qwen-image-edit

README.md

@@ -4,7 +4,7 @@ emoji: 🔥
 colorFrom: indigo
 colorTo: gray
 sdk: gradio
-sdk_version: 6.25.0
+sdk_version: 6.3.0
 app_file: app.py
 pinned: true
 license: apache-2.0

app.py

@@ -1,41 +1,105 @@
 import os
 import gc
 import gradio as gr
-from gradio import Server
-from fastapi.responses import HTMLResponse
 import numpy as np
 import spaces
 import torch
 import random
-import base64
-import json
-from io import BytesIO
 from PIL import Image
+from typing import Iterable
+from gradio.themes import Soft
+from gradio.themes.utils import colors, fonts, sizes
+
+colors.orange_red = colors.Color(
+    name="orange_red",
+    c50="#FFF0E5",
+    c100="#FFE0CC",
+    c200="#FFC299",
+    c300="#FFA366",
+    c400="#FF8533",
+    c500="#FF4500",
+    c600="#E63E00",
+    c700="#CC3700",
+    c800="#B33000",
+    c900="#992900",
+    c950="#802200",
+)
+
+class OrangeRedTheme(Soft):
+    def __init__(
+        self,
+        *,
+        primary_hue: colors.Color | str = colors.gray,
+        secondary_hue: colors.Color | str = colors.orange_red,
+        neutral_hue: colors.Color | str = colors.slate,
+        text_size: sizes.Size | str = sizes.text_lg,
+        font: fonts.Font | str | Iterable[fonts.Font | str] = (
+            fonts.GoogleFont("Outfit"), "Arial", "sans-serif",
+        ),
+        font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
+            fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace",
+        ),
+    ):
+        super().__init__(
+            primary_hue=primary_hue,
+            secondary_hue=secondary_hue,
+            neutral_hue=neutral_hue,
+            text_size=text_size,
+            font=font,
+            font_mono=font_mono,
+        )
+        super().set(
+            background_fill_primary="*primary_50",
+            background_fill_primary_dark="*primary_900",
+            body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)",
+            body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)",
+            button_primary_text_color="white",
+            button_primary_text_color_hover="white",
+            button_primary_background_fill="linear-gradient(90deg, *secondary_500, *secondary_600)",
+            button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_700)",
+            button_primary_background_fill_dark="linear-gradient(90deg, *secondary_600, *secondary_700)",
+            button_primary_background_fill_hover_dark="linear-gradient(90deg, *secondary_500, *secondary_600)",
+            button_secondary_text_color="black",
+            button_secondary_text_color_hover="white",
+            button_secondary_background_fill="linear-gradient(90deg, *primary_300, *primary_300)",
+            button_secondary_background_fill_hover="linear-gradient(90deg, *primary_400, *primary_400)",
+            button_secondary_background_fill_dark="linear-gradient(90deg, *primary_500, *primary_600)",
+            button_secondary_background_fill_hover_dark="linear-gradient(90deg, *primary_500, *primary_500)",
+            slider_color="*secondary_500",
+            slider_color_dark="*secondary_600",
+            block_title_text_weight="600",
+            block_border_width="3px",
+            block_shadow="*shadow_drop_lg",
+            button_primary_shadow="*shadow_drop_lg",
+            button_large_padding="11px",
+            color_accent_soft="*primary_100",
+            block_label_background_fill="*primary_200",
+        )
+
+orange_red_theme = OrangeRedTheme()
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
+print("torch.__version__ =", torch.__version__)
+print("Using device:", device)
 
 from diffusers import FlowMatchEulerDiscreteScheduler
 from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
 from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
 from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
 
-MAX_SEED = np.iinfo(np.int32).max
-LANCZOS = getattr(Image, "Resampling", Image).LANCZOS
-
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 dtype = torch.bfloat16
 
-print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
-print("torch.__version__ =", torch.__version__)
-print("Using device:", device)
-
-print("Loading FLUX.2 Klein 9B model base...")
 pipe = QwenImageEditPlusPipeline.from_pretrained(
     "Qwen/Qwen-Image-Edit-2509",
     transformer=QwenImageTransformer2DModel.from_pretrained(
         "prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V19",
+        #subfolder='transformer',
         torch_dtype=dtype,
-        device_map="cuda",
+        device_map='cuda'
     ),
-    torch_dtype=dtype,
+    torch_dtype=dtype
 ).to(device)
 
 try:
@@ -44,7 +108,8 @@ try:
 except Exception as e:
     print(f"Warning: Could not set FA3 processor: {e}")
 
-# ── LoRA adapter registry ──────────────────────────────────────────────────────
+MAX_SEED = np.iinfo(np.int32).max
+
 ADAPTER_SPECS = {
     "Qwen-Image-Edit-2511-Object-Adder": {
         "repo": "prithivMLmods/Qwen-Image-Edit-2511-Object-Adder",
@@ -78,144 +143,86 @@ ADAPTER_SPECS = {
     }, 
 }
 
-LOADED_ADAPTERS: set = set()
-ADAPTER_NAMES = list(ADAPTER_SPECS.keys())
-
-EXAMPLES_CONFIG = [
-    {"images": ["examples/D.jpg"], "prompt": "Add the batman logo to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "lora": "Qwen-Image-Edit-2511-Object-Adder"},
-    {"images": ["examples/A.jpg"], "prompt": "Add the slim rectangular transparent frame sunglasses to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "lora": "Qwen-Image-Edit-2511-Object-Adder"},
-    {"images": ["examples/B.jpeg"], "prompt": "Remove the necklace and goggles from the image while preserving the background and remaining elements, maintaining realism and original details.", "lora": "Qwen-Image-Edit-2511-Object-Remover"},
-    {"images": ["examples/DL2.jpg"], "prompt": "add the nike tick design inside the red marked area.", "lora": "Outfit-Design-Layout"},
-    {"images": ["examples/DL1.jpg"], "prompt": "add the akatsuki cloud design inside the red marked area.", "lora": "Outfit-Design-Layout"},
-    {"images": ["examples/C.png"], "prompt": "Add the leather cowboy cap to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "lora": "Qwen-Image-Edit-2511-Object-Adder"},
-    {"images": ["examples/ZM.jpg"], "prompt": "Zoom into the red highlighted area.", "lora": "Zoom-Master"},
-    {"images": ["examples/OBJ1.jpg"], "prompt": "Remove the red highlighted object from the scene.", "lora": "QIE-2511-Object-Remover-v2"},
-    {"images": ["examples/OBJ2.jpg"], "prompt": "Remove the red highlighted object from the scene.", "lora": "QIE-2511-Object-Remover-v2"},
-    {"images": ["examples/OE.jpg"], "prompt": "Extract the clothing and create a flat mockup.", "lora": "Extract-Outfit"},
-]
-
-def make_thumb_b64(path, max_dim=220):
-    if not os.path.exists(path):
-        return ""
-    try:
-        img = Image.open(path).convert("RGB")
-        img.thumbnail((max_dim, max_dim), LANCZOS)
-        buf = BytesIO()
-        img.save(buf, format="JPEG", quality=65)
-        return f"data:image/jpeg;base64,{base64.b64encode(buf.getvalue()).decode()}"
-    except Exception as e:
-        return ""
-
-def encode_full_image(path):
-    if not os.path.exists(path):
-        return ""
-    try:
-        with open(path, "rb") as f:
-            data = f.read()
-        ext = path.rsplit(".", 1)[-1].lower()
-        mime = {"jpg": "image/jpeg", "jpeg": "image/jpeg", "png": "image/png", "webp": "image/webp"}.get(ext, "image/jpeg")
-        return f"data:{mime};base64,{base64.b64encode(data).decode()}"
-    except Exception as e:
-        return ""
-
-def build_client_config():
-    examples = []
-    for i, ex in enumerate(EXAMPLES_CONFIG):
-        examples.append({
-            "idx": i,
-            "thumbs": [make_thumb_b64(p) for p in ex["images"]],
-            "n_images": len(ex["images"]),
-            "lora": ex["lora"],
-            "prompt": ex["prompt"],
-        })
-    return {
-        "loras": ADAPTER_NAMES,
-        "default_lora": "Qwen-Image-Edit-2511-Object-Adder",
-        "examples": examples,
-    }
-
-print("Building client config (example thumbnails)…")
-CLIENT_CONFIG = build_client_config()
-print(f"Built config with {len(EXAMPLES_CONFIG)} examples and {len(ADAPTER_NAMES)} LoRAs.")
-
-def b64_to_pil_list(b64_json_str):
-    if not b64_json_str or b64_json_str.strip() in ("", "[]"):
-        return []
-    try:
-        b64_list = json.loads(b64_json_str)
-    except Exception:
-        return []
-    pil_images = []
-    for b64_str in b64_list:
-        if not b64_str or not isinstance(b64_str, str):
-            continue
-        try:
-            if b64_str.startswith("data:image"):
-                _, data = b64_str.split(",", 1)
-            else:
-                data = b64_str
-            image_data = base64.b64decode(data)
-            pil_images.append(Image.open(BytesIO(image_data)).convert("RGB"))
-        except Exception as e:
-            print(f"Error decoding image: {e}")
-    return pil_images
-
-def pil_to_b64_png(image: Image.Image) -> str:
-    buf = BytesIO()
-    image.save(buf, format="PNG")
-    return f"data:image/png;base64,{base64.b64encode(buf.getvalue()).decode()}"
+LOADED_ADAPTERS = set()
 
 def update_dimensions_on_upload(image):
     if image is None:
         return 1024, 1024
-    w, h = image.size
-    if w > h:
-        nw = 1024
-        nh = int(nw * h / w)
+    
+    original_width, original_height = image.size
+    
+    if original_width > original_height:
+        new_width = 1024
+        aspect_ratio = original_height / original_width
+        new_height = int(new_width * aspect_ratio)
     else:
-        nh = 1024
-        nw = int(nh * w / h)
-    return (nw // 8) * 8, (nh // 8) * 8
-
-# ── Gradio Server (Server mode): FastAPI + Gradio queue/API engine ────────────
-app = Server(title="Qwen-Image-Edit-Object-Manipulator")
-
-@app.mcp.tool(name="edit_image")
-@app.api(name="edit_image")
-@spaces.GPU(size="xlarge")
+        new_height = 1024
+        aspect_ratio = original_width / original_height
+        new_width = int(new_height * aspect_ratio)
+        
+    new_width = (new_width // 8) * 8
+    new_height = (new_height // 8) * 8
+    
+    return new_width, new_height
+
+@spaces.GPU
 def infer(
-    images_b64_json: str,
-    prompt: str,
-    lora_adapter: str,
-    seed: int,
-    randomize_seed: bool,
-    guidance_scale: float,
-    steps: int,
-) -> dict:
-    """Edit one or more images with Qwen-Image-Edit + a lazily-loaded LoRA."""
+    images,
+    prompt,
+    lora_adapter,
+    seed,
+    randomize_seed,
+    guidance_scale,
+    steps,
+    progress=gr.Progress(track_tqdm=True)
+):
     gc.collect()
     torch.cuda.empty_cache()
 
-    pil_images = b64_to_pil_list(images_b64_json)
-    if not pil_images:
+    if not images:
         raise gr.Error("Please upload at least one image to edit.")
-    if not prompt or prompt.strip() == "":
-        raise gr.Error("Please enter an edit prompt.")
+
+    pil_images = []
+    if images is not None:
+        for item in images:
+            try:
+                if isinstance(item, tuple) or isinstance(item, list):
+                    path_or_img = item[0]
+                else:
+                    path_or_img = item
+
+                if isinstance(path_or_img, str):
+                    pil_images.append(Image.open(path_or_img).convert("RGB"))
+                elif isinstance(path_or_img, Image.Image):
+                    pil_images.append(path_or_img.convert("RGB"))
+                else:
+                    pil_images.append(Image.open(path_or_img.name).convert("RGB"))
+            except Exception as e:
+                print(f"Skipping invalid image item: {e}")
+                continue
+
+    if not pil_images:
+        raise gr.Error("Could not process uploaded images.")
 
     spec = ADAPTER_SPECS.get(lora_adapter)
     if not spec:
         raise gr.Error(f"Configuration not found for: {lora_adapter}")
 
     adapter_name = spec["adapter_name"]
+
     if adapter_name not in LOADED_ADAPTERS:
         print(f"--- Downloading and Loading Adapter: {lora_adapter} ---")
         try:
-            pipe.load_lora_weights(spec["repo"], weight_name=spec["weights"], adapter_name=adapter_name)
+            pipe.load_lora_weights(
+                spec["repo"], 
+                weight_name=spec["weights"], 
+                adapter_name=adapter_name
+            )
             LOADED_ADAPTERS.add(adapter_name)
         except Exception as e:
             raise gr.Error(f"Failed to load adapter {lora_adapter}: {e}")
     else:
-        print(f"--- Adapter {lora_adapter} already loaded. ---")
+        print(f"--- Adapter {lora_adapter} is already loaded. ---")
 
     pipe.set_adapters([adapter_name], adapter_weights=[1.0])
 
@@ -223,10 +230,8 @@ def infer(
         seed = random.randint(0, MAX_SEED)
 
     generator = torch.Generator(device=device).manual_seed(seed)
-    negative_prompt = (
-        "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, "
-        "extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry"
-    )
+    negative_prompt = "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry"
+
     width, height = update_dimensions_on_upload(pil_images[0])
 
     try:
@@ -240,45 +245,111 @@ def infer(
             generator=generator,
             true_cfg_scale=guidance_scale,
         ).images[0]
-        return {"image": pil_to_b64_png(result_image), "seed": seed}
+        
+        return result_image, seed
+
     except Exception as e:
         raise e
     finally:
         gc.collect()
         torch.cuda.empty_cache()
 
+@spaces.GPU
+def infer_example(images, prompt, lora_adapter):
+    if not images:
+        return None, 0
+    
+    if isinstance(images, str):
+        images_list = [images]
+    else:
+        images_list = images
+        
+    result, seed = infer(
+        images=images_list,
+        prompt=prompt,
+        lora_adapter=lora_adapter,
+        seed=0,
+        randomize_seed=True,
+        guidance_scale=1.0,
+        steps=4
+    )
+    return result, seed
 
-@app.api(name="load_example", queue=False)
-def load_example(idx: float) -> dict:
-    """Return base64-encoded example images + prompt + LoRA for a given example index."""
-    try:
-        i = int(idx)
-    except (ValueError, TypeError):
-        i = -1
-    if i < 0 or i >= len(EXAMPLES_CONFIG):
-        return {"images": [], "prompt": "", "lora": "", "names": [], "status": "error"}
-    ex = EXAMPLES_CONFIG[i]
-    b64_list, names = [], []
-    for path in ex["images"]:
-        b64 = encode_full_image(path)
-        if b64:
-            b64_list.append(b64)
-            names.append(os.path.basename(path))
-    return {"images": b64_list, "prompt": ex["prompt"], "lora": ex["lora"], "names": names, "status": "ok"}
-
-
-@app.get("/api/config")
-def client_config():
-    """Plain FastAPI route: LoRA choices + example card data for the frontend."""
-    return CLIENT_CONFIG
-
-
-@app.get("/", response_class=HTMLResponse)
-async def homepage():
-    html_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "index.html")
-    with open(html_path, "r", encoding="utf-8") as f:
-        return f.read()
-
+css="""
+#col-container {
+    margin: 0 auto;
+    max-width: 1000px;
+}
+#main-title h1 {font-size: 2.3em !important;}
+"""
+
+with gr.Blocks() as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown("# **Qwen-Image-Edit-Object-Manipulator**", elem_id="main-title")
+        gr.Markdown("Perform diverse image edits using specialized [LoRA](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image-Edit-2511) adapters. Upload one or more images.")
+
+        with gr.Row(equal_height=True):
+            with gr.Column():
+                images = gr.Gallery(
+                    label="Upload Images", 
+                    type="filepath", 
+                    columns=2, 
+                    rows=1, 
+                    height=300,
+                    allow_preview=True
+                )
+                
+                prompt = gr.Text(
+                    label="Edit Prompt",
+                    show_label=True,
+                    placeholder="e.g., transform into anime..",
+                )
+
+                run_button = gr.Button("Edit Image", variant="primary")
+
+            with gr.Column():
+                output_image = gr.Image(label="Output Image", interactive=False, format="png", height=363)
+                
+                with gr.Row():
+                    lora_adapter = gr.Dropdown(
+                        label="Choose Manipulator",
+                        choices=list(ADAPTER_SPECS.keys()),
+                        value="Qwen-Image-Edit-2511-Object-Adder"
+                    )
+                
+                with gr.Accordion("Advanced Settings", open=False, visible=False):
+                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
+                    randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
+                    guidance_scale = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0)
+                    steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=4)
+        
+        gr.Examples(
+            examples=[
+                [["examples/D.jpg"], "Add the batman logo to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "Qwen-Image-Edit-2511-Object-Adder"],
+                [["examples/A.jpg"], "Add the slim rectangular transparent frame sunglasses to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "Qwen-Image-Edit-2511-Object-Adder"],
+                [["examples/B.jpeg"], "Remove the necklace and goggles from the image while preserving the background and remaining elements, maintaining realism and original details.", "Qwen-Image-Edit-2511-Object-Remover"],
+                [["examples/DL2.jpg"], "add the nike tick design inside the red marked area.", "Outfit-Design-Layout"],
+                [["examples/DL1.jpg"], "add the akatsuki cloud design inside the red marked area.", "Outfit-Design-Layout"],
+                [["examples/C.png"], "Add the leather cowboy cap to the image while preserving the background lighting and surrounding elements maintaining realism and original details.", "Qwen-Image-Edit-2511-Object-Adder"],
+                [["examples/ZM.jpg"], "Zoom into the red highlighted area.", "Zoom-Master"],
+                [["examples/OBJ1.jpg"], "Remove the red highlighted object from the scene.", "QIE-2511-Object-Remover-v2"],
+                [["examples/OBJ2.jpg"], "Remove the red highlighted object from the scene.", "QIE-2511-Object-Remover-v2"],
+                [["examples/OE.jpg"], "Extract the clothing and create a flat mockup.", "Extract-Outfit"],
+            ],
+            inputs=[images, prompt, lora_adapter],
+            outputs=[output_image, seed],
+            fn=infer_example,
+            cache_examples=False,
+            label="Examples"
+        )
+        
+        gr.Markdown("[*](https://huggingface.co/spaces/prithivMLmods/Qwen-Image-Edit-2511-LoRAs-Fast)This is still an experimental Space for Qwen-Image-Edit-2511.")
+
+    run_button.click(
+        fn=infer,
+        inputs=[images, prompt, lora_adapter, seed, randomize_seed, guidance_scale, steps],
+        outputs=[output_image, seed]
+    )
 
 if __name__ == "__main__":
-    app.launch(show_error=True, mcp_server=True)
+    demo.queue(max_size=30).launch(css=css, theme=orange_red_theme, mcp_server=True, ssr_mode=False, show_error=True)

index.html

@@ -1,1383 +0,0 @@
-<!DOCTYPE html>
-<html lang="en">
-<head>
-<meta charset="UTF-8">
-<meta name="viewport" content="width=device-width, initial-scale=1.0">
-<title>Qwen Studio · Object Manipulator</title>
-<style>
-@import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700;800&family=JetBrains+Mono:wght@400;500;600&display=swap');
-*{box-sizing:border-box;margin:0;padding:0}
-:root{
-    --bg:#0a0a0c;--panel:#131316;--panel2:#0e0e11;--canvas:#08080a;
-    --border:#2a1f24;--border2:#3f2a32;
-    --text:#e4e4e7;--muted:#a1a1aa;--dim:#6b6b75;--faint:#4a4a52;
-    --accent:#FF1493;--accent2:#FF40A3;--accent3:#FF80C5;
-    --menubar-h:48px;--statusbar-h:30px;--rail-w:52px;--inspector-w:360px;--filmstrip-h:132px;
-}
-html,body{height:100%}
-body{
-    background:var(--bg);font-family:'Inter',system-ui,-apple-system,sans-serif;
-    font-size:13px;color:var(--text);overflow:hidden;
-}
-
-/* ══ Layout grid ══ */
-.studio{
-    display:grid;height:100vh;
-    grid-template-rows:var(--menubar-h) 1fr var(--statusbar-h);
-    grid-template-columns:var(--rail-w) 1fr var(--inspector-w);
-    grid-template-areas:
-        "menubar menubar menubar"
-        "rail    workspace inspector"
-        "statusbar statusbar statusbar";
-}
-
-/* ══ Menu bar ══ */
-.menubar{
-    grid-area:menubar;display:flex;align-items:center;gap:14px;
-    background:var(--panel);border-bottom:1px solid var(--border);
-    padding:0 16px;z-index:50;
-}
-.menubar-logo{
-    width:28px;height:28px;background:linear-gradient(135deg,#FF1493,#FF40A3,#FF80C5);
-    border-radius:8px;display:flex;align-items:center;justify-content:center;
-    box-shadow:0 2px 8px rgba(255,20,147,.35);flex-shrink:0;
-}
-.menubar-logo svg{width:15px;height:15px;fill:#fff}
-.menubar-title{font-size:14px;font-weight:700;letter-spacing:-.2px;white-space:nowrap}
-.menubar-title span{color:var(--dim);font-weight:500}
-.menubar-badge{
-    font-size:10px;font-weight:700;padding:2px 8px;border-radius:12px;
-    background:rgba(255,20,147,.14);color:var(--accent2);border:1px solid rgba(255,20,147,.25);
-    letter-spacing:.4px;white-space:nowrap;
-}
-.menubar-badge.fast{background:rgba(34,197,94,.12);color:#4ade80;border-color:rgba(34,197,94,.25)}
-.menubar-spacer{flex:1}
-.menubar-status{
-    display:flex;align-items:center;gap:7px;font-family:'JetBrains Mono',monospace;
-    font-size:11px;color:var(--dim);
-}
-.menubar-status .dot{width:7px;height:7px;border-radius:50%;background:#4ade80;box-shadow:0 0 6px rgba(74,222,128,.6)}
-.menubar-status .dot.off{background:#f87171;box-shadow:0 0 6px rgba(248,113,113,.5)}
-.gh-btn{
-    display:inline-flex;align-items:center;gap:6px;
-    padding:5px 13px;border-radius:7px;text-decoration:none;
-    font-size:12px;font-weight:600;background:var(--accent);color:#fff;
-    border:1px solid rgba(255,255,255,.15);
-    transition:all .15s ease;flex-shrink:0;
-}
-.gh-btn:hover{background:var(--accent2)}
-.gh-btn svg{fill:#fff;width:13px;height:13px}
-
-/* ══ Left icon rail ══ */
-.rail{
-    grid-area:rail;background:var(--panel);border-right:1px solid var(--border);
-    display:flex;flex-direction:column;align-items:center;padding:10px 0;gap:4px;
-    overflow-y:auto;z-index:40;
-}
-.rail-btn{
-    width:38px;height:38px;border-radius:9px;border:1px solid transparent;
-    background:transparent;color:var(--dim);cursor:pointer;
-    display:flex;align-items:center;justify-content:center;
-    transition:all .15s ease;position:relative;
-}
-.rail-btn:hover:not(:disabled){background:rgba(255,20,147,.12);color:var(--accent3);border-color:rgba(255,20,147,.25)}
-.rail-btn.active{background:rgba(255,20,147,.2);color:var(--accent2);border-color:rgba(255,20,147,.4)}
-.rail-btn:disabled{opacity:.3;cursor:not-allowed}
-.rail-btn svg{width:17px;height:17px;fill:none;stroke:currentColor;stroke-width:2;stroke-linecap:round;stroke-linejoin:round}
-.rail-btn svg.fill{fill:currentColor;stroke:none}
-.rail-sep{width:24px;height:1px;background:var(--border);margin:6px 0}
-.rail-label{
-    font-size:8px;font-weight:700;color:var(--faint);letter-spacing:1px;
-    text-transform:uppercase;margin:8px 0 2px;user-select:none;
-}
-
-/* ══ Workspace ══ */
-.workspace{
-    grid-area:workspace;display:flex;flex-direction:column;min-width:0;min-height:0;
-    background:var(--canvas);
-}
-.canvas-head{
-    display:flex;align-items:center;gap:10px;padding:8px 16px;
-    border-bottom:1px solid var(--border);background:var(--panel2);flex-shrink:0;
-}
-.seg-control{
-    display:inline-flex;background:var(--panel);border:1px solid var(--border);
-    border-radius:8px;padding:2px;gap:2px;
-}
-.seg-btn{
-    padding:4px 14px;font-size:11px;font-weight:600;color:var(--dim);
-    background:transparent;border:none;border-radius:6px;cursor:pointer;
-    font-family:'Inter',sans-serif;transition:all .15s;letter-spacing:.2px;
-}
-.seg-btn:hover:not(.active){color:var(--muted)}
-.seg-btn.active{background:rgba(255,20,147,.18);color:var(--accent2)}
-.canvas-meta{
-    font-family:'JetBrains Mono',monospace;font-size:11px;color:var(--faint);
-    overflow:hidden;white-space:nowrap;text-overflow:ellipsis;flex:1;
-}
-.canvas-actions{display:flex;align-items:center;gap:6px}
-.out-seed-chip{
-    display:none;font-family:'JetBrains Mono',monospace;font-size:10px;font-weight:500;
-    color:var(--dim);background:var(--panel);border:1px solid var(--border);
-    border-radius:6px;padding:4px 9px;cursor:pointer;transition:all .15s;
-}
-.out-seed-chip.visible{display:inline-flex}
-.out-seed-chip:hover{color:var(--accent3);border-color:rgba(255,20,147,.35)}
-
-.canvas-stage{
-    flex:1;position:relative;overflow:hidden;min-height:0;
-    display:flex;align-items:center;justify-content:center;
-    background:
-        repeating-conic-gradient(#0c0c0f 0% 25%, #08080a 0% 50%) 0 0/24px 24px;
-}
-.canvas-stage img.modern-out-img{
-    max-width:calc(100% - 48px);max-height:calc(100% - 48px);
-    box-shadow:0 12px 48px rgba(0,0,0,.7),0 0 0 1px var(--border);
-    border-radius:4px;cursor:zoom-in;image-rendering:auto;
-    animation:canvasIn .3s ease;
-}
-@keyframes canvasIn{from{opacity:0;transform:scale(.985)}to{opacity:1;transform:scale(1)}}
-.canvas-placeholder{
-    text-align:center;color:var(--faint);user-select:none;
-    display:flex;flex-direction:column;align-items:center;gap:14px;padding:24px;
-}
-.canvas-placeholder svg{width:72px;height:72px;opacity:.5}
-.canvas-placeholder .cp-title{font-size:15px;font-weight:600;color:var(--dim)}
-.canvas-placeholder .cp-sub{font-size:12px;line-height:1.7;max-width:340px}
-.canvas-placeholder kbd{
-    padding:1px 6px;background:var(--panel);border:1px solid var(--border2);
-    border-radius:4px;font-family:'JetBrains Mono',monospace;font-size:10px;color:var(--muted);
-}
-
-/* Compare slider */
-.compare-wrap{
-    position:relative;max-width:calc(100% - 48px);max-height:calc(100% - 48px);overflow:hidden;
-    display:none;user-select:none;touch-action:none;cursor:ew-resize;
-    box-shadow:0 12px 48px rgba(0,0,0,.7),0 0 0 1px var(--border);border-radius:4px;
-}
-.compare-wrap.visible{display:block}
-.compare-wrap img{display:block;max-width:100%;max-height:100%;pointer-events:none}
-.compare-top{position:absolute;top:0;left:0;width:100%;height:100%;overflow:hidden}
-.compare-top img{position:absolute;top:0;left:0;height:100%;width:auto;max-width:none!important;max-height:none!important}
-.compare-divider{position:absolute;top:0;bottom:0;width:2px;background:#fff;box-shadow:0 0 8px rgba(0,0,0,.6);pointer-events:none}
-.compare-handle{
-    position:absolute;top:50%;transform:translate(-50%,-50%);
-    width:32px;height:32px;border-radius:50%;background:#fff;
-    display:flex;align-items:center;justify-content:center;
-    box-shadow:0 2px 8px rgba(0,0,0,.5);color:#111;font-size:13px;font-weight:800;
-}
-.compare-label{
-    position:absolute;top:8px;padding:2px 8px;border-radius:4px;
-    background:rgba(0,0,0,.65);color:#fff;font-size:10px;font-weight:600;
-    font-family:'JetBrains Mono',monospace;pointer-events:none;
-}
-.compare-label.before{left:8px}
-.compare-label.after{right:8px}
-
-/* Loader */
-.modern-loader{
-    display:none;position:absolute;inset:0;background:rgba(8,8,10,.9);
-    z-index:15;flex-direction:column;align-items:center;justify-content:center;gap:16px;backdrop-filter:blur(5px);
-}
-.modern-loader.active{display:flex}
-.loader-spinner{
-    width:38px;height:38px;border:3px solid var(--border);border-top-color:var(--accent);
-    border-radius:50%;animation:spin .8s linear infinite;
-}
-@keyframes spin{to{transform:rotate(360deg)}}
-.loader-text{font-size:13px;color:var(--muted);font-weight:500;text-align:center;padding:0 16px}
-.loader-bar-track{width:220px;height:4px;background:var(--border);border-radius:2px;overflow:hidden}
-.loader-bar-fill{
-    height:100%;width:100%;background:linear-gradient(90deg,#FF1493,#FF40A3,#FF1493);
-    background-size:200% 100%;animation:shimmer 1.5s ease-in-out infinite;border-radius:2px;
-    transition:width .3s ease;
-}
-.loader-bar-fill.determinate{animation:none;background:var(--accent)}
-@keyframes shimmer{0%{background-position:200% 0}100%{background-position:-200% 0}}
-
-/* ══ Filmstrip ══ */
-.filmstrip{
-    height:var(--filmstrip-h);flex-shrink:0;background:var(--panel);
-    border-top:1px solid var(--border);display:flex;min-height:0;overflow:hidden;
-}
-.filmstrip-section{
-    display:flex;flex-direction:column;min-width:0;padding:8px 0 8px 14px;
-}
-.filmstrip-section.inputs{flex:1.2;border-right:1px solid var(--border)}
-.filmstrip-section.history{flex:1}
-.filmstrip-label{
-    font-size:9px;font-weight:700;color:var(--faint);letter-spacing:1.2px;
-    text-transform:uppercase;margin-bottom:6px;display:flex;align-items:center;gap:8px;
-    user-select:none;flex-shrink:0;
-}
-.filmstrip-label .count{
-    font-family:'JetBrains Mono',monospace;font-weight:500;color:var(--dim);
-    background:var(--panel2);border:1px solid var(--border);border-radius:4px;padding:0 5px;font-size:9px;
-}
-.filmstrip-row{
-    display:flex;gap:8px;overflow-x:auto;overflow-y:hidden;flex:1;
-    align-items:flex-start;padding-bottom:4px;padding-right:14px;
-}
-.filmstrip-row::-webkit-scrollbar{height:5px}
-.filmstrip-row::-webkit-scrollbar-thumb{background:var(--border);border-radius:3px}
-.fs-thumb{
-    position:relative;flex-shrink:0;width:76px;height:76px;border-radius:8px;overflow:hidden;
-    border:2px solid var(--border);cursor:pointer;transition:all .15s;background:var(--panel2);
-    animation:thumbIn .25s ease;
-}
-@keyframes thumbIn{from{opacity:0;transform:scale(.92)}to{opacity:1;transform:scale(1)}}
-.fs-thumb:hover{border-color:var(--border2);transform:translateY(-2px)}
-.fs-thumb.selected{border-color:var(--accent);box-shadow:0 0 0 2px rgba(255,20,147,.25)}
-.fs-thumb img{width:100%;height:100%;object-fit:cover}
-.fs-badge{
-    position:absolute;bottom:4px;left:4px;background:rgba(0,0,0,.75);color:#fff;
-    padding:1px 6px;border-radius:4px;font-family:'JetBrains Mono',monospace;font-size:9px;font-weight:600;
-}
-.fs-remove{
-    position:absolute;top:4px;right:4px;width:20px;height:20px;background:rgba(0,0,0,.8);
-    color:#fff;border:1px solid rgba(255,255,255,.2);border-radius:50%;cursor:pointer;
-    display:none;align-items:center;justify-content:center;font-size:10px;transition:all .15s;line-height:1;
-}
-.fs-thumb:hover .fs-remove{display:flex}
-.fs-remove:hover{background:#ef4444;border-color:#ef4444}
-.fs-add{
-    flex-shrink:0;width:76px;height:76px;border-radius:8px;border:2px dashed var(--border2);
-    display:flex;flex-direction:column;align-items:center;justify-content:center;
-    cursor:pointer;transition:all .2s;background:rgba(255,20,147,.03);gap:3px;
-}
-.fs-add:hover{border-color:var(--accent);background:rgba(255,20,147,.08)}
-.fs-add .add-icon{font-size:22px;color:var(--dim);font-weight:300}
-.fs-add .add-text{font-size:9px;color:var(--dim);font-weight:600;text-transform:uppercase;letter-spacing:.5px}
-.fs-empty{
-    flex-shrink:0;display:flex;align-items:center;height:76px;
-    color:var(--faint);font-size:11px;font-style:italic;padding-right:12px;
-}
-
-/* ══ Inspector ══ */
-.inspector{
-    grid-area:inspector;background:var(--panel);border-left:1px solid var(--border);
-    display:flex;flex-direction:column;min-height:0;z-index:40;
-}
-.inspector-tabs{
-    display:flex;border-bottom:1px solid var(--border);flex-shrink:0;
-}
-.insp-tab{
-    flex:1;padding:11px 8px;font-size:11px;font-weight:700;color:var(--dim);
-    background:transparent;border:none;border-bottom:2px solid transparent;
-    cursor:pointer;font-family:'Inter',sans-serif;text-transform:uppercase;letter-spacing:.8px;
-    transition:all .15s;
-}
-.insp-tab:hover:not(.active){color:var(--muted)}
-.insp-tab.active{color:var(--accent2);border-bottom-color:var(--accent)}
-.insp-page{display:none;flex:1;overflow-y:auto;min-height:0;flex-direction:column}
-.insp-page.active{display:flex}
-
-.insp-section{border-bottom:1px solid var(--border);padding:14px 18px}
-.insp-section-title{
-    font-size:10px;font-weight:700;color:var(--dim);text-transform:uppercase;
-    letter-spacing:1px;margin-bottom:10px;display:flex;align-items:center;justify-content:space-between;
-}
-.char-count{font-size:9px;font-weight:500;color:var(--faint);font-family:'JetBrains Mono',monospace;text-transform:none;letter-spacing:0}
-.modern-textarea{
-    width:100%;background:var(--panel2);border:1px solid var(--border);border-radius:8px;
-    padding:10px 13px;font-family:'Inter',sans-serif;font-size:13px;color:var(--text);
-    resize:vertical;outline:none;min-height:64px;transition:border-color .2s;line-height:1.55;
-}
-.modern-textarea:focus{border-color:var(--accent);box-shadow:0 0 0 3px rgba(255,20,147,.15)}
-.modern-textarea::placeholder{color:var(--faint)}
-.modern-textarea.error-flash{
-    border-color:#ef4444!important;box-shadow:0 0 0 3px rgba(239,68,68,.2)!important;animation:shake .4s ease;
-}
-@keyframes shake{0%,100%{transform:translateX(0)}20%,60%{transform:translateX(-4px)}40%,80%{transform:translateX(4px)}}
-
-.suggestions-wrap{display:flex;flex-wrap:wrap;gap:5px}
-.suggestion-chip{
-    display:inline-flex;align-items:center;padding:4px 11px;
-    background:rgba(255,20,147,.07);border:1px solid rgba(255,20,147,.18);border-radius:14px;
-    color:var(--accent3);font-size:11px;font-weight:500;font-family:'Inter',sans-serif;
-    cursor:pointer;transition:all .15s;white-space:nowrap;
-}
-.suggestion-chip:hover{background:rgba(255,20,147,.15);border-color:rgba(255,20,147,.35);color:var(--accent2)}
-.suggestion-chip.active{background:rgba(255,20,147,.25);border-color:var(--accent);color:#fff}
-
-.lora-native-select{
-    width:100%;background:var(--panel2);
-    border:1px solid var(--border2);border-radius:8px;
-    padding:9px 34px 9px 13px;
-    font-family:'Inter',sans-serif;font-size:12px;font-weight:600;color:var(--text);
-    outline:none;appearance:none;-webkit-appearance:none;
-    background-image:url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='12' height='12' viewBox='0 0 12 12'%3E%3Cpath fill='%23FF1493' d='M6 8L1 3h10z'/%3E%3C/svg%3E");
-    background-repeat:no-repeat;background-position:right 11px center;
-    cursor:pointer;transition:border-color .2s;color-scheme:dark;
-}
-.lora-native-select:focus{border-color:var(--accent);box-shadow:0 0 0 3px rgba(255,20,147,.15)}
-.lora-native-select option{background:var(--panel);color:var(--text)}
-
-.btn-run{
-    display:flex;align-items:center;justify-content:center;gap:8px;width:100%;
-    background:linear-gradient(135deg,#FF1493,#C01079);border:none;border-radius:9px;
-    padding:12px 24px;cursor:pointer;font-size:14px;font-weight:700;font-family:'Inter',sans-serif;
-    color:#fff;transition:all .2s ease;letter-spacing:-.2px;
-    box-shadow:0 4px 16px rgba(255,20,147,.3),inset 0 1px 0 rgba(255,255,255,.1);
-}
-.btn-run:hover:not(:disabled){
-    background:linear-gradient(135deg,#FF40A3,#FF1493);transform:translateY(-1px);
-    box-shadow:0 6px 24px rgba(255,20,147,.45),inset 0 1px 0 rgba(255,255,255,.15);
-}
-.btn-run:active:not(:disabled){transform:translateY(0)}
-.btn-run:disabled{opacity:.55;cursor:not-allowed}
-.btn-cancel{
-    display:flex;align-items:center;justify-content:center;gap:6px;width:100%;
-    margin-top:8px;background:transparent;border:1px solid var(--border2);border-radius:9px;
-    padding:8px 24px;cursor:pointer;font-size:12px;font-weight:600;font-family:'Inter',sans-serif;
-    color:var(--muted);transition:all .15s;
-}
-.btn-cancel:hover{border-color:#ef4444;color:#ef4444;background:rgba(239,68,68,.08)}
-.run-hint{
-    text-align:center;font-size:10px;color:var(--faint);margin-top:8px;
-    font-family:'JetBrains Mono',monospace;
-}
-
-.slider-row{display:flex;align-items:center;gap:9px;min-height:26px;margin-bottom:10px}
-.slider-row:last-child{margin-bottom:0}
-.slider-row label{font-size:12px;font-weight:500;color:var(--muted);min-width:64px;flex-shrink:0}
-.slider-row input[type="range"]{
-    flex:1;-webkit-appearance:none;appearance:none;height:5px;background:var(--border);
-    border-radius:3px;outline:none;min-width:0;
-}
-.slider-row input[type="range"]::-webkit-slider-thumb{
-    -webkit-appearance:none;width:14px;height:14px;background:linear-gradient(135deg,#FF1493,#C01079);
-    border-radius:50%;cursor:pointer;box-shadow:0 2px 6px rgba(255,20,147,.4);transition:transform .15s;
-}
-.slider-row input[type="range"]::-webkit-slider-thumb:hover{transform:scale(1.2)}
-.slider-row input[type="range"]::-moz-range-thumb{
-    width:14px;height:14px;background:linear-gradient(135deg,#FF1493,#C01079);
-    border-radius:50%;cursor:pointer;border:none;box-shadow:0 2px 6px rgba(255,20,147,.4);
-}
-.slider-val{
-    min-width:48px;text-align:right;font-family:'JetBrains Mono',monospace;font-size:11px;
-    font-weight:500;padding:2px 7px;background:var(--panel2);border:1px solid var(--border);
-    border-radius:5px;color:var(--muted);flex-shrink:0;
-}
-.dice-btn{
-    display:inline-flex;align-items:center;justify-content:center;width:24px;height:24px;
-    background:var(--panel2);border:1px solid var(--border);border-radius:5px;cursor:pointer;
-    color:var(--dim);transition:all .15s;flex-shrink:0;padding:0;
-}
-.dice-btn:hover{border-color:var(--accent);color:var(--accent3)}
-.dice-btn svg{width:13px;height:13px;fill:currentColor}
-.checkbox-row{display:flex;align-items:center;gap:8px;font-size:12px;color:var(--muted);margin-bottom:10px}
-.checkbox-row input[type="checkbox"]{accent-color:var(--accent);width:15px;height:15px;cursor:pointer}
-.checkbox-row label{color:var(--muted);font-size:12px;cursor:pointer}
-
-/* Examples page */
-.examples-list{padding:12px 14px;display:flex;flex-direction:column;gap:10px}
-.example-card{
-    background:var(--panel2);border:1px solid var(--border);
-    border-radius:10px;overflow:hidden;cursor:pointer;transition:all .2s ease;flex-shrink:0;
-    display:flex;align-items:stretch;
-}
-.example-card:hover{border-color:var(--accent);box-shadow:0 4px 14px rgba(255,20,147,.15)}
-.example-card.loading{opacity:.5;pointer-events:none}
-.example-thumb-wrap{width:96px;flex-shrink:0;background:var(--panel);overflow:hidden;display:flex}
-.example-thumb-wrap img{width:100%;height:100%;object-fit:cover}
-.example-thumb-placeholder{
-    width:100%;height:100%;display:flex;align-items:center;justify-content:center;
-    background:var(--panel);color:var(--border2);font-size:10px;
-}
-.example-meta{padding:6px 10px 2px;display:flex;align-items:center;gap:5px;flex-wrap:wrap}
-.example-lora-badge{
-    display:inline-flex;padding:2px 7px;background:rgba(255,20,147,.1);border-radius:4px;
-    font-size:9px;font-weight:600;color:var(--accent2);font-family:'JetBrains Mono',monospace;
-    white-space:nowrap;border:1px solid rgba(255,20,147,.25);
-}
-.example-body{flex:1;display:flex;flex-direction:column;justify-content:center;min-width:0}
-.example-prompt-text{
-    padding:4px 10px 9px;font-size:11px;color:var(--muted);line-height:1.45;
-    display:-webkit-box;-webkit-line-clamp:2;-webkit-box-orient:vertical;overflow:hidden;
-}
-
-/* ══ Status bar ══ */
-.statusbar{
-    grid-area:statusbar;background:var(--panel);border-top:1px solid var(--border);
-    display:flex;align-items:center;gap:4px;padding:0 14px;font-size:11px;z-index:50;
-}
-.statusbar .sb-section{
-    padding:0 10px;display:flex;align-items:center;font-family:'JetBrains Mono',monospace;
-    font-size:11px;color:var(--faint);overflow:hidden;white-space:nowrap;gap:6px;
-}
-.statusbar .sb-right{margin-left:auto;display:flex;align-items:center;gap:10px}
-.sb-pill{
-    padding:2px 12px;border-radius:5px;font-weight:600;font-size:10px;
-    background:rgba(255,20,147,.1);color:var(--accent2);
-}
-.sb-pill.error{background:rgba(239,68,68,.12);color:#f87171}
-.sb-pill.done{background:rgba(34,197,94,.12);color:#4ade80}
-.sb-link{color:var(--faint);text-decoration:none;font-family:'JetBrains Mono',monospace;font-size:10px}
-.sb-link:hover{color:var(--accent3)}
-
-/* ══ Toast ══ */
-.toast-notification{
-    position:fixed;top:60px;left:50%;transform:translateX(-50%) translateY(-140%);
-    z-index:9999;padding:10px 22px;border-radius:9px;font-family:'Inter',sans-serif;
-    font-size:13px;font-weight:600;display:flex;align-items:center;gap:8px;
-    box-shadow:0 8px 24px rgba(0,0,0,.5);
-    transition:transform .35s cubic-bezier(.34,1.56,.64,1),opacity .35s ease;opacity:0;pointer-events:none;
-}
-.toast-notification.visible{transform:translateX(-50%) translateY(0);opacity:1;pointer-events:auto}
-.toast-notification.error{background:linear-gradient(135deg,#dc2626,#b91c1c);color:#fff;border:1px solid rgba(255,255,255,.15)}
-.toast-notification.warning{background:linear-gradient(135deg,#d97706,#b45309);color:#fff;border:1px solid rgba(255,255,255,.15)}
-.toast-notification.info{background:linear-gradient(135deg,#FF1493,#C01079);color:#fff;border:1px solid rgba(255,255,255,.15)}
-.toast-notification .toast-icon{font-size:15px;line-height:1}
-
-/* ══ Lightbox ══ */
-.lightbox{
-    position:fixed;inset:0;background:rgba(0,0,0,.92);z-index:10000;
-    display:none;align-items:center;justify-content:center;cursor:zoom-out;
-    backdrop-filter:blur(6px);animation:fadeIn .2s ease;
-}
-.lightbox.visible{display:flex}
-.lightbox img{max-width:94vw;max-height:94vh;border-radius:6px;box-shadow:0 20px 60px rgba(0,0,0,.8)}
-@keyframes fadeIn{from{opacity:0}to{opacity:1}}
-
-/* ══ Drop overlay ══ */
-.drop-overlay{
-    position:fixed;inset:0;z-index:9000;background:rgba(8,8,10,.85);
-    display:none;align-items:center;justify-content:center;pointer-events:none;
-    backdrop-filter:blur(3px);
-}
-.drop-overlay.visible{display:flex}
-.drop-overlay-inner{
-    border:3px dashed var(--accent);border-radius:22px;padding:56px 76px;
-    font-size:19px;font-weight:700;color:var(--accent2);background:rgba(255,20,147,.06);
-}
-
-/* ══ Scrollbars ══ */
-::-webkit-scrollbar{width:8px;height:8px}
-::-webkit-scrollbar-track{background:transparent}
-::-webkit-scrollbar-thumb{background:var(--border);border-radius:4px}
-::-webkit-scrollbar-thumb:hover{background:var(--border2)}
-
-/* ══ Touch devices ══ */
-@media (pointer:coarse){
-    .fs-remove{display:flex}
-    .rail-btn{width:44px;height:44px}
-    .rail-btn svg{width:19px;height:19px}
-}
-
-/* ══ Responsive ══ */
-@media(max-width:960px){
-    html,body{height:auto}
-    body{overflow:auto!important}
-    .studio{
-        display:flex;flex-direction:column;height:auto;min-height:100vh;
-    }
-    .menubar{
-        flex-wrap:wrap;gap:8px;padding:10px 12px;min-height:var(--menubar-h);
-        position:sticky;top:0;
-    }
-    .menubar-title{font-size:13px}
-    .menubar-status{display:none}
-    .menubar-spacer{flex:1}
-    .gh-btn{padding:6px 10px}
-    .gh-btn span{display:none}
-    .rail{
-        flex-direction:row;justify-content:flex-start;flex-wrap:nowrap;
-        overflow-x:auto;overflow-y:hidden;border-right:none;
-        border-bottom:1px solid var(--border);padding:6px 10px;gap:6px;
-    }
-    .rail-sep{width:1px;height:26px;margin:0 4px;flex-shrink:0}
-    .rail-label{display:none}
-    .rail-btn{flex-shrink:0}
-    .workspace{min-height:0}
-    .canvas-head{flex-wrap:wrap;gap:8px;padding:8px 12px}
-    .canvas-meta{flex-basis:100%;order:3;font-size:10px}
-    .canvas-stage{min-height:42vh}
-    .canvas-stage img.modern-out-img{max-width:calc(100% - 20px);max-height:52vh}
-    .compare-wrap{max-width:calc(100% - 20px);max-height:52vh}
-    .canvas-placeholder .cp-sub{font-size:11px;max-width:280px}
-    .filmstrip{flex-direction:column;height:auto;max-height:none}
-    .filmstrip-section.inputs{border-right:none;border-bottom:1px solid var(--border)}
-    .filmstrip-section{padding:8px 0 8px 12px}
-    .fs-thumb{width:68px;height:68px}
-    .fs-add{width:68px;height:68px}
-    .fs-empty{height:68px}
-    .inspector{border-left:none;border-top:1px solid var(--border)}
-    .inspector-tabs{position:sticky;top:0;background:var(--panel);z-index:30}
-    .insp-tab{padding:13px 8px;font-size:12px}
-    .insp-page{overflow-y:visible}
-    .insp-section{padding:14px 16px}
-    .modern-textarea{font-size:16px}
-    .btn-run{padding:14px 24px;font-size:15px}
-    .suggestion-chip{padding:6px 13px;font-size:12px}
-    .statusbar{flex-wrap:wrap;height:auto;min-height:var(--statusbar-h);padding:4px 10px;gap:2px}
-    .sb-link{display:none}
-    .toast-notification{max-width:calc(100vw - 32px);font-size:12px;padding:10px 16px}
-    .drop-overlay-inner{padding:32px 28px;font-size:15px;border-radius:16px;margin:0 16px}
-    .example-card{flex-direction:column}
-    .example-thumb-wrap{width:100%;height:96px}
-}
-
-@media(max-width:400px){
-    .menubar-badge.fast{display:none}
-    .seg-btn{padding:4px 12px}
-    .canvas-stage{min-height:36vh}
-}
-</style>
-</head>
-<body>
-<div class="studio">
-
-  <!-- ══ Menu bar ══ -->
-  <div class="menubar">
-    <div class="menubar-logo">
-      <svg viewBox="0 0 24 24" fill="white" xmlns="http://www.w3.org/2000/svg"><path d="M12 23c-3.6 0-8-2.69-8-7.5 0-3.5 3-6.5 4.5-8 .27-.27.75-.08.75.28v2.44c0 .42.5.63.72.28C12.28 7.5 13 3 13 1c0-.42.48-.64.8-.35C18 4.5 20 9 20 12c0 5.5-3.5 11-8 11z"/></svg>
-    </div>
-    <div class="menubar-title">Qwen Studio <span>/ Object Manipulator</span></div>
-    <span class="menubar-badge">2511</span>
-    <span class="menubar-badge fast">4-STEP FAST</span>
-    <div class="menubar-spacer"></div>
-    <div class="menubar-status"><span class="dot" id="conn-dot"></span><span id="conn-text">connecting…</span></div>
-    <a href="https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-Object-Manipulator"
-       target="_blank" rel="noopener" class="gh-btn">
-      <svg viewBox="0 0 16 16" xmlns="http://www.w3.org/2000/svg"><path d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0016 8c0-4.42-3.58-8-8-8z"/></svg>
-      <span>GitHub</span>
-    </a>
-  </div>
-
-  <!-- ══ Left icon rail ══ -->
-  <div class="rail">
-    <div class="rail-label">Input</div>
-    <button class="rail-btn" id="tb-upload" title="Upload images">
-      <svg viewBox="0 0 24 24"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="17 8 12 3 7 8"/><line x1="12" y1="3" x2="12" y2="15"/></svg>
-    </button>
-    <button class="rail-btn" id="tb-remove" title="Remove selected image" disabled>
-      <svg viewBox="0 0 24 24"><circle cx="12" cy="12" r="10"/><line x1="15" y1="9" x2="9" y2="15"/><line x1="9" y1="9" x2="15" y2="15"/></svg>
-    </button>
-    <button class="rail-btn" id="tb-clear" title="Clear all images" disabled>
-      <svg viewBox="0 0 24 24"><polyline points="3 6 5 6 21 6"/><path d="M19 6v14a2 2 0 01-2 2H7a2 2 0 01-2-2V6m3 0V4a2 2 0 012-2h4a2 2 0 012 2v2"/><line x1="10" y1="11" x2="10" y2="17"/><line x1="14" y1="11" x2="14" y2="17"/></svg>
-    </button>
-    <div class="rail-sep"></div>
-    <div class="rail-label">Result</div>
-    <button class="rail-btn" id="compare-btn" title="Compare before / after" disabled>
-      <svg viewBox="0 0 24 24"><rect x="3" y="3" width="18" height="18" rx="2"/><line x1="12" y1="3" x2="12" y2="21"/></svg>
-    </button>
-    <button class="rail-btn" id="use-as-input-btn" title="Use result as input (chain edits)" disabled>
-      <svg viewBox="0 0 24 24"><polyline points="17 1 21 5 17 9"/><path d="M3 11V9a4 4 0 014-4h14"/><polyline points="7 23 3 19 7 15"/><path d="M21 13v2a4 4 0 01-4 4H3"/></svg>
-    </button>
-    <button class="rail-btn" id="dl-btn-output" title="Download result" disabled>
-      <svg viewBox="0 0 24 24"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>
-    </button>
-    <button class="rail-btn" id="zoom-btn" title="Zoom result" disabled>
-      <svg viewBox="0 0 24 24"><circle cx="11" cy="11" r="8"/><line x1="21" y1="21" x2="16.65" y2="16.65"/><line x1="11" y1="8" x2="11" y2="14"/><line x1="8" y1="11" x2="14" y2="11"/></svg>
-    </button>
-  </div>
-
-  <!-- ══ Workspace ══ -->
-  <div class="workspace">
-    <div class="canvas-head">
-      <div class="seg-control">
-        <button class="seg-btn active" id="seg-result">Result</button>
-        <button class="seg-btn" id="seg-input">Input</button>
-      </div>
-      <div class="canvas-meta" id="canvas-meta">No image loaded</div>
-      <div class="canvas-actions">
-        <span id="out-seed-chip" class="out-seed-chip" title="Click to copy seed"></span>
-      </div>
-    </div>
-
-    <div class="canvas-stage" id="output-image-container">
-      <div class="modern-loader" id="output-loader">
-        <div class="loader-spinner"></div>
-        <div class="loader-text" id="loader-text">Processing image&hellip;</div>
-        <div class="loader-bar-track"><div class="loader-bar-fill" id="loader-bar-fill"></div></div>
-      </div>
-
-      <div class="canvas-placeholder" id="output-placeholder">
-        <svg viewBox="0 0 80 80" fill="none" xmlns="http://www.w3.org/2000/svg">
-          <rect x="8" y="14" width="64" height="52" rx="6" fill="none" stroke="#3f2a32" stroke-width="2" stroke-dasharray="4 3"/>
-          <polygon points="12,62 30,40 42,50 54,34 68,62" fill="rgba(255,20,147,0.1)" stroke="#3f2a32" stroke-width="1.5"/>
-          <circle cx="28" cy="30" r="6" fill="rgba(255,20,147,0.12)" stroke="#3f2a32" stroke-width="1.5"/>
-        </svg>
-        <div class="cp-title">Canvas is empty</div>
-        <div class="cp-sub">
-          Drop images anywhere, paste with <kbd>⌘/Ctrl+V</kbd>, or use the rail to upload.
-          Then write an instruction and press <kbd>⌘/Ctrl</kbd>+<kbd>↵</kbd>.
-        </div>
-      </div>
-
-      <div class="compare-wrap" id="compare-wrap">
-        <img id="compare-after-img" alt="after">
-        <div class="compare-top" id="compare-top"><img id="compare-before-img" alt="before"></div>
-        <div class="compare-divider" id="compare-divider"><div class="compare-handle">⇄</div></div>
-        <span class="compare-label before">Before</span>
-        <span class="compare-label after">After</span>
-      </div>
-    </div>
-
-    <!-- Filmstrip -->
-    <div class="filmstrip">
-      <div class="filmstrip-section inputs">
-        <div class="filmstrip-label">Inputs <span class="count" id="tb-image-count">0</span></div>
-        <div class="filmstrip-row" id="image-gallery-grid"></div>
-      </div>
-      <div class="filmstrip-section history">
-        <div class="filmstrip-label">History <span class="count" id="history-count">0</span></div>
-        <div class="filmstrip-row" id="history-strip"></div>
-      </div>
-    </div>
-  </div>
-
-  <!-- ══ Inspector ══ -->
-  <div class="inspector">
-    <div class="inspector-tabs">
-      <button class="insp-tab active" id="tab-edit">Edit</button>
-      <button class="insp-tab" id="tab-examples">Examples</button>
-    </div>
-
-    <!-- Edit page -->
-    <div class="insp-page active" id="page-edit">
-      <div class="insp-section">
-        <div class="insp-section-title"><span>Instruction</span><span class="char-count" id="char-count">0</span></div>
-        <textarea id="custom-prompt-input" class="modern-textarea" rows="3"
-                  placeholder="e.g., Add the batman logo, Remove the necklace, Zoom in..."></textarea>
-      </div>
-
-      <div class="insp-section">
-        <div class="insp-section-title"><span>Quick Prompts</span></div>
-        <div class="suggestions-wrap" id="suggestions-wrap"></div>
-      </div>
-
-      <div class="insp-section">
-        <div class="insp-section-title"><span>Style / LoRA</span></div>
-        <select id="custom-lora-select" class="lora-native-select"></select>
-      </div>
-
-      <div class="insp-section">
-        <button id="custom-run-btn" class="btn-run">
-          <span id="run-btn-label">Edit Image</span>
-        </button>
-        <button id="custom-cancel-btn" class="btn-cancel" style="display:none;">Cancel</button>
-        <div class="run-hint">⌘/Ctrl + Enter</div>
-      </div>
-
-      <div class="insp-section">
-        <div class="insp-section-title"><span>Advanced</span></div>
-        <div class="slider-row">
-          <label>Seed</label>
-          <input type="range" id="custom-seed" min="0" max="2147483647" step="1" value="0">
-          <span class="slider-val" id="custom-seed-val">0</span>
-          <button class="dice-btn" id="dice-btn" title="Random seed">
-            <svg viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M19 3H5c-1.1 0-2 .9-2 2v14c0 1.1.9 2 2 2h14c1.1 0 2-.9 2-2V5c0-1.1-.9-2-2-2zM7.5 18c-.83 0-1.5-.67-1.5-1.5S6.67 15 7.5 15s1.5.67 1.5 1.5S8.33 18 7.5 18zm0-9C6.67 9 6 8.33 6 7.5S6.67 6 7.5 6 9 6.67 9 7.5 8.33 9 7.5 9zm4.5 4.5c-.83 0-1.5-.67-1.5-1.5s.67-1.5 1.5-1.5 1.5.67 1.5 1.5-.67 1.5-1.5 1.5zm4.5 4.5c-.83 0-1.5-.67-1.5-1.5s.67-1.5 1.5-1.5 1.5.67 1.5 1.5-.67 1.5-1.5 1.5zm0-9c-.83 0-1.5-.67-1.5-1.5S15.67 6 16.5 6s1.5.67 1.5 1.5S17.33 9 16.5 9z"/></svg>
-          </button>
-        </div>
-        <div class="checkbox-row">
-          <input type="checkbox" id="custom-randomize" checked>
-          <label for="custom-randomize">Randomize seed</label>
-        </div>
-        <div class="slider-row">
-          <label>Guidance</label>
-          <input type="range" id="custom-guidance" min="1" max="10" step="0.1" value="1.0">
-          <span class="slider-val" id="custom-guidance-val">1.0</span>
-        </div>
-        <div class="slider-row">
-          <label>Steps</label>
-          <input type="range" id="custom-steps" min="1" max="50" step="1" value="4">
-          <span class="slider-val" id="custom-steps-val">4</span>
-        </div>
-      </div>
-    </div>
-
-    <!-- Examples page -->
-    <div class="insp-page" id="page-examples">
-      <div class="examples-list" id="examples-scroll"></div>
-    </div>
-  </div>
-
-  <!-- ══ Status bar ══ -->
-  <div class="statusbar">
-    <div class="sb-section" id="sb-image-count">No inputs</div>
-    <div class="sb-right">
-      <a class="sb-link" href="https://huggingface.co/Qwen/Qwen-Image-Edit-2509" target="_blank" rel="noopener">Qwen-Image-Edit-2509</a>
-      <span class="sb-pill" id="sb-status">Ready</span>
-    </div>
-  </div>
-
-</div>
-
-<input id="custom-file-input" type="file" accept="image/*" multiple style="display:none;" />
-
-<!-- Lightbox -->
-<div class="lightbox" id="lightbox"><img id="lightbox-img" alt="zoom"></div>
-
-<!-- Global drop overlay -->
-<div class="drop-overlay" id="drop-overlay"><div class="drop-overlay-inner">Drop images to upload</div></div>
-
-<script type="module">
-import { Client } from "https://cdn.jsdelivr.net/npm/@gradio/client/dist/index.min.js";
-
-/* ── DOM refs ── */
-const galleryGrid  = document.getElementById('image-gallery-grid');
-const fileInput    = document.getElementById('custom-file-input');
-const btnUpload    = document.getElementById('tb-upload');
-const btnRemove    = document.getElementById('tb-remove');
-const btnClear     = document.getElementById('tb-clear');
-const promptInput  = document.getElementById('custom-prompt-input');
-const charCount    = document.getElementById('char-count');
-const loraSelect   = document.getElementById('custom-lora-select');
-const runBtn       = document.getElementById('custom-run-btn');
-const runBtnLabel  = document.getElementById('run-btn-label');
-const cancelBtn    = document.getElementById('custom-cancel-btn');
-const imgCountTb   = document.getElementById('tb-image-count');
-const imgCountSb   = document.getElementById('sb-image-count');
-const sbStatus     = document.getElementById('sb-status');
-const loader       = document.getElementById('output-loader');
-const loaderText   = document.getElementById('loader-text');
-const loaderFill   = document.getElementById('loader-bar-fill');
-const outBody      = document.getElementById('output-image-container');
-const outPh        = document.getElementById('output-placeholder');
-const dlBtn        = document.getElementById('dl-btn-output');
-const compareBtn   = document.getElementById('compare-btn');
-const useAsInputBtn= document.getElementById('use-as-input-btn');
-const zoomBtn      = document.getElementById('zoom-btn');
-const seedChip     = document.getElementById('out-seed-chip');
-const historyStrip = document.getElementById('history-strip');
-const historyCount = document.getElementById('history-count');
-const seedSlider   = document.getElementById('custom-seed');
-const seedVal      = document.getElementById('custom-seed-val');
-const lightbox     = document.getElementById('lightbox');
-const lightboxImg  = document.getElementById('lightbox-img');
-const dropOverlay  = document.getElementById('drop-overlay');
-const compareWrap  = document.getElementById('compare-wrap');
-const compareTop   = document.getElementById('compare-top');
-const compareDivider = document.getElementById('compare-divider');
-const compareBeforeImg = document.getElementById('compare-before-img');
-const compareAfterImg  = document.getElementById('compare-after-img');
-const canvasMeta   = document.getElementById('canvas-meta');
-const segResult    = document.getElementById('seg-result');
-const segInput     = document.getElementById('seg-input');
-const connDot      = document.getElementById('conn-dot');
-const connText     = document.getElementById('conn-text');
-
-/* ── State (multi-image) ── */
-let images = [];
-let selectedIdx = -1;
-let toastTimer = null;
-let running = false;
-let currentJob = null;
-let currentResult = null;   // {b64, seed, prompt, beforeB64}
-let history = [];
-let historyIdx = -1;
-let compareMode = false;
-let viewMode = 'result';    // 'result' | 'input'
-
-const SUGGESTIONS = [
-    ['Add Logo', 'Add the batman logo to the image while preserving the background lighting and surrounding elements maintaining realism and original details.'],
-    ['Add Glasses', 'Add the slim rectangular transparent frame sunglasses to the image while preserving the background lighting and surrounding elements maintaining realism and original details.'],
-    ['Remove Objects', 'Remove the necklace and goggles from the image while preserving the background and remaining elements, maintaining realism and original details.'],
-    ['Zoom', 'Zoom into the red highlighted area.'],
-    ['Extract', 'Extract the clothing and create a flat mockup.'],
-];
-
-/* ── Toast ── */
-function showToast(message, type) {
-    let toast = document.getElementById('app-toast');
-    if (!toast) {
-        toast = document.createElement('div');
-        toast.id = 'app-toast';
-        toast.className = 'toast-notification';
-        toast.innerHTML = '<span class="toast-icon"></span><span class="toast-text"></span>';
-        document.body.appendChild(toast);
-    }
-    const icon = toast.querySelector('.toast-icon');
-    const text = toast.querySelector('.toast-text');
-    toast.className = 'toast-notification ' + (type || 'error');
-    icon.textContent = type === 'warning' ? '⚠' : type === 'info' ? 'ℹ' : '✗';
-    text.textContent = message;
-    if (toastTimer) clearTimeout(toastTimer);
-    void toast.offsetWidth;
-    toast.classList.add('visible');
-    toastTimer = setTimeout(() => toast.classList.remove('visible'), 3500);
-}
-
-function flashPromptError() {
-    promptInput.classList.add('error-flash');
-    promptInput.focus();
-    setTimeout(() => promptInput.classList.remove('error-flash'), 800);
-}
-
-function setStatus(text, cls) {
-    sbStatus.textContent = text;
-    sbStatus.className = 'sb-pill' + (cls ? ' ' + cls : '');
-}
-
-/* ── View mode ── */
-function setViewMode(mode) {
-    viewMode = mode;
-    segResult.classList.toggle('active', mode === 'result');
-    segInput.classList.toggle('active', mode === 'input');
-    renderCanvas();
-}
-segResult.addEventListener('click', () => setViewMode('result'));
-segInput.addEventListener('click', () => setViewMode('input'));
-
-function getCanvasImg() { return outBody.querySelector('img.modern-out-img'); }
-
-function renderCanvas() {
-    exitCompareMode();
-    let img = getCanvasImg();
-    const showResult = viewMode === 'result' && currentResult;
-    const showInput  = viewMode === 'input' && images.length > 0;
-
-    if (showResult || showInput) {
-        outPh.style.display = 'none';
-        if (!img) {
-            img = document.createElement('img');
-            img.className = 'modern-out-img';
-            img.addEventListener('click', () => {
-                if (img.src) { lightboxImg.src = img.src; lightbox.classList.add('visible'); }
-            });
-            outBody.appendChild(img);
-        }
-        img.style.display = '';
-        if (showResult) {
-            img.src = currentResult.b64;
-            canvasMeta.textContent = (currentResult.lora || 'result') + ' · seed ' + currentResult.seed;
-        } else {
-            const i = selectedIdx >= 0 ? selectedIdx : 0;
-            img.src = images[i].b64;
-            canvasMeta.textContent = 'input #' + (i + 1) + ' · ' + (images[i].name || 'image');
-        }
-    } else {
-        if (img) img.style.display = 'none';
-        outPh.style.display = '';
-        canvasMeta.textContent = images.length > 0
-            ? images.length + ' input' + (images.length > 1 ? 's' : '') + ' ready — run an edit'
-            : 'No image loaded';
-    }
-    updateRailState();
-}
-
-function updateRailState() {
-    const hasResult = !!currentResult;
-    dlBtn.disabled = !hasResult;
-    zoomBtn.disabled = !hasResult && images.length === 0;
-    useAsInputBtn.disabled = !hasResult;
-    compareBtn.disabled = !(hasResult && currentResult.beforeB64);
-    btnRemove.disabled = selectedIdx < 0;
-    btnClear.disabled = images.length === 0;
-}
-
-/* ── Inputs filmstrip ── */
-function updateCounts() {
-    const n = images.length;
-    imgCountTb.textContent = n;
-    imgCountSb.textContent = n > 0 ? n + ' input' + (n > 1 ? 's' : '') + ' · ' + history.length + ' result' + (history.length === 1 ? '' : 's') : 'No inputs';
-}
-
-function addImage(b64, name) {
-    images.push({id: Date.now() + Math.random(), b64, name});
-    if (selectedIdx < 0) selectedIdx = 0;
-    renderFilmstrip(); updateCounts();
-    if (!currentResult || viewMode === 'input') renderCanvas();
-}
-
-function removeImage(idx) {
-    images.splice(idx, 1);
-    if (selectedIdx === idx) selectedIdx = images.length ? 0 : -1;
-    else if (selectedIdx > idx) selectedIdx--;
-    renderFilmstrip(); updateCounts(); renderCanvas();
-}
-
-function clearAll() {
-    images = []; selectedIdx = -1;
-    renderFilmstrip(); updateCounts(); renderCanvas();
-}
-
-function renderFilmstrip() {
-    galleryGrid.innerHTML = '';
-    if (images.length === 0) {
-        const empty = document.createElement('div');
-        empty.className = 'fs-empty';
-        empty.textContent = 'Drop or paste images…';
-        galleryGrid.appendChild(empty);
-    }
-    images.forEach((img, i) => {
-        const t = document.createElement('div');
-        t.className = 'fs-thumb' + (i === selectedIdx ? ' selected' : '');
-        t.innerHTML = '<img src="' + img.b64 + '" alt="">' +
-                      '<span class="fs-badge">#' + (i + 1) + '</span>' +
-                      '<button class="fs-remove">✕</button>';
-        t.addEventListener('click', (e) => {
-            if (e.target.closest('.fs-remove')) return;
-            selectedIdx = i;
-            renderFilmstrip();
-            setViewMode('input');
-        });
-        t.querySelector('.fs-remove').addEventListener('click', (e) => {
-            e.stopPropagation(); removeImage(i);
-        });
-        galleryGrid.appendChild(t);
-    });
-    const add = document.createElement('div');
-    add.className = 'fs-add';
-    add.innerHTML = '<span class="add-icon">+</span><span class="add-text">Add</span>';
-    add.addEventListener('click', () => fileInput.click());
-    galleryGrid.appendChild(add);
-    updateRailState();
-}
-
-function processFiles(files) {
-    let added = 0;
-    Array.from(files).forEach(file => {
-        if (!file.type.startsWith('image/')) return;
-        added++;
-        const reader = new FileReader();
-        reader.onload = (e) => addImage(e.target.result, file.name);
-        reader.readAsDataURL(file);
-    });
-    return added;
-}
-
-fileInput.addEventListener('change', (e) => { processFiles(e.target.files); e.target.value = ''; });
-btnUpload.addEventListener('click',  () => fileInput.click());
-btnRemove.addEventListener('click',  () => { if (selectedIdx >= 0) removeImage(selectedIdx); });
-btnClear.addEventListener('click',   clearAll);
-
-/* ── Drop anywhere ── */
-let dragDepth = 0;
-window.addEventListener('dragenter', (e) => {
-    if (!e.dataTransfer || !Array.from(e.dataTransfer.types).includes('Files')) return;
-    e.preventDefault();
-    dragDepth++;
-    dropOverlay.classList.add('visible');
-});
-window.addEventListener('dragleave', (e) => {
-    if (!e.dataTransfer || !Array.from(e.dataTransfer.types).includes('Files')) return;
-    dragDepth = Math.max(0, dragDepth - 1);
-    if (dragDepth === 0) dropOverlay.classList.remove('visible');
-});
-window.addEventListener('dragover', (e) => { e.preventDefault(); });
-window.addEventListener('drop', (e) => {
-    e.preventDefault();
-    dragDepth = 0;
-    dropOverlay.classList.remove('visible');
-    if (e.dataTransfer.files.length) {
-        const n = processFiles(e.dataTransfer.files);
-        if (n > 0) showToast('Added ' + n + ' image' + (n > 1 ? 's' : ''), 'info');
-    }
-});
-
-/* ── Paste ── */
-document.addEventListener('paste', (e) => {
-    if (e.target === promptInput) return;
-    const items = e.clipboardData && e.clipboardData.items;
-    if (!items) return;
-    const files = [];
-    for (const item of items) {
-        if (item.kind === 'file' && item.type.startsWith('image/')) {
-            const f = item.getAsFile();
-            if (f) files.push(f);
-        }
-    }
-    if (files.length) {
-        e.preventDefault();
-        processFiles(files);
-        showToast('Pasted ' + files.length + ' image' + (files.length > 1 ? 's' : ''), 'info');
-    }
-});
-
-/* ── Inspector tabs ── */
-const tabEdit = document.getElementById('tab-edit');
-const tabExamples = document.getElementById('tab-examples');
-const pageEdit = document.getElementById('page-edit');
-const pageExamples = document.getElementById('page-examples');
-tabEdit.addEventListener('click', () => {
-    tabEdit.classList.add('active'); tabExamples.classList.remove('active');
-    pageEdit.classList.add('active'); pageExamples.classList.remove('active');
-});
-tabExamples.addEventListener('click', () => {
-    tabExamples.classList.add('active'); tabEdit.classList.remove('active');
-    pageExamples.classList.add('active'); pageEdit.classList.remove('active');
-});
-
-/* ── Suggestion chips ── */
-const suggWrap = document.getElementById('suggestions-wrap');
-SUGGESTIONS.forEach(([label, prompt]) => {
-    const chip = document.createElement('button');
-    chip.className = 'suggestion-chip';
-    chip.textContent = label;
-    chip.addEventListener('click', () => {
-        promptInput.value = prompt;
-        updateCharCount();
-        suggWrap.querySelectorAll('.suggestion-chip').forEach(c => c.classList.remove('active'));
-        chip.classList.add('active');
-        promptInput.focus();
-    });
-    suggWrap.appendChild(chip);
-});
-
-function updateCharCount() { charCount.textContent = promptInput.value.length; }
-promptInput.addEventListener('input', () => {
-    updateCharCount();
-    suggWrap.querySelectorAll('.suggestion-chip').forEach(c => c.classList.remove('active'));
-});
-updateCharCount();
-
-/* ── Sliders ── */
-function bindSlider(id) {
-    const slider  = document.getElementById(id);
-    const valSpan = document.getElementById(id + '-val');
-    slider.addEventListener('input', () => { valSpan.textContent = slider.value; });
-}
-bindSlider('custom-seed');
-bindSlider('custom-guidance');
-bindSlider('custom-steps');
-
-document.getElementById('dice-btn').addEventListener('click', () => {
-    const s = Math.floor(Math.random() * 2147483647);
-    seedSlider.value = s; seedVal.textContent = s;
-});
-
-/* ── Loader ── */
-function showLoader(text) {
-    loaderText.textContent = text || 'Processing image…';
-    loaderFill.classList.remove('determinate');
-    loaderFill.style.width = '100%';
-    loader.classList.add('active');
-    setStatus('Processing…');
-}
-function setLoaderProgress(text, frac) {
-    loaderText.textContent = text;
-    if (frac != null && frac >= 0 && frac <= 1) {
-        loaderFill.classList.add('determinate');
-        loaderFill.style.width = Math.max(4, Math.round(frac * 100)) + '%';
-    }
-}
-function hideLoader(statusText, cls) {
-    loader.classList.remove('active');
-    setStatus(statusText || 'Done', cls || 'done');
-}
-
-/* ── Result / history ── */
-function showResult(entry) {
-    currentResult = entry;
-    setViewMode('result');
-    dlBtn.disabled = false;
-    useAsInputBtn.disabled = false;
-    zoomBtn.disabled = false;
-    if (entry.beforeB64) compareBtn.disabled = false;
-    if (entry.seed != null) {
-        seedChip.textContent = 'seed ' + entry.seed;
-        seedChip.classList.add('visible');
-    }
-}
-
-function renderHistory() {
-    historyStrip.innerHTML = '';
-    historyCount.textContent = history.length;
-    if (history.length === 0) {
-        const empty = document.createElement('div');
-        empty.className = 'fs-empty';
-        empty.textContent = 'No results yet';
-        historyStrip.appendChild(empty);
-        return;
-    }
-    history.forEach((entry, i) => {
-        const t = document.createElement('div');
-        t.className = 'fs-thumb' + (i === historyIdx ? ' selected' : '');
-        t.title = (entry.lora || '') + ' · seed ' + entry.seed;
-        const im = document.createElement('img');
-        im.src = entry.b64;
-        t.appendChild(im);
-        t.addEventListener('click', () => { historyIdx = i; showResult(history[i]); renderHistory(); });
-        historyStrip.appendChild(t);
-    });
-}
-
-function pushHistory(entry) {
-    history.push(entry);
-    historyIdx = history.length - 1;
-    renderHistory();
-    updateCounts();
-}
-
-/* ── Compare slider ── */
-function enterCompareMode() {
-    if (!currentResult || !currentResult.beforeB64) return;
-    compareMode = true;
-    compareBtn.classList.add('active');
-    const img = getCanvasImg();
-    if (img) img.style.display = 'none';
-    compareAfterImg.src = currentResult.b64;
-    compareBeforeImg.src = currentResult.beforeB64;
-    compareWrap.classList.add('visible');
-    requestAnimationFrame(() => setComparePos(0.5));
-}
-function exitCompareMode() {
-    if (!compareMode) return;
-    compareMode = false;
-    compareBtn.classList.remove('active');
-    compareWrap.classList.remove('visible');
-    const img = getCanvasImg();
-    if (img && (viewMode === 'result' && currentResult || viewMode === 'input' && images.length)) img.style.display = '';
-}
-function setComparePos(frac) {
-    frac = Math.max(0, Math.min(1, frac));
-    const w = compareWrap.clientWidth;
-    const x = frac * w;
-    compareTop.style.clipPath = 'inset(0 ' + (w - x) + 'px 0 0)';
-    compareDivider.style.left = x + 'px';
-}
-compareBtn.addEventListener('click', () => { compareMode ? exitCompareMode() : enterCompareMode(); });
-
-let compareDragging = false;
-function compareMove(clientX) {
-    const rect = compareWrap.getBoundingClientRect();
-    setComparePos((clientX - rect.left) / rect.width);
-}
-compareWrap.addEventListener('pointerdown', (e) => {
-    compareDragging = true;
-    compareWrap.setPointerCapture(e.pointerId);
-    compareMove(e.clientX);
-});
-compareWrap.addEventListener('pointermove', (e) => { if (compareDragging) compareMove(e.clientX); });
-compareWrap.addEventListener('pointerup',   () => { compareDragging = false; });
-compareWrap.addEventListener('pointercancel', () => { compareDragging = false; });
-
-/* ── Lightbox ── */
-lightbox.addEventListener('click', () => lightbox.classList.remove('visible'));
-document.addEventListener('keydown', (e) => { if (e.key === 'Escape') lightbox.classList.remove('visible'); });
-zoomBtn.addEventListener('click', () => {
-    const img = getCanvasImg();
-    const src = (viewMode === 'result' && currentResult) ? currentResult.b64
-              : (images.length ? images[selectedIdx >= 0 ? selectedIdx : 0].b64 : null);
-    if (src) { lightboxImg.src = src; lightbox.classList.add('visible'); }
-});
-
-/* ── Download / chain / seed ── */
-dlBtn.addEventListener('click', () => {
-    if (currentResult && currentResult.b64) {
-        const a = document.createElement('a');
-        a.href = currentResult.b64;
-        a.download = 'qwen_manipulator_' + (currentResult.seed != null ? currentResult.seed : 'output') + '.png';
-        document.body.appendChild(a); a.click(); document.body.removeChild(a);
-    }
-});
-useAsInputBtn.addEventListener('click', () => {
-    if (!currentResult) return;
-    addImage(currentResult.b64, 'edit_seed_' + (currentResult.seed != null ? currentResult.seed : 'x') + '.png');
-    showToast('Result added as input image', 'info');
-});
-seedChip.addEventListener('click', () => {
-    if (currentResult && currentResult.seed != null) {
-        navigator.clipboard.writeText(String(currentResult.seed)).then(
-            () => showToast('Seed copied: ' + currentResult.seed, 'info'), () => {});
-    }
-});
-
-/* ── Backend connection ── */
-let client = null;
-try {
-    client = await Client.connect(window.location.origin);
-    setStatus('Ready');
-    connDot.classList.remove('off');
-    connText.textContent = 'connected';
-} catch (e) {
-    console.error('Failed to connect to Gradio backend:', e);
-    setStatus('Offline', 'error');
-    connDot.classList.add('off');
-    connText.textContent = 'offline';
-    showToast('Could not connect to the backend API', 'error');
-}
-
-/* ── Config: LoRAs + examples ── */
-function renderConfig(cfg) {
-    loraSelect.innerHTML = '';
-    cfg.loras.forEach(name => {
-        const opt = document.createElement('option');
-        opt.value = name; opt.textContent = name;
-        loraSelect.appendChild(opt);
-    });
-    if (cfg.default_lora) loraSelect.value = cfg.default_lora;
-
-    const list = document.getElementById('examples-scroll');
-    list.innerHTML = '';
-    cfg.examples.forEach(ex => {
-        const card = document.createElement('div');
-        card.className = 'example-card';
-        const thumbs = ex.thumbs.map(t =>
-            t ? '<img src="' + t + '" alt="">' : '<div class="example-thumb-placeholder">Preview</div>'
-        ).join('');
-        const promptShort = ex.prompt.length > 90 ? ex.prompt.slice(0, 90) + '…' : ex.prompt;
-        card.innerHTML =
-            '<div class="example-thumb-wrap">' + thumbs + '</div>'
-            + '<div class="example-body">'
-            +   '<div class="example-meta">'
-            +     '<span class="example-lora-badge">' + ex.n_images + ' img' + (ex.n_images > 1 ? 's' : '') + '</span>'
-            +   '</div>'
-            +   '<div class="example-prompt-text"></div>'
-            + '</div>';
-        card.querySelector('.example-prompt-text').textContent = promptShort;
-        card.addEventListener('click', () => loadExample(card, ex.idx));
-        list.appendChild(card);
-    });
-}
-
-async function loadExample(card, idx) {
-    if (!client) { showToast('Backend not connected', 'error'); return; }
-    document.querySelectorAll('.example-card.loading').forEach(c => c.classList.remove('loading'));
-    card.classList.add('loading');
-    showToast('Loading example…', 'info');
-    try {
-        const result = await client.predict('/load_example', { idx: idx });
-        const data = result.data[0];
-        if (data && data.status === 'ok' && data.images && data.images.length > 0) {
-            clearAll();
-            if (data.prompt) { promptInput.value = data.prompt; updateCharCount(); }
-            if (data.lora)   loraSelect.value = data.lora;
-            data.images.forEach((b64, i) => {
-                const name = (data.names && data.names[i]) ? data.names[i] : ('example_' + (i + 1) + '.jpg');
-                addImage(b64, name);
-            });
-            tabEdit.click();
-            showToast('Example loaded — ' + data.images.length + ' image(s)', 'info');
-        } else {
-            showToast('Could not load example images', 'error');
-        }
-    } catch (e) {
-        console.error('Example load error:', e);
-        showToast('Could not load example images', 'error');
-    } finally {
-        card.classList.remove('loading');
-    }
-}
-
-try {
-    const cfg = await fetch('api/config').then(r => r.json());
-    renderConfig(cfg);
-} catch (e) {
-    console.error('Failed to load config:', e);
-}
-
-/* ── Run ── */
-function validateBeforeRun() {
-    const promptVal = promptInput.value.trim();
-    const hasImages = images.length > 0;
-    if (!hasImages && !promptVal) { showToast('Please upload an image and enter a prompt', 'error'); flashPromptError(); return false; }
-    if (!hasImages)  { showToast('Please upload at least one image', 'error'); return false; }
-    if (!promptVal)  { showToast('Please enter an edit prompt', 'warning'); flashPromptError(); return false; }
-    return true;
-}
-
-function setRunningUI(on) {
-    running = on;
-    runBtn.disabled = on;
-    runBtnLabel.textContent = on ? 'Editing…' : 'Edit Image';
-    cancelBtn.style.display = on ? 'flex' : 'none';
-}
-
-async function runEdit() {
-    if (running) return;
-    if (!client) { showToast('Backend not connected', 'error'); return; }
-    if (!validateBeforeRun()) return;
-
-    setRunningUI(true);
-    setViewMode('result');
-    showLoader('Submitting to queue…');
-
-    const beforeB64 = images[0] ? images[0].b64 : null;
-    const usedPrompt = promptInput.value;
-    const usedLora = loraSelect.value;
-
-    const payload = {
-        images_b64_json: JSON.stringify(images.map(img => img.b64)),
-        prompt: usedPrompt,
-        lora_adapter: usedLora,
-        seed: parseInt(seedSlider.value) || 0,
-        randomize_seed: document.getElementById('custom-randomize').checked,
-        guidance_scale: parseFloat(document.getElementById('custom-guidance').value),
-        steps: parseInt(document.getElementById('custom-steps').value),
-    };
-
-    try {
-        currentJob = client.submit('/edit_image', payload);
-        for await (const msg of currentJob) {
-            if (msg.type === 'status') {
-                if (msg.status === 'pending') {
-                    const pos = (msg.position != null && msg.position >= 0)
-                        ? 'In queue — position ' + (msg.position + 1) + (msg.queue_size ? ' of ' + msg.queue_size : '')
-                        : 'Waiting in queue…';
-                    setLoaderProgress(pos, null);
-                    setStatus('Queued');
-                } else if (msg.status === 'generating') {
-                    setStatus('Processing…');
-                    const pd = msg.progress_data;
-                    if (pd && pd.length > 0) {
-                        const p = pd[pd.length - 1];
-                        if (p.index != null && p.length) {
-                            const pct = Math.round((p.index / p.length) * 100);
-                            setLoaderProgress('Generating… ' + p.index + '/' + p.length + ' (' + pct + '%)', p.index / p.length);
-                        } else {
-                            setLoaderProgress('Generating…', null);
-                        }
-                    } else {
-                        setLoaderProgress('Generating image…', null);
-                    }
-                } else if (msg.status === 'error') {
-                    throw new Error(msg.message || 'Generation failed');
-                }
-            } else if (msg.type === 'data') {
-                const out = msg.data && msg.data[0];
-                if (out && out.image) {
-                    const entry = { b64: out.image, seed: out.seed, prompt: usedPrompt, lora: usedLora, beforeB64 };
-                    showResult(entry);
-                    pushHistory(entry);
-                    if (out.seed != null) {
-                        seedSlider.value = out.seed;
-                        seedVal.textContent = out.seed;
-                    }
-                    hideLoader('Done', 'done');
-                    showToast('Image edited successfully', 'info');
-                } else {
-                    hideLoader('Done', 'done');
-                }
-            }
-        }
-    } catch (e) {
-        console.error('Edit error:', e);
-        hideLoader('Error', 'error');
-        const msg = (e && e.message) ? e.message : 'Generation failed';
-        showToast(msg.length > 120 ? msg.slice(0, 120) + '…' : msg, 'error');
-    } finally {
-        setRunningUI(false);
-        currentJob = null;
-    }
-}
-
-runBtn.addEventListener('click', runEdit);
-
-cancelBtn.addEventListener('click', async () => {
-    if (currentJob) {
-        try { await currentJob.cancel(); } catch (e) { console.warn('Cancel error:', e); }
-        currentJob = null;
-    }
-    setRunningUI(false);
-    hideLoader('Cancelled');
-    showToast('Generation cancelled', 'warning');
-});
-
-document.addEventListener('keydown', (e) => {
-    if ((e.metaKey || e.ctrlKey) && e.key === 'Enter') {
-        e.preventDefault();
-        runEdit();
-    }
-});
-
-renderFilmstrip();
-renderHistory();
-updateCounts();
-renderCanvas();
-</script>
-</body>
-</html>

pre-requirements.txt

@@ -1 +1 @@
-pip==26.1.2
+pip>=23.0.0

qwenimage/qwen_fa3_processor.py

@@ -1,183 +1,89 @@
 """
 Paired with a good language model. Thanks!
-
-FA3 is currently broken on Blackwell (sm_100) GPUs; this module detects that
-at import time and falls back to PyTorch scaled-dot-product attention (SDPA)
-automatically.  The public class name / call signature are unchanged.
 """
 
 import torch
-import torch.nn.functional as F
 from typing import Optional, Tuple
 from diffusers.models.transformers.transformer_qwenimage import apply_rotary_emb_qwen
 
-
-# ---------------------------------------------------------------------------
-# FA3 availability check
-# ---------------------------------------------------------------------------
-
-def _is_blackwell() -> bool:
-    """Return True when the current default CUDA device is an sm_100 (Blackwell) GPU."""
-    if not torch.cuda.is_available():
-        return False
-    cap = torch.cuda.get_device_capability()
-    # Blackwell → compute capability 10.x  (sm_100)
-    return cap[0] >= 10
-
-
-_fa3_available: bool = False
-_fa3_unavailable_reason: str = ""
-_flash_attn_func = None
-
-if _is_blackwell():
-    _fa3_unavailable_reason = (
-        "FlashAttention-3 is not yet supported on Blackwell (sm_100) GPUs. "
-        "Falling back to scaled-dot-product attention (SDPA)."
-    )
-else:
-    try:
-        from kernels import get_kernel
-        _k = get_kernel("kernels-community/vllm-flash-attn3")
-        _flash_attn_func = _k.flash_attn_func
-        _fa3_available = True
-    except Exception as e:
-        _fa3_unavailable_reason = (
-            "FlashAttention-3 via Hugging Face `kernels` is unavailable. "
-            f"Tried `get_kernel('kernels-community/vllm-flash-attn3')` and failed with:\n{e}\n"
-            "Falling back to scaled-dot-product attention (SDPA)."
+try:
+    from kernels import get_kernel
+    _k = get_kernel("kernels-community/vllm-flash-attn3")
+    _flash_attn_func = _k.flash_attn_func
+except Exception as e:
+    _flash_attn_func = None
+    _kernels_err = e
+
+
+def _ensure_fa3_available():
+    if _flash_attn_func is None:
+        raise ImportError(
+            "FlashAttention-3 via Hugging Face `kernels` is required. "
+            "Tried `get_kernel('kernels-community/vllm-flash-attn3')` and failed with:\n"
+            f"{_kernels_err}"
         )
 
-
-# ---------------------------------------------------------------------------
-# FA3 custom op (registered only when the kernel loaded successfully)
-# ---------------------------------------------------------------------------
-
-if _fa3_available:
-    @torch.library.custom_op("flash::flash_attn_func", mutates_args=())
-    def flash_attn_func(
-        q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, causal: bool = False
-    ) -> torch.Tensor:
-        # _flash_attn_func returns (output, softmax_lse); we only need output.
-        output, _lse = _flash_attn_func(q, k, v, causal=causal)
-        return output
-
-    @flash_attn_func.register_fake
-    def _flash_attn_func_fake(q, k, v, causal=False):
-        # output shape mirrors q: (batch, seq_len, num_heads, head_dim)
-        return torch.empty_like(q).contiguous()
-
-else:
-    # Provide a stub so call-sites that import the symbol don't break at
-    # module load; the processor will route around it at runtime.
-    def flash_attn_func(
-        q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, causal: bool = False
-    ) -> torch.Tensor:
-        raise RuntimeError(_fa3_unavailable_reason)
-
-
-# ---------------------------------------------------------------------------
-# SDPA fallback helper
-# ---------------------------------------------------------------------------
-
-def _sdpa_attention(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    causal: bool = False,
+@torch.library.custom_op("flash::flash_attn_func", mutates_args=())
+def flash_attn_func(
+    q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, causal: bool = False
 ) -> torch.Tensor:
-    """
-    Scaled dot-product attention using torch.nn.functional.scaled_dot_product_attention.
-
-    Input / output layout: (B, S, H, D_h)  — same as the FA3 kernel.
-    """
-    # SDPA expects (B, H, S, D_h)
-    q = q.transpose(1, 2)
-    k = k.transpose(1, 2)
-    v = v.transpose(1, 2)
+    outputs, lse = _flash_attn_func(q, k, v, causal=causal)
+    return outputs
 
-    out = F.scaled_dot_product_attention(q, k, v, is_causal=causal)
+@flash_attn_func.register_fake
+def _(q, k, v, **kwargs):
+    # two outputs:
+    # 1. output: (batch, seq_len, num_heads, head_dim)
+    # 2. softmax_lse: (batch, num_heads, seq_len) with dtype=torch.float32
+    meta_q = torch.empty_like(q).contiguous()
+    return meta_q #, q.new_empty((q.size(0), q.size(2), q.size(1)), dtype=torch.float32)
 
-    # Back to (B, S, H, D_h)
-    return out.transpose(1, 2)
-
-
-# ---------------------------------------------------------------------------
-# Attention processor
-# ---------------------------------------------------------------------------
 
 class QwenDoubleStreamAttnProcessorFA3:
     """
-    Attention processor for the Qwen double-stream architecture.
-
-    Preferred backend: vLLM FlashAttention-3 via Hugging Face ``kernels``.
-    Automatic fallback: PyTorch ``scaled_dot_product_attention`` (SDPA) when
-    FA3 is unavailable — e.g. on Blackwell (sm_100) GPUs where FA3 is not yet
-    supported, or when the ``kernels`` package is absent.
-
-    Notes / limitations
-    -------------------
-    - Arbitrary attention masks are not supported on the FA3 path.  Pass
-      ``attention_mask=None`` (the default) to stay on the fast path.
-    - On the SDPA path, ``attention_mask`` is likewise ignored; add explicit
-      support here if you need it.
-    - ``encoder_hidden_states`` (text stream) is required.
+    FA3-based attention processor for Qwen double-stream architecture.
+    Computes joint attention over concatenated [text, image] streams using vLLM FlashAttention-3
+    accessed via Hugging Face `kernels`.
+
+    Notes / limitations:
+    - General attention masks are not supported here (FA3 path). `is_causal=False` and no arbitrary mask.
+    - Optional windowed attention / sink tokens / softcap can be plumbed through if you use those features.
+    - Expects an available `apply_rotary_emb_qwen` in scope (same as your non-FA3 processor).
     """
 
-    _attention_backend: str  # set in __init__ after capability detection
+    _attention_backend = "fa3"  # for parity with your other processors, not used internally
 
     def __init__(self):
-        if _fa3_available:
-            self._attention_backend = "fa3"
-        else:
-            import warnings
-            warnings.warn(
-                f"QwenDoubleStreamAttnProcessorFA3: {_fa3_unavailable_reason}",
-                stacklevel=2,
-            )
-            self._attention_backend = "sdpa"
-
-    def _attend(
-        self,
-        q: torch.Tensor,
-        k: torch.Tensor,
-        v: torch.Tensor,
-        causal: bool = False,
-    ) -> torch.Tensor:
-        """Dispatch to FA3 or SDPA depending on what is available."""
-        if self._attention_backend == "fa3":
-            return flash_attn_func(q, k, v, causal=causal)
-        return _sdpa_attention(q, k, v, causal=causal)
+        _ensure_fa3_available()
 
     @torch.no_grad()
     def __call__(
         self,
-        attn,
-        hidden_states: torch.FloatTensor,                          # (B, S_img, D_model)
-        encoder_hidden_states: torch.FloatTensor = None,           # (B, S_txt, D_model)
-        encoder_hidden_states_mask: torch.FloatTensor = None,      # unused
-        attention_mask: Optional[torch.FloatTensor] = None,        # unsupported on FA3 path
-        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        attn,  # Attention module with to_q/to_k/to_v/add_*_proj, norms, to_out, to_add_out, and .heads
+        hidden_states: torch.FloatTensor,                 # (B, S_img, D_model)  image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # (B, S_txt, D_model)  text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,  # unused in FA3 path
+        attention_mask: Optional[torch.FloatTensor] = None,    # unused in FA3 path
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # (img_freqs, txt_freqs)
     ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
-
         if encoder_hidden_states is None:
-            raise ValueError(
-                "QwenDoubleStreamAttnProcessorFA3 requires encoder_hidden_states (text stream)."
-            )
-        if attention_mask is not None and self._attention_backend == "fa3":
-            raise NotImplementedError(
-                "attention_mask is not supported on the FA3 path. "
-                "Either drop the mask or let the processor fall back to SDPA."
-            )
+            raise ValueError("QwenDoubleStreamAttnProcessorFA3 requires encoder_hidden_states (text stream).")
+        if attention_mask is not None:
+            # FA3 kernel path here does not consume arbitrary masks; fail fast to avoid silent correctness issues.
+            raise NotImplementedError("attention_mask is not supported in this FA3 implementation.")
+
+        _ensure_fa3_available()
 
         B, S_img, _ = hidden_states.shape
         S_txt = encoder_hidden_states.shape[1]
 
-        # ---- QKV projections ----
-        img_q = attn.to_q(hidden_states)
+        # ---- QKV projections (image/sample stream) ----
+        img_q = attn.to_q(hidden_states)   # (B, S_img, D)
         img_k = attn.to_k(hidden_states)
         img_v = attn.to_v(hidden_states)
 
-        txt_q = attn.add_q_proj(encoder_hidden_states)
+        # ---- QKV projections (text/context stream) ----
+        txt_q = attn.add_q_proj(encoder_hidden_states)  # (B, S_txt, D)
         txt_k = attn.add_k_proj(encoder_hidden_states)
         txt_v = attn.add_v_proj(encoder_hidden_states)
 
@@ -191,7 +97,7 @@ class QwenDoubleStreamAttnProcessorFA3:
         txt_k = txt_k.unflatten(-1, (H, -1))
         txt_v = txt_v.unflatten(-1, (H, -1))
 
-        # ---- Q/K normalization ----
+        # ---- Q/K normalization (per your module contract) ----
         if getattr(attn, "norm_q", None) is not None:
             img_q = attn.norm_q(img_q)
         if getattr(attn, "norm_k", None) is not None:
@@ -204,22 +110,25 @@ class QwenDoubleStreamAttnProcessorFA3:
         # ---- RoPE (Qwen variant) ----
         if image_rotary_emb is not None:
             img_freqs, txt_freqs = image_rotary_emb
+            # expects tensors shaped (B, S, H, D_h)
             img_q = apply_rotary_emb_qwen(img_q, img_freqs, use_real=False)
             img_k = apply_rotary_emb_qwen(img_k, img_freqs, use_real=False)
             txt_q = apply_rotary_emb_qwen(txt_q, txt_freqs, use_real=False)
             txt_k = apply_rotary_emb_qwen(txt_k, txt_freqs, use_real=False)
 
         # ---- Joint attention over [text, image] along sequence axis ----
-        q = torch.cat([txt_q, img_q], dim=1)  # (B, S_txt + S_img, H, D_h)
+        # Shapes: (B, S_total, H, D_h)
+        q = torch.cat([txt_q, img_q], dim=1)
         k = torch.cat([txt_k, img_k], dim=1)
         v = torch.cat([txt_v, img_v], dim=1)
 
-        out = self._attend(q, k, v, causal=False)  # (B, S_total, H, D_h)
+        # FlashAttention-3 path expects (B, S, H, D_h) and returns (out, softmax_lse)
+        out = flash_attn_func(q, k, v, causal=False)  # out: (B, S_total, H, D_h)
 
         # ---- Back to (B, S, D_model) ----
         out = out.flatten(2, 3).to(q.dtype)
 
-        # ---- Split text / image segments ----
+        # Split back to text / image segments
         txt_attn_out = out[:, :S_txt, :]
         img_attn_out = out[:, S_txt:, :]
 

requirements.txt

@@ -1,13 +1,14 @@
---extra-index-url https://download.pytorch.org/whl/cu130
-
-torch==2.11.0
-torchvision==0.26.0
-transformers==5.14.1
-accelerate==1.14.0
-diffusers==0.39.0
-peft==0.19.1
-gradio==6.25.0
-av==17.1.0
-spaces==0.51.1
-huggingface-hub==1.24.0
-kernels==0.16.0
+git+https://github.com/huggingface/transformers.git@v4.57.3
+git+https://github.com/huggingface/accelerate.git
+git+https://github.com/huggingface/diffusers.git
+git+https://github.com/huggingface/peft.git
+huggingface_hub
+sentencepiece 
+torchvision
+supervision
+kernels
+spaces
+hf_xet
+torch
+numpy
+av