    fs = HfFileSystem()
+    try:
+        list_of_files = fs.ls(link, detail=False)
+        
+        # Find safetensors file
+        safetensors_name = None
+        for file in list_of_files:
+            filename = file.split("/")[-1]
+            if filename.endswith(".safetensors"):
+                safetensors_name = filename
+                break
+
+        if not safetensors_name:
+            raise Exception("No valid *.safetensors file found in the repository.")
+
+    except Exception as e:
+        print(e)
+        raise Exception("You didn't include a valid Hugging Face repository with a *.safetensors LoRA")
+    
+    return split_link[1], link, safetensors_name, trigger_word, image_url
+
+def check_custom_model(link):
+    print(f"Checking a custom model on: {link}")
+    if link.startswith("https://"):
+        if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
+            link_split = link.split("huggingface.co/")
+            return get_huggingface_safetensors(link_split[1])
+    else: 
+        return get_huggingface_safetensors(link)
+
+def add_custom_lora(custom_lora):
+    global loras
+    if custom_lora:
+        try:
+            title, repo, path, trigger_word, image = check_custom_model(custom_lora)
+            print(f"Loaded custom LoRA: {repo}")
+            card = f'''
+            <div class="custom_lora_card">
+              <span>Loaded custom LoRA:</span>
+              <div class="card_internal">
+                <img src="{image}" />
+                <div>
+                    <h3>{title}</h3>
+                    <small>{"Using: <code><b>"+trigger_word+"</code></b> as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}<br></small>
+                </div>
+              </div>
+            </div>
+            '''
+            existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
+            if not existing_item_index:
+                new_item = {
+                    "image": image,
+                    "title": title,
+                    "repo": repo,
+                    "weights": path,
+                    "trigger_word": trigger_word
+                }
+                print(new_item)
+                existing_item_index = len(loras)
+                loras.append(new_item)
+        
+            return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
+        except Exception as e:
+            gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-Qwen-Image LoRA, this was the issue: {e}")
+            return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-Qwen-Image LoRA"), gr.update(visible=True), gr.update(), "", None, ""
+    else:
+        return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
+
+def remove_custom_lora():
+    return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
+
+run_lora.zerogpu = True
+
+css = '''
+#gen_btn{height: 100%}
+#gen_column{align-self: stretch}
+#title{text-align: center}
+#title h1{font-size: 3em; display:inline-flex; align-items:center}
+#title img{width: 100px; margin-right: 0.5em}
+#gallery .grid-wrap{height: 10vh}
+#lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%}
+.card_internal{display: flex;height: 100px;margin-top: .5em}
+.card_internal img{margin-right: 1em}
+.styler{--form-gap-width: 0px !important}
+#speed_status{padding: .5em; border-radius: 5px; margin: 1em 0}
+'''
+
+with gr.Blocks(theme=gr.themes.Soft(), css=css, delete_cache=(60, 60)) as app:
+    title = gr.HTML(
+        """<h1><img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/qwen_image_logo.png" alt="Qwen-Image" style="height: 100px; margin-right: 10px; vertical-align: middle;"> Qwen-Image LoRA Explorer</h1>""",
+        elem_id="title",
+    )
+    
+    selected_index = gr.State(None)
+    
+    with gr.Row():
+        with gr.Column(scale=3):
+            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Type a prompt after selecting a LoRA")
+        with gr.Column(scale=1, elem_id="gen_column"):
+            generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
+    
+    with gr.Row():
+        with gr.Column():
+            selected_info = gr.Markdown("")
+            gallery = gr.Gallery(
+                [(item["image"], item["title"]) for item in loras],
+                label="LoRA Gallery",
+                allow_preview=False,
+                columns=3,
+                elem_id="gallery",
+                show_share_button=False
+            )
+            with gr.Group():
+                custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="username/qwen-image-custom-lora")
+                gr.Markdown("[Check Qwen-Image LoRAs](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image)", elem_id="lora_list")
+            custom_lora_info = gr.HTML(visible=False)
+            custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
+        
+        with gr.Column():
+            result = gr.Image(label="Generated Image")
+            
+            with gr.Row():
+                speed_mode = gr.Radio(
+                    label="Generation Mode",
+                    choices=["Speed (8 steps)", "Quality (28 steps)"],
+                    value="Quality (28 steps)",
+                    info="Speed mode uses Lightning LoRA for faster generation"
+                )
+            
+            speed_status = gr.Markdown("Quality mode active", elem_id="speed_status")
+
+    with gr.Row():
+        with gr.Accordion("Advanced Settings", open=False):
+            with gr.Column():
+                with gr.Row():
+                    aspect_ratio = gr.Radio(
+                        label="Aspect Ratio",
+                        choices=["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
+                        value="1:1"
+                    )
+                    
+                with gr.Row():
+                    cfg_scale = gr.Slider(
+                        label="Guidance Scale (True CFG)", 
+                        minimum=1.0, 
+                        maximum=5.0, 
+                        step=0.1, 
+                        value=3.5,
+                        info="Lower for speed mode, higher for quality"
+                    )
+                    steps = gr.Slider(
+                        label="Steps", 
+                        minimum=4, 
+                        maximum=50, 
+                        step=1, 
+                        value=28,
+                        info="Automatically set by speed mode"
+                    )
+                
+                with gr.Row():
+                    randomize_seed = gr.Checkbox(True, label="Randomize seed")
+                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
+                    lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
+
+    # Event handlers
+    gallery.select(
+        update_selection,
+        inputs=[aspect_ratio],
+        outputs=[prompt, selected_info, selected_index, aspect_ratio]
+    )
+    
+    speed_mode.change(
+        handle_speed_mode,
+        inputs=[speed_mode],
+        outputs=[speed_status, steps, cfg_scale]
+    )
+    
+    custom_lora.input(
+        add_custom_lora,
+        inputs=[custom_lora],
+        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt]
+    )
+    
+    custom_lora_button.click(
+        remove_custom_lora,
+        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
+    )
+    
+    gr.on(
+        triggers=[generate_button.click, prompt.submit],
+        fn=run_lora,
+        inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode],
+        outputs=[result, seed]
+    )
+
+app.queue()
+app.launch()