Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,78 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/011d56ec-477d-4fe8-b9f1-12d08c109bdf.webp filter=lfs diff=lfs merge=lfs -text
+examples/01d3fe21-162f-4dc0-8bf6-dcc4165125bb.webp filter=lfs diff=lfs merge=lfs -text
+examples/02d7daf5-a732-4c65-a10b-2778f159d193.webp filter=lfs diff=lfs merge=lfs -text
+examples/05620500-6fb8-406b-9ea9-7b9e1afe699e.webp filter=lfs diff=lfs merge=lfs -text
+examples/0f2d052b-6ccf-46e2-821d-d073ead679ac.webp filter=lfs diff=lfs merge=lfs -text
+examples/0f5d38f9-d04a-498e-aaef-9dae01da8c6d.webp filter=lfs diff=lfs merge=lfs -text
+examples/13233320-a579-489d-8946-7b3274775a8f.webp filter=lfs diff=lfs merge=lfs -text
+examples/16708ba4-0aac-4247-9ccc-3d8c4b4e27d8.webp filter=lfs diff=lfs merge=lfs -text
+examples/25ed3f18-2597-4530-ae92-6eb30b855225.webp filter=lfs diff=lfs merge=lfs -text
+examples/27696b53-3f93-4f15-a7c0-0aecc425b543.webp filter=lfs diff=lfs merge=lfs -text
+examples/2d02ee93-6278-4aa9-8fb9-3ca8d4c89dbb.webp filter=lfs diff=lfs merge=lfs -text
+examples/2fac1fad-886d-4344-b6b2-97ed05cf1b35.webp filter=lfs diff=lfs merge=lfs -text
+examples/386bf1a3-a74f-414e-956d-3532c2d7b1ec.webp filter=lfs diff=lfs merge=lfs -text
+examples/3acf3276-0572-490e-b8ce-334833a6ca20.webp filter=lfs diff=lfs merge=lfs -text
+examples/3e7b1d18-236a-45d7-b919-4bc4504b4a7c.webp filter=lfs diff=lfs merge=lfs -text
+examples/447da49d-dcd0-46ec-9ca8-20127c0f07d1.webp filter=lfs diff=lfs merge=lfs -text
+examples/469fc64a-e35d-484a-8362-f12fcb7212fa.webp filter=lfs diff=lfs merge=lfs -text
+examples/49fd6d29-231a-4254-a507-304c1b6de119.webp filter=lfs diff=lfs merge=lfs -text
+examples/4ab2fd4e-8d3f-4539-842b-87c6d45fbe18.webp filter=lfs diff=lfs merge=lfs -text
+examples/4f703275-173e-405e-9a86-c86cc8fde01b.webp filter=lfs diff=lfs merge=lfs -text
+examples/530f0874-145f-498d-aed4-b6f642137419.webp filter=lfs diff=lfs merge=lfs -text
+examples/57f4af16-40dd-4022-89b4-5cd6e7b8e0b5.webp filter=lfs diff=lfs merge=lfs -text
+examples/5a0bfab1-3527-410c-a98f-c44cc88ea750.webp filter=lfs diff=lfs merge=lfs -text
+examples/6059360b-9b94-4821-9e68-3f05a1e91070.webp filter=lfs diff=lfs merge=lfs -text
+examples/6a84c701-3388-498d-abf8-eea1cf1d1667.webp filter=lfs diff=lfs merge=lfs -text
+examples/6ae02ff5-e65f-45c9-81e9-daca1931b62b.webp filter=lfs diff=lfs merge=lfs -text
+examples/6c04c931-ef74-4d1b-bd26-61dcc9dc56f6.webp filter=lfs diff=lfs merge=lfs -text
+examples/7794a497-5176-4a5e-9aa4-0366dc81e410.webp filter=lfs diff=lfs merge=lfs -text
+examples/7922c916-518b-4c07-a7d1-819ecc4ae88c.webp filter=lfs diff=lfs merge=lfs -text
+examples/7adabfc5-0028-4df3-94f9-e9490ccdeb76.webp filter=lfs diff=lfs merge=lfs -text
+examples/849c51b8-87da-4572-9ec9-d0e514c22ba5.webp filter=lfs diff=lfs merge=lfs -text
+examples/895cfefb-8b12-4d57-89ea-e396f1554440.webp filter=lfs diff=lfs merge=lfs -text
+examples/8c08b1b8-ddfd-428e-86bc-7b07d789f7fa.webp filter=lfs diff=lfs merge=lfs -text
+examples/8c265cf7-51ec-44fd-9436-30fbd96491de.webp filter=lfs diff=lfs merge=lfs -text
+examples/8e26fdb3-8623-4418-b68d-089e760006fe.webp filter=lfs diff=lfs merge=lfs -text
+examples/8ffd565d-7915-4353-b78d-cc11111c71da.webp filter=lfs diff=lfs merge=lfs -text
+examples/90972d17-5686-4a88-a605-6b175137f655.webp filter=lfs diff=lfs merge=lfs -text
+examples/93ada464-6ce0-420c-9f7d-0d6c810c741a.webp filter=lfs diff=lfs merge=lfs -text
+examples/9664b431-48dc-4c5c-9873-3de0468cc717.webp filter=lfs diff=lfs merge=lfs -text
+examples/9bd2638a-d848-4dbc-8a62-78039487bae4.webp filter=lfs diff=lfs merge=lfs -text
+examples/9d4b89d1-ad3b-4527-acb4-c956e01ec078.webp filter=lfs diff=lfs merge=lfs -text
+examples/a64d5702-f199-4667-baea-f0ab25bcf706.webp filter=lfs diff=lfs merge=lfs -text
+examples/a65548fa-e7a8-4a8a-9471-6fb102de66b5.webp filter=lfs diff=lfs merge=lfs -text
+examples/aa74e9d4-7ace-4b1f-9a68-a166976fb1fb.webp filter=lfs diff=lfs merge=lfs -text
+examples/adea5906-f3be-4283-9b31-b357a37800c9.webp filter=lfs diff=lfs merge=lfs -text
+examples/bf8274b8-1db3-4d3e-b185-8a4ee7c98168.webp filter=lfs diff=lfs merge=lfs -text
+examples/c6f9bab3-a041-4bcd-a9ef-5246a65b39d9.webp filter=lfs diff=lfs merge=lfs -text
+examples/c74003dd-cde6-41ed-a66c-d8eb12767b67.webp filter=lfs diff=lfs merge=lfs -text
+examples/d085337f-fdc8-419c-8f6f-fd21f32a612e.webp filter=lfs diff=lfs merge=lfs -text
+examples/d194ae64-f876-4a67-9190-52391a1ada78.webp filter=lfs diff=lfs merge=lfs -text
+examples/df248326-4cc6-4413-a2fe-06d0f3235842.webp filter=lfs diff=lfs merge=lfs -text
+examples/e636800f-257b-4fa5-adb3-853b647f75b1.webp filter=lfs diff=lfs merge=lfs -text
+examples/edf53ce0-2f5b-4f1e-9a36-3aca54237698.webp filter=lfs diff=lfs merge=lfs -text
+examples/eed82398-ad69-4328-b5a3-0cd8952b25b7.webp filter=lfs diff=lfs merge=lfs -text
+examples/ef7735f9-96ff-4da7-a2ba-64dd63fd0250.webp filter=lfs diff=lfs merge=lfs -text
+examples/f263765e-e4d7-4683-87dd-b18fa77280a7.webp filter=lfs diff=lfs merge=lfs -text
+examples/f53a2acb-be77-4799-b041-964c7906459c.webp filter=lfs diff=lfs merge=lfs -text
+examples/f7b1d96d-8f47-40c2-b166-1eaf59bd20ba.webp filter=lfs diff=lfs merge=lfs -text
+examples/f85dbf4f-8504-4b6a-b30d-4681738270a0.webp filter=lfs diff=lfs merge=lfs -text
+examples/fa6cb3bf-2367-4f03-a9dd-a451e7afbe74.webp filter=lfs diff=lfs merge=lfs -text
+examples/fe74ca55-e4ff-49a7-92b5-3b9abe89c72d.webp filter=lfs diff=lfs merge=lfs -text
+examples/ff689e81-11c8-415e-852e-4e08ac0d0ea3.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_en_1.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_en_2.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_en_3.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_en_4.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_running_shoes.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_1.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_2.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_3.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_4.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_5.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_6.webp filter=lfs diff=lfs merge=lfs -text
+examples/generated_zh_7.webp filter=lfs diff=lfs merge=lfs -text
+examples/transparent_anime_bride.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,56 @@
 ---
-title: Qwen Image 2 1
-emoji: 🚀
-colorFrom: green
-colorTo: blue
+title: Qwen-Image-2.1
+emoji: 🖼️
+colorFrom: indigo
+colorTo: pink
 sdk: gradio
 sdk_version: 6.28.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: Generate and edit images with Qwen-Image-2.1
+python_version: "3.12"
+startup_duration_timeout: 1h
+license: other
+license_name: qwen-research
+license_link: https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Qwen-Image-2.1
+
+Text-to-image generation and multi-image editing with
+[`Qwen/Qwen-Image-2.1`](https://huggingface.co/Qwen/Qwen-Image-2.1), running on ZeroGPU
+through the `QwenImage21Pipeline` in `diffusers`.
+
+* **Text to image** — leave the input gallery empty and describe what you want.
+* **Image editing** — upload up to 10 images and refer to them as `<image1>` … `<image10>`
+  in the prompt (upload order).
+* **Transparency** — the model natively decodes RGBA. Ask for it in the prompt, e.g.
+  *"This is an RGBA image with transparency. … The image has alpha channel and the
+  background is transparent."*
+
+## Prompt enhancement
+
+The **Enhance prompt** checkbox calls a companion Space,
+[`hugging-apps/qwen-image-2-1-prompt-enhancer`](https://huggingface.co/spaces/hugging-apps/qwen-image-2-1-prompt-enhancer),
+which keeps both official rewriters (`Qwen-Image-2.1-PE-T2I` and `Qwen-Image-2.1-PE-I2I`)
+warm and resident on their own GPU worker. Splitting them out keeps this Space's GPU
+allocation dedicated to diffusion.
+
+## Safety
+
+Editing requests (any submission with input images) are screened with
+[`hfmlsoc/ncii-guard-v02`](https://huggingface.co/hfmlsoc/ncii-guard-v02) before anything
+is generated.
+
+## Example assets
+
+The images under `examples/` are the official demo assets from the
+[`Qwen/Qwen-Image-2.1`](https://huggingface.co/spaces/Qwen/Qwen-Image-2.1) Space,
+redistributed here under the Qwen Research License for non-commercial demonstration.
+
+## License
+
+> Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026
+> Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
+
+Research / evaluation use only — see the
+[license](https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE).

app.py

@@ -0,0 +1,453 @@
+"""Qwen-Image-2.1 — unified text-to-image generation and image editing on ZeroGPU.
+
+Pipeline for one request:
+
+1. NCII guard (CPU, main process) — image-input requests have their prompt screened
+   with ``hfmlsoc/ncii-guard-v02`` before any GPU time is reserved.
+2. Optional prompt enhancement — delegated over the Gradio API to a companion Space
+   that keeps ``Qwen-Image-2.1-PE-T2I`` and ``Qwen-Image-2.1-PE-I2I`` warm.
+3. Generation — ``QwenImage21Pipeline`` inside ``@spaces.GPU``.
+
+Steps 1 and 2 deliberately run outside the GPU fork so a rejected or enhanced prompt
+costs the visitor no ZeroGPU quota.
+"""
+
+import os
+
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+import spaces  # noqa: E402  — must precede torch / any CUDA-touching import
+
+import json  # noqa: E402
+import math  # noqa: E402
+import random  # noqa: E402
+import tempfile  # noqa: E402
+import time  # noqa: E402
+
+import gradio as gr  # noqa: E402
+import torch  # noqa: E402
+from diffusers import QwenImage21Pipeline  # noqa: E402
+from PIL import Image  # noqa: E402
+from transformers import AutoModelForSequenceClassification, AutoTokenizer  # noqa: E402
+
+MODEL_ID = "Qwen/Qwen-Image-2.1"
+GUARD_ID = "hfmlsoc/ncii-guard-v02"
+PE_SPACE_ID = os.environ.get("PE_SPACE_ID", "hugging-apps/qwen-image-2-1-prompt-enhancer")
+
+MAX_SEED = 2**31 - 1
+MAX_INPUT_IMAGES = 10
+EXAMPLE_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "examples")
+
+# Shown verbatim when the guard rejects a prompt — deliberately non-explicit.
+GUARD_REJECTION_MESSAGE = "prompt invalid based on our classifiers, try again"
+# See the model card's threshold sweep: 0.5 is the balanced precision/recall point.
+GUARD_THRESHOLD = 0.5
+
+# The aspect ratios the model card lists, at its 2048-base resolution.
+BASE_RESOLUTION = 2048
+ASPECT_RATIOS = {
+    "1:1": (2048, 2048),
+    "4:3": (2400, 1792),
+    "3:4": (1792, 2400),
+    "3:2": (2528, 1696),
+    "2:3": (1696, 2528),
+    "16:9": (2752, 1536),
+    "9:16": (1536, 2752),
+}
+RESOLUTION_CHOICES = [("1K — fastest", 1024), ("1.5K", 1536), ("2K — model default", 2048)]
+
+TRANSPARENCY_INFO = (
+    "Describe what to generate, or how to edit the input images (refer to them as "
+    "<image1>…<image10> in upload order). For a transparent RGBA result, phrase the prompt as: "
+    "“This is an RGBA image with transparency. … The image has alpha channel and the "
+    "background is transparent.”"
+)
+
+# --------------------------------------------------------------------------------------
+# Models
+# --------------------------------------------------------------------------------------
+
+print(f"[load] {MODEL_ID} …", flush=True)
+pipe = QwenImage21Pipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
+pipe = pipe.to("cuda")
+print("[load] pipeline ready", flush=True)
+
+# The guard is a 270M classifier and stays on the CPU in the main process, so a prompt can
+# be rejected before any GPU allocation is requested. Its tokenizer carries the homoglyph /
+# zero-width normalizer, so it must come from this repo.
+print(f"[load] {GUARD_ID} …", flush=True)
+guard_tokenizer = AutoTokenizer.from_pretrained(GUARD_ID)
+guard_model = AutoModelForSequenceClassification.from_pretrained(
+    GUARD_ID, dtype=torch.float32
+).eval()
+print("[load] guard ready", flush=True)
+
+
+def guard_score(text: str) -> float:
+    """Probability that ``text`` is a request for non-consensual intimate imagery."""
+    batch = guard_tokenizer(
+        text or "", truncation=True, max_length=256, padding=True, return_tensors="pt"
+    )
+    with torch.no_grad():
+        logits = guard_model(**batch).logits.float()
+    return torch.softmax(logits, dim=-1)[0, 1].item()
+
+
+# --------------------------------------------------------------------------------------
+# Prompt enhancement (delegated to the companion Space)
+# --------------------------------------------------------------------------------------
+
+_pe_client = None
+
+
+def get_pe_client():
+    """Lazily connect to the prompt-enhancement Space (its models are already warm)."""
+    global _pe_client
+    if _pe_client is None:
+        from gradio_client import Client
+
+        token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
+        _pe_client = Client(
+            PE_SPACE_ID, hf_token=token, httpx_kwargs={"timeout": 900}, verbose=False
+        )
+    return _pe_client
+
+
+def enhance_prompt(prompt: str, image_paths: list[str]) -> tuple[str, str]:
+    """Rewrite ``prompt`` with the official PE models. Returns (prompt, wh_ratio)."""
+    from gradio_client import handle_file
+
+    client = get_pe_client()
+    started = time.perf_counter()
+    result = client.predict(
+        prompt=prompt,
+        image_paths=[handle_file(p) for p in image_paths],
+        max_new_tokens=1536,
+        enable_thinking=True,
+        seed=0,
+        randomize_seed=True,
+        api_name="/enhance",
+    )
+    rewritten, wh_ratio = str(result[0]).strip(), str(result[1]).strip()
+    print(f"[enhance] {time.perf_counter() - started:.1f}s -> ratio={wh_ratio!r}", flush=True)
+    return (rewritten or prompt), wh_ratio
+
+
+# --------------------------------------------------------------------------------------
+# Helpers
+# --------------------------------------------------------------------------------------
+
+
+def normalize_gallery(gallery) -> list[str]:
+    """Reduce a ``gr.Gallery`` value to a flat list of local file paths."""
+    paths: list[str] = []
+    for item in gallery or []:
+        if isinstance(item, (tuple, list)):
+            item = item[0]
+        if isinstance(item, dict):
+            item = item.get("path") or item.get("image") or item.get("name")
+            if isinstance(item, dict):
+                item = item.get("path")
+        if isinstance(item, Image.Image):
+            tmp = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
+            item.save(tmp.name)
+            item = tmp.name
+        if isinstance(item, str) and item:
+            paths.append(item)
+    return paths
+
+
+def resolve_size(resolution: int, aspect_ratio: str) -> tuple[int | None, int | None]:
+    """Map a resolution tier + aspect ratio onto the model card's width/height pairs."""
+    if not aspect_ratio or aspect_ratio == "Auto" or aspect_ratio not in ASPECT_RATIOS:
+        return None, None
+    base_w, base_h = ASPECT_RATIOS[aspect_ratio]
+    scale = int(resolution) / BASE_RESOLUTION
+    round32 = lambda v: max(32, int(round(v * scale / 32)) * 32)  # noqa: E731
+    return round32(base_w), round32(base_h)
+
+
+def estimate_duration(
+    input_images=None,
+    prompt="",
+    negative_prompt="",
+    true_cfg_scale=1.0,
+    num_inference_steps=40,
+    seed=0,
+    resolution=1024,
+    aspect_ratio="Auto",
+    *args,
+    **kwargs,
+):
+    """Budget GPU seconds. Cost is ~linear in steps and ~quadratic in side length."""
+    try:
+        steps = int(num_inference_steps)
+        res = int(resolution)
+        cfg = float(true_cfg_scale)
+    except (TypeError, ValueError):
+        steps, res, cfg = 40, 1024, 1.0
+    n_images = len(input_images or [])
+    tokens = (res / 16.0) ** 2
+    per_step = 9.0e-5 * tokens * (2.0 if cfg > 1.0 else 1.0)
+    overhead = 18.0 + 4.0 * n_images + 6.0e-6 * res * res
+    return int(min(360, math.ceil(overhead + steps * per_step)))
+
+
+# --------------------------------------------------------------------------------------
+# Inference
+# --------------------------------------------------------------------------------------
+
+
+@spaces.GPU(duration=estimate_duration)
+def generate_image(
+    input_images=None,
+    prompt: str = "",
+    negative_prompt: str = "",
+    true_cfg_scale: float = 1.0,
+    num_inference_steps: int = 40,
+    seed: int = 0,
+    resolution: int = 1024,
+    aspect_ratio: str = "Auto",
+    progress=gr.Progress(track_tqdm=True),
+):
+    """Generate or edit an image with Qwen-Image-2.1.
+
+    With no input images this is text-to-image; with one to ten input images it is
+    image editing, and the prompt may refer to them as <image1>…<image10>.
+
+    Args:
+        input_images: Up to 10 condition images. Leave empty for text-to-image.
+        prompt: What to generate, or how to edit the inputs.
+        negative_prompt: What to avoid. Only applied when true_cfg_scale > 1.
+        true_cfg_scale: Classifier-free guidance scale; 1.0 disables it.
+        num_inference_steps: Number of flow-matching steps.
+        seed: Random seed.
+        resolution: Target side length in pixels (1024, 1536 or 2048).
+        aspect_ratio: One of the model's supported ratios, or "Auto".
+
+    Returns:
+        The generated image, as RGBA (the model natively supports transparency).
+    """
+    prompt = (prompt or "").strip()
+    if not prompt:
+        raise gr.Error("Please enter a prompt.")
+
+    image_paths = normalize_gallery(input_images)
+    if len(image_paths) > MAX_INPUT_IMAGES:
+        raise gr.Error(f"Up to {MAX_INPUT_IMAGES} input images are supported.")
+
+    condition_images = [Image.open(p) for p in image_paths] or None
+    width, height = resolve_size(resolution, aspect_ratio)
+
+    negative_prompt = (negative_prompt or "").strip()
+    call_kwargs = {}
+    if negative_prompt and float(true_cfg_scale) > 1.0:
+        call_kwargs["negative_prompt"] = negative_prompt
+        call_kwargs["true_cfg_scale"] = float(true_cfg_scale)
+
+    started = time.perf_counter()
+    image = pipe(
+        prompt=prompt,
+        image=condition_images,
+        height=height,
+        width=width,
+        output_resolution=int(resolution),
+        num_inference_steps=int(num_inference_steps),
+        generator=torch.Generator(device="cuda").manual_seed(int(seed) % (MAX_SEED + 1)),
+        **call_kwargs,
+    ).images[0]
+    print(
+        f"[generate] images={len(image_paths)} steps={num_inference_steps} "
+        f"res={resolution} ratio={aspect_ratio} size={image.size} "
+        f"elapsed={time.perf_counter() - started:.1f}s",
+        flush=True,
+    )
+    return image
+
+
+def prepare(prompt, input_images, enhance, seed, randomize_seed, aspect_ratio):
+    """Screen the prompt, optionally rewrite it, and resolve the seed — all off-GPU."""
+    prompt = (prompt or "").strip()
+    if not prompt:
+        raise gr.Error("Please enter a prompt.")
+
+    image_paths = normalize_gallery(input_images)
+    if len(image_paths) > MAX_INPUT_IMAGES:
+        raise gr.Error(f"Up to {MAX_INPUT_IMAGES} input images are supported.")
+
+    # Guard image-input requests before any GPU time is reserved.
+    if image_paths:
+        score = guard_score(prompt)
+        print(f"[guard] score={score:.3f} threshold={GUARD_THRESHOLD}", flush=True)
+        if score >= GUARD_THRESHOLD:
+            raise gr.Error(GUARD_REJECTION_MESSAGE)
+
+    resolved_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+
+    final_prompt, rewritten_display = prompt, ""
+    if enhance:
+        try:
+            final_prompt, wh_ratio = enhance_prompt(prompt, image_paths)
+            rewritten_display = final_prompt
+            # Honour the rewriter's ratio only when the user left the choice open and
+            # there is no input image whose aspect ratio the pipeline should follow.
+            if wh_ratio and aspect_ratio == "Auto" and not image_paths:
+                if wh_ratio in ASPECT_RATIOS:
+                    aspect_ratio = wh_ratio
+        except Exception as exc:  # noqa: BLE001 — never let PE take the whole request down
+            print(f"[enhance] failed: {exc!r}", flush=True)
+            gr.Warning(f"Prompt enhancement unavailable, using the original prompt ({exc}).")
+
+    return final_prompt, rewritten_display, resolved_seed, aspect_ratio
+
+
+# --------------------------------------------------------------------------------------
+# Examples (Qwen's own demo cases, bundled under examples/)
+# --------------------------------------------------------------------------------------
+
+with open(os.path.join(EXAMPLE_DIR, "cases.json"), encoding="utf-8") as fh:
+    DEMO_CASES = json.load(fh)
+
+T2I_EXAMPLES = [[c["prompt"]] for c in DEMO_CASES if not c["inputs"]]
+T2I_LABELS = [c["title_en"] for c in DEMO_CASES if not c["inputs"]]
+EDIT_EXAMPLES = [
+    [[os.path.join(EXAMPLE_DIR, name) for name in c["inputs"]], c["prompt"]]
+    for c in DEMO_CASES
+    if c["inputs"]
+]
+EDIT_LABELS = [c["title_en"] for c in DEMO_CASES if c["inputs"]]
+
+
+# --------------------------------------------------------------------------------------
+# UI
+# --------------------------------------------------------------------------------------
+
+css = """
+#col-container { margin: 0 auto; max-width: 1100px; }
+"""
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=css, title="Qwen-Image-2.1") as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.HTML(
+            '<img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/image2.1/logo.png"'
+            ' alt="Qwen-Image logo" width="380" style="display:block;margin:0 auto;">'
+        )
+
+        with gr.Row():
+            with gr.Column():
+                input_images = gr.Gallery(
+                    label="Input images (editing, up to 10)",
+                    type="filepath",
+                    interactive=True,
+                    columns=5,
+                    height=300,
+                    show_label=True,
+                )
+            with gr.Column():
+                result = gr.Image(
+                    label="Result",
+                    type="pil",
+                    image_mode="RGBA",
+                    format="png",
+                    height=300,
+                    show_label=True,
+                )
+
+        prompt = gr.Textbox(
+            label="Prompt",
+            info=TRANSPARENCY_INFO,
+            lines=5,
+            max_lines=24,
+            placeholder="Describe what you want to generate or edit…",
+        )
+
+        with gr.Row():
+            enhance = gr.Checkbox(label="Enhance prompt", value=True, scale=1)
+            generate_button = gr.Button("Generate image", variant="primary", scale=2)
+
+        rewritten_prompt_output = gr.Textbox(
+            label="Rewritten prompt",
+            lines=4,
+            max_lines=20,
+            interactive=False,
+            show_copy_button=True,
+            placeholder="With prompt enhancement on, the rewritten prompt appears here.",
+        )
+
+        with gr.Accordion("Advanced settings", open=False):
+            with gr.Row():
+                resolution = gr.Dropdown(
+                    label="Resolution", choices=RESOLUTION_CHOICES, value=1024
+                )
+                aspect_ratio = gr.Dropdown(
+                    label="Aspect ratio",
+                    choices=["Auto"] + list(ASPECT_RATIOS),
+                    value="Auto",
+                    info="Auto follows the input image, or squares up for text-to-image.",
+                )
+            num_inference_steps = gr.Slider(
+                label="Inference steps", minimum=8, maximum=60, step=1, value=40
+            )
+            with gr.Row():
+                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
+                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+            negative_prompt = gr.Textbox(
+                label="Negative prompt",
+                value="",
+                lines=2,
+                info="Only applied when guidance is above 1.0.",
+            )
+            true_cfg_scale = gr.Slider(
+                label="True CFG scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0
+            )
+
+        final_prompt_state = gr.State("")
+
+        gr.Markdown("### Examples")
+        gr.Markdown("**Text to image**")
+        gr.Examples(
+            examples=T2I_EXAMPLES,
+            example_labels=T2I_LABELS,
+            inputs=[prompt],
+            fn=generate_image,
+            outputs=[result],
+            cache_examples=True,
+            cache_mode="lazy",
+            label="Text-to-image cases",
+        )
+        gr.Markdown("**Image editing** — the example loads its reference images too")
+        gr.Examples(
+            examples=EDIT_EXAMPLES,
+            example_labels=EDIT_LABELS,
+            inputs=[input_images, prompt],
+            fn=generate_image,
+            outputs=[result],
+            cache_examples=True,
+            cache_mode="lazy",
+            label="Image-editing cases",
+        )
+
+    generate_button.click(
+        fn=prepare,
+        inputs=[prompt, input_images, enhance, seed, randomize_seed, aspect_ratio],
+        outputs=[final_prompt_state, rewritten_prompt_output, seed, aspect_ratio],
+    ).then(
+        fn=generate_image,
+        inputs=[
+            input_images,
+            final_prompt_state,
+            negative_prompt,
+            true_cfg_scale,
+            num_inference_steps,
+            seed,
+            resolution,
+            aspect_ratio,
+        ],
+        outputs=[result],
+        api_name="generate",
+    )
+
+demo.queue(default_concurrency_limit=1, max_size=20)
+
+if __name__ == "__main__":
+    demo.launch(mcp_server=True, allowed_paths=[EXAMPLE_DIR])

examples/README.md

@@ -0,0 +1,15 @@
+# Demo cases
+
+`cases.json` contains the 25 examples imported from `2.1 Demo Case.md`: 21 image-editing cases and 4 text-to-image cases. Input images retain their order from the document, including the 10-image room-furnishing example.
+
+Images are bundled locally as lossless WebP files. Decoded RGB/RGBA pixels were verified against the downloaded originals. Reference outputs are displayed separately from generated results; the running-shoe reference was generated separately because the source document did not supply one.
+
+Where the document supplies both `raw_prompt` and `positive_prompt`, the example uses the complete `positive_prompt`. Original fields remain in `source_prompts`. Markdown code fences, table escapes, and encoded line breaks are removed from the displayed prompt. These curated prompts load with prompt enhancement disabled so they are sent unchanged.
+
+Titles have Chinese and English variants. Switching the interface language preserves the original language and content of each prompt.
+
+## Generated reference outputs
+
+The running-shoe case and all 11 original Chinese/English text-to-image examples now include outputs generated with `pre-qwen-image-2.1-eval-preprocess-0917`, seed 42. Their prompt enhancement settings match the corresponding example buttons. `generated_references.json` records the 11 text-to-image prompts, output filenames and generation settings; the running-shoe metadata is in `cases.json`. Generated outputs are stored as lossless WebP. No API keys or temporary download URLs are stored here.
+
+The transparent anime bride example uses the supplied Chinese prompt verbatim with enhancement disabled. Its reference is an original RGBA PNG with verified transparent and semi-transparent pixels; generation metadata includes the request ID.