Comparison tab (33 Qwen Space examples, turbo 6 steps vs base 40 steps, image slider); 5 examples from the Qwen Space; header: very competitive with the base (part 2)

.gitattributes

@@ -166,3 +166,33 @@ compare/img/case25/base.webp filter=lfs diff=lfs merge=lfs -text
 compare/img/case25/in1.webp filter=lfs diff=lfs merge=lfs -text
 compare/img/case25/ref.webp filter=lfs diff=lfs merge=lfs -text
 compare/img/case25/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_1/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_1/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_2/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_2/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_3/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_3/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_3/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_3/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_4/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_4/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_4/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/en_4/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_1/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_1/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_1/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_1/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_2/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_2/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_2/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_2/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_7/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_7/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/zh_7/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+examples/qwen_mask_edit.png filter=lfs diff=lfs merge=lfs -text
+examples/qwen_mask_edit_in1.webp filter=lfs diff=lfs merge=lfs -text
+examples/qwen_perfume.png filter=lfs diff=lfs merge=lfs -text
+examples/qwen_poster.png filter=lfs diff=lfs merge=lfs -text
+examples/qwen_rgba_bride.png filter=lfs diff=lfs merge=lfs -text
+examples/qwen_storyboard.png filter=lfs diff=lfs merge=lfs -text
+examples/qwen_storyboard_in1.webp filter=lfs diff=lfs merge=lfs -text

compare.py

@@ -0,0 +1,141 @@
+"""The demo's Comparison tab: 33 of the 37 examples of the official Qwen/Qwen-Image-2.1 Space, rendered with viggle-turbo
+v0.2.1 in 6 steps (and in 8 steps on the 5 dense-text examples) and with the base model in 40 steps. Both get the same
+prompt, input images and seed 42, with prompt enhancement off. No GPU: everything is pre-rendered. compare/ (WebP images
++ cases.json) is written by release/build_compare_space.py, and upload_space.sh builds it into the Space."""
+import json
+import statistics
+from pathlib import Path
+
+import gradio as gr
+from PIL import Image
+
+DIR = Path(__file__).parent / "compare"
+CASES = json.loads((DIR / "cases.json").read_text(encoding="utf-8")) if (DIR / "cases.json").exists() else []
+BY_ID = {case["id"]: case for case in CASES}
+FIRST = "case01"  # Character style infographic: a dense layout where we prefer the turbo's render
+KINDS = {
+    "All": lambda case: True,
+    "Editing": lambda case: len(case["inputs"]) > 0,
+    "Multi-image editing": lambda case: len(case["inputs"]) > 1,
+    "Text-to-image": lambda case: not case["inputs"],
+    "Dense text (+8 steps)": lambda case: "turbo8" in case,
+}
+
+
+def arms(case):
+    """(menu label, key) for every render of an example; the render time is in the label."""
+    menu = [(f"viggle-turbo · 6 steps · {case['turbo_s']:.1f} s", "turbo")]
+    if "turbo8" in case:
+        menu.append((f"viggle-turbo · 8 steps · {case['turbo8_s']:.1f} s", "turbo8"))
+    menu.append((f"Qwen-Image-2.1 base · 40 steps · {case['base_s']:.1f} s", "base"))
+    if case["reference"]:
+        menu.append(("Qwen API reference (bundled with Qwen's Space)", "reference"))
+    return menu
+
+
+def image(case, key):
+    if key == "reference":  # Qwen's API output has its own size; match the renders so the two slider halves line up
+        return Image.open(DIR / case["reference"]["src"]).resize(tuple(case["size"]), Image.LANCZOS)
+    return str(DIR / case[key]["src"])
+
+
+def panels(case, left, right):
+    """Slider pair, caption, input images and prompt for one example."""
+    names = dict((key, label) for label, key in arms(case))
+    width, height = case["size"]
+    title = case["title"] if case["title_zh"] == case["title"] else f"{case['title']} · {case['title_zh']}"
+    caption = (f"### {title}\n"
+               f"{width} × {height} · **left:** {names[left]} · **right:** {names[right]} · drag the handle to compare")
+    inputs = [(str(DIR / item["src"]), f"input {i + 1}") for i, item in enumerate(case["inputs"])]
+    return (image(case, left), image(case, right)), caption, inputs, case["prompt"]
+
+
+def thumbs(ids):
+    return [(str(DIR / BY_ID[case_id]["turbo"]["thumb"]), BY_ID[case_id]["title"]) for case_id in ids]
+
+
+def render():
+    """Builds the tab inside the caller's gr.Blocks / gr.Tab context."""
+    edits = [case for case in CASES if case["inputs"]]
+    t2is = [case for case in CASES if not case["inputs"]]
+    median = lambda cases, key: statistics.median(case[key] for case in cases)  # noqa: E731
+    speedup = sum(case["base_s"] for case in CASES) / sum(case["turbo_s"] for case in CASES)
+    gr.Markdown(
+        f"**viggle-turbo (6 steps) vs Qwen-Image-2.1 (40 steps)** on {len(CASES)} of the 37 examples of the official "
+        "[Qwen/Qwen-Image-2.1 Space](https://huggingface.co/spaces/Qwen/Qwen-Image-2.1). Both models get the same prompt, "
+        "input images and seed, with prompt enhancement off, one sample each and no seed picking. "
+        f"At **{speedup:.1f}× less time** the turbo is very competitive with the base model, and on some examples "
+        "(the character style infographic, shown first) we prefer its render; complicated edits can still fall short of the base. "
+        "Pick an example on the left, then drag the slider. The menus also offer the 8-step turbo on the dense-text examples "
+        "and Qwen's own API reference where the Space ships one.\n\n"
+        f"Median time, editing: **{median(edits, 'turbo_s'):.1f} s vs {median(edits, 'base_s'):.1f} s** · "
+        f"text-to-image (~4 MP): **{median(t2is, 'turbo_s'):.1f} s vs {median(t2is, 'base_s'):.1f} s** · "
+        f"all {len(CASES)} examples: **{speedup:.1f}× faster** (one NVIDIA B200, one pipeline call each)"
+    )
+    case = BY_ID[FIRST]
+    pair, caption, inputs, prompt = panels(case, "turbo", "base")
+    ids = gr.State([item["id"] for item in CASES])
+    with gr.Row():
+        with gr.Column(scale=1, min_width=280):
+            kind = gr.Radio([(f"{name} ({sum(map(test, CASES))})", name) for name, test in KINDS.items()], value="All", label="Show")
+            gallery = gr.Gallery(thumbs(ids.value), columns=3, height=760, object_fit="cover", allow_preview=False,
+                                 show_label=False, show_download_button=False, show_fullscreen_button=False)
+        with gr.Column(scale=3):
+            head = gr.Markdown(caption)
+            with gr.Row():
+                left = gr.Dropdown(arms(case), value="turbo", label="Left")
+                right = gr.Dropdown(arms(case), value="base", label="Right")
+            slider = gr.ImageSlider(pair, type="filepath", interactive=False, max_height=820, show_label=False)
+            with gr.Row():
+                shown_inputs = gr.Gallery(inputs, label="Input images", columns=3, height=220, visible=bool(inputs), scale=1)
+                shown_prompt = gr.Textbox(prompt, label="Prompt (sent as written, prompt enhancement off)", lines=8, max_lines=8,
+                                          interactive=False, show_copy_button=True, scale=2)
+    with gr.Accordion("How these were made", open=False):
+        gr.Markdown(
+            "- **Turbo:** Qwen-Image-2.1 + the viggle-turbo v0.2.1 LoRA (rank 256), 6 steps on sigmas "
+            "`[1, 0.9375, 0.875, 0.75, 0.5, 0.25]`, no classifier-free guidance, exactly as the Generate tab runs it.\n"
+            "- **Turbo, 8 steps** (the 5 dense-text examples): sigmas `[1, 0.9375, 0.875, 0.75, 0.625, 0.5, 0.25, 0.125]`. "
+            "At 6 steps small Latin text can print twice, like a double exposure (decided in the 0.75 → 0.5 step), and small "
+            "Chinese strokes can get colour blotches (the last 0.25 → 0 step); each added sigma splits one of the two. In our "
+            "OCR tests 8 steps raise word recall from 0.71 to 0.83 on the academic infographic's caption (16 seeds) and from "
+            "0.76 to 0.90 on the architecture board's Chinese labels (4 seeds); the base model scores 0.95 on both.\n"
+            "- **Base:** Qwen-Image-2.1, 40 steps, default scheduler, `true_cfg_scale=1.0` (the pipeline default).\n"
+            "- **Both:** diffusers `QwenImage21Pipeline` in bf16, seed 42. Input images are encoded at 1024² pixels; the output keeps "
+            "the aspect ratio of Qwen's reference at 2048² pixels for text-to-image and 1536² for editing, in multiples of 32.\n"
+            "- **Time:** one pipeline call on one NVIDIA B200 (text encoding, denoising and VAE decoding), after warm-up.\n"
+            "- **Left out (4 of 37):** three Chinese infographics and a subtitled storyboard whose short prompts leave the "
+            "on-image text to Qwen's prompt enhancement; without it both models fill them with made-up characters.\n"
+            "- **Qwen API reference:** the output bundled with the example in Qwen's Space, made with Qwen's API, which may "
+            "differ from the open weights. It is omitted for the 7 text-to-image examples Qwen generated with prompt enhancement on.\n\n"
+            "Prompts, input images and reference outputs come from the "
+            "[Qwen/Qwen-Image-2.1 Space](https://huggingface.co/spaces/Qwen/Qwen-Image-2.1) (Qwen Research License)."
+        )
+
+    def show(case_id, left_key, right_key, fixed="left"):
+        """Keeps the chosen pair when the new example has it (turbo8 and reference are per-example), else turbo vs base.
+        If both menus land on the same render, the one the user just changed (`fixed`) wins and the other moves off it."""
+        case = BY_ID[case_id]
+        keys = [key for _, key in arms(case)]
+        left_key = left_key if left_key in keys else "turbo"
+        right_key = right_key if right_key in keys else "base"
+        if left_key == right_key:
+            other = next(key for key in ("base", "turbo") if key != left_key)
+            left_key, right_key = (left_key, other) if fixed == "left" else (other, right_key)
+        pair, caption, inputs, prompt = panels(case, left_key, right_key)
+        return (case_id, gr.update(choices=arms(case), value=left_key), gr.update(choices=arms(case), value=right_key),
+                pair, caption, gr.update(value=inputs, visible=bool(inputs)), prompt)
+
+    def pick(ids, left_key, right_key, evt: gr.SelectData):
+        return show(ids[evt.index], left_key, right_key)
+
+    def filter_cases(name):
+        ids = [case["id"] for case in CASES if KINDS[name](case)]
+        return ids, thumbs(ids)
+
+    current = gr.State(FIRST)
+    outputs = [current, left, right, slider, head, shown_inputs, shown_prompt]
+    kind.change(filter_cases, inputs=kind, outputs=[ids, gallery])
+    gallery.select(pick, inputs=[ids, left, right], outputs=outputs)
+    # .input, not .change: show() rewrites both menus, which must not fire another show()
+    left.input(show, inputs=[current, left, right], outputs=outputs)
+    right.input(lambda case_id, left_key, right_key: show(case_id, left_key, right_key, "right"), inputs=[current, left, right], outputs=outputs)

examples/manifest.json

@@ -27,6 +27,76 @@
   "size_label": "3:2 · 2496×1664 (2048² area)",
   "enhance": false
  },
+ {
+  "prompt": "Design a refined travel poster for a fictional night train. Render the headline exactly as \"THE MIDNIGHT EXPRESS\" and the subtitle \"A journey under the stars\". A silver train curves through dark blue mountains beneath a crescent moon. Art Deco geometry, ivory and gold lettering, clear typographic hierarchy, generous margins, print-ready composition.",
+  "refs": [
+   null,
+   null,
+   null
+  ],
+  "result": "qwen_poster.png",
+  "info": "Pre-rendered example · seed `42` · 1664×2496 · 6 steps · v0.2.1 LoRA r256 · prompt sent as written (enhancement off)",
+  "used_prompt": "Design a refined travel poster for a fictional night train. Render the headline exactly as \"THE MIDNIGHT EXPRESS\" and the subtitle \"A journey under the stars\". A silver train curves through dark blue mountains beneath a crescent moon. Art Deco geometry, ivory and gold lettering, clear typographic hierarchy, generous margins, print-ready composition.",
+  "steps": 6,
+  "size_label": "2:3 · 1664×2496 (2048² area)",
+  "enhance": false
+ },
+ {
+  "prompt": "生成一张无缝的分镜插画，**现代真实偶像剧风格**（电影感写实摄影，柔和自然光与暖色调，浅景深虚化，韩系都市偶像剧质感，画面精致高级）。\n\n【版面（必须严格遵守）】\n- 恰好 6 个分镜格，**等宽**，排成**单独一横排（1 行 6 列）**；不是网格、不是两行，绝不要把任何一格再拆成上下两块。\n- 每格都是**竖长条矩形**（高明显大于宽，近似 9:16 竖条）。\n- 从左到右依次编号 1、2、3、4、5、6，每格左上角一个圆形数字徽标；**每个数字只出现一次，不重复、不跳号、不缺号。**\n- **每一格里都必须画出角色本人**在做该格的动作，角色占该格画面主体、完整清晰（半身或全身，约占该格高度 2/3 以上）；**严禁出现只有背景、没有人物的空格。**\n- 6 格之间画风、光影氛围、人物一致。输出为一张扁平整图，6 格并排合并，不要输出多张分离文件。\n\n【角色】以输入参考图（含面部特写与正/侧/背三视全身立绘）为角色设计基准：现代年轻女性，棕色波浪长发带空气刘海，妆容清透，穿粉色方领修身长袖针织上衣、棕色灯芯绒过膝半裙配棕色皮带、棕色皮质机车靴。每一格里都是同一个人——面容、发型、服饰、身材比例保持一致，只改变姿势、表情、动作与所处场景；参考图仅定义外观，不要照搬其站姿。把角色自然地融入各自的真实场景中，透视比例正确、光影贴合。\n\n【逐格场景与内容】\n第1格：清晨阳光洒入的现代简约公寓，女孩站在落地窗边，双手捧着马克杯，望向窗外的城市天际线，神情恬静。\n第2格：白天繁华的都市街道人行道，两旁是时尚店铺橱窗和梧桐树，女孩背着单肩包边走边微微回头，浅浅微笑。\n第3格：文艺清新的咖啡馆靠窗卡座，女孩坐着低头看手机，嘴角带笑，窗外光线柔和，桌上放着一杯拿铁。\n第4格：暖色调的独立书店书架走廊，女孩踮脚从书架上抽出一本书，侧脸专注，光线温柔。\n第5格：傍晚金色晚霞下的城市人行天桥，女孩倚着栏杆回眸，背景是虚化的车流与高楼，暖光洒在脸上。\n第6格：夜晚霓虹灯光的街角，女孩仰头微笑看向前方，暖黄路灯与散景光斑环绕，都市夜景氛围。",
+  "refs": [
+   "qwen_storyboard_in1.webp",
+   null,
+   null
+  ],
+  "result": "qwen_storyboard.png",
+  "info": "Pre-rendered example · seed `42` · 2048×1152 · 6 steps · v0.2.1 LoRA r256 · prompt sent as written (enhancement off)",
+  "used_prompt": "生成一张无缝的分镜插画，**现代真实偶像剧风格**（电影感写实摄影，柔和自然光与暖色调，浅景深虚化，韩系都市偶像剧质感，画面精致高级）。\n\n【版面（必须严格遵守）】\n- 恰好 6 个分镜格，**等宽**，排成**单独一横排（1 行 6 列）**；不是网格、不是两行，绝不要把任何一格再拆成上下两块。\n- 每格都是**竖长条矩形**（高明显大于宽，近似 9:16 竖条）。\n- 从左到右依次编号 1、2、3、4、5、6，每格左上角一个圆形数字徽标；**每个数字只出现一次，不重复、不跳号、不缺号。**\n- **每一格里都必须画出角色本人**在做该格的动作，角色占该格画面主体、完整清晰（半身或全身，约占该格高度 2/3 以上）；**严禁出现只有背景、没有人物的空格。**\n- 6 格之间画风、光影氛围、人物一致。输出为一张扁平整图，6 格并排合并，不要输出多张分离文件。\n\n【角色】以输入参考图（含面部特写与正/侧/背三视全身立绘）为角色设计基准：现代年轻女性，棕色波浪长发带空气刘海，妆容清透，穿粉色方领修身长袖���织上衣、棕色灯芯绒过膝半裙配棕色皮带、棕色皮质机车靴。每一格里都是同一个人——面容、发型、服饰、身材比例保持一致，只改变姿势、表情、动作与所处场景；参考图仅定义外观，不要照搬其站姿。把角色自然地融入各自的真实场景中，透视比例正确、光影贴合。\n\n【逐格场景与内容】\n第1格：清晨阳光洒入的现代简约公寓，女孩站在落地窗边，双手捧着马克杯，望向窗外的城市天际线，神情恬静。\n第2格：白天繁华的都市街道人行道，两旁是时尚店铺橱窗和梧桐树，女孩背着单肩包边走边微微回头，浅浅微笑。\n第3格：文艺清新的咖啡馆靠窗卡座，女孩坐着低头看手机，嘴角带笑，窗外光线柔和，桌上放着一杯拿铁。\n第4格：暖色调的独立书店书架走廊，女孩踮脚从书架上抽出一本书，侧脸专注，光线温柔。\n第5格：傍晚金色晚霞下的城市人行天桥，女孩倚着栏杆回眸，背景是虚化的车流与高楼，暖光洒在脸上。\n第6格：夜晚霓虹灯光的街角，女孩仰头微笑看向前方，暖黄路灯与散景光斑环绕，都市夜景氛围。",
+  "steps": 6,
+  "size_label": "16:9 · 2048×1152 (1536² area)",
+  "enhance": false
+ },
+ {
+  "prompt": "Photograph a translucent emerald perfume bottle on pale limestone beside a shallow pool. Rippling sunlight reflects through the glass onto the stone. A single olive branch frames the upper left corner. Luxury product photography, realistic refraction, crisp bottle edges, soft shadows, uncluttered composition, no logo or text.",
+  "refs": [
+   null,
+   null,
+   null
+  ],
+  "result": "qwen_perfume.png",
+  "info": "Pre-rendered example · seed `42` · 2496×1664 · 6 steps · v0.2.1 LoRA r256 · prompt sent as written (enhancement off)",
+  "used_prompt": "Photograph a translucent emerald perfume bottle on pale limestone beside a shallow pool. Rippling sunlight reflects through the glass onto the stone. A single olive branch frames the upper left corner. Luxury product photography, realistic refraction, crisp bottle edges, soft shadows, uncluttered composition, no logo or text.",
+  "steps": 6,
+  "size_label": "3:2 · 2496×1664 (2048² area)",
+  "enhance": false
+ },
+ {
+  "prompt": "图中画圈标注的地方需要补上一名骑坐姿态的西部牛仔男子。他头戴棕色宽檐牛仔帽，蓄着浓密的络腮胡，脸侧向画面左方；上身穿棕色帆布夹克，里面搭配深蓝色牛仔衬衫，颈间系着浅棕色围巾；下身穿着带流苏的棕色皮质护腿，脚穿皮靴踩进马镫，一只手搭在鞍部附近，整体呈现自然的骑乘状态。",
+  "refs": [
+   "qwen_mask_edit_in1.webp",
+   "qwen_mask_edit_in2.webp",
+   null
+  ],
+  "result": "qwen_mask_edit.png",
+  "info": "Pre-rendered example · seed `42` · 1248×1888 · 6 steps · v0.2.1 LoRA r256 · prompt sent as written (enhancement off)",
+  "used_prompt": "图中画圈标注的地方需要补上一名骑坐姿态的西部牛仔男子。他头戴棕色宽檐牛仔帽，蓄着浓密的络腮胡，脸侧向画面左方；上身穿棕色帆布夹克，里面搭配深蓝色牛仔衬衫，颈间系着浅棕色围巾；下身穿着带流苏的棕色皮质护腿，脚穿皮靴踩进马镫，一只手搭在鞍部附近，整体呈现自然的骑乘状态。",
+  "steps": 6,
+  "size_label": "2:3 · 1248×1888 (1536² area)",
+  "enhance": false
+ },
+ {
+  "prompt": "这是一张带有透明度的RGBA格式图像，呈现一位身着华丽婚纱的动漫风格女性角色。画面中无任何可识别的文字信息。角色为年轻女性，拥有白皙肌肤、红润脸颊与明亮红色眼眸，面带温柔微笑，表情甜美而幸福。她头戴精致银色皇冠与白色蕾丝头纱，金色长发自然垂落肩头，发丝柔顺且富有光泽。身穿多层荷叶边设计的粉白渐变抹胸婚纱，裙摆蓬松飘逸，层次丰富，材质轻盈如花瓣般展开；胸前手持一束由白色玫瑰与绿色叶片组成的捧花，花束饱满，绿叶点缀其间，增添清新感。双腿修长，穿着透明薄纱质感的过膝袜与饰有玫瑰图案的白色高跟鞋，姿态优雅站立，双手轻握捧花置于腹前。整体构图为全身正面视角，色彩柔和梦幻，以粉、白、绿为主色调，光影细腻，具有典型的日系二次元插画美学风格，适用于游戏立绘、视觉小说或数字艺术收藏等场景。 该图像具有alpha通道，背景是透明的。",
+  "refs": [
+   null,
+   null,
+   null
+  ],
+  "result": "qwen_rgba_bride.png",
+  "info": "Pre-rendered example · seed `42` · 1664×2496 · 6 steps · v0.2.1 LoRA r256 · prompt sent as written (enhancement off)",
+  "used_prompt": "这是一张带有透明度的RGBA格式图像，呈现一位身着华丽婚纱的动漫风格女性角色。画面中无任何可识别的文字信息。角色为年轻女性，拥有白皙肌肤、红润脸颊与明亮红色眼眸，面带温柔微笑，表情甜美而幸福。她头戴精致银色皇冠与白色蕾丝头纱，金色长发自然垂落肩头，发丝柔顺且富有光泽。身穿多层荷叶边设计的粉白渐变抹胸婚纱，裙摆蓬松飘逸，层次丰富，材质轻盈如花瓣般展开；胸前手持一束由白色玫瑰与绿色叶片组成的捧花，花束饱满，绿叶点缀其间，增添清新感。双腿修长，穿着透明薄纱质感的过膝袜与饰有玫瑰图案的白色高跟鞋，姿态优雅站立，双手轻握捧花置于腹前。整体构图为全身正面视角，色彩柔和梦幻，以粉、白、绿为主色调，光影细腻，具有典型的日系二次元插画美学风格，适用于游戏立绘、视觉小说或数字艺术收藏等场景。 该图像具有alpha通道，背景是透明的。",
+  "steps": 6,
+  "size_label": "2:3 · 1664×2496 (2048² area)",
+  "enhance": false
+ },
  {
   "prompt": "Soaking wet capybara taking shelter under a banana leaf in the rainy jungle, close up photo",
   "refs": [

examples/qwen_prompts.json

@@ -0,0 +1,7 @@
+{
+ "qwen_poster": "Design a refined travel poster for a fictional night train. Render the headline exactly as \"THE MIDNIGHT EXPRESS\" and the subtitle \"A journey under the stars\". A silver train curves through dark blue mountains beneath a crescent moon. Art Deco geometry, ivory and gold lettering, clear typographic hierarchy, generous margins, print-ready composition.",
+ "qwen_storyboard": "生成一张无缝的分镜插画，**现代真实偶像剧风格**（电影感写实摄影，柔和自然光与暖色调，浅景深虚化，韩系都市偶像剧质感，画面精致高级）。\n\n【版面（必须严格遵守）】\n- 恰好 6 个分镜格，**等宽**，排成**单独一横排（1 行 6 列）**；不是网格、不是两行，绝不要把任何一格再拆成上下两块。\n- 每格都是**竖长条矩形**（高明显大于宽，近似 9:16 竖条）。\n- 从左到右依次编号 1、2、3、4、5、6，每格左上角一个圆形数字徽标；**每个数字只出现一次，不重复、不跳号、不缺号。**\n- **每一格里都必须画出角色本人**在做该格的动作，角色占该格画面主体、完整清晰（半身或全身，约占该格高度 2/3 以上）；**严禁出现只有背景、没有人物的空格。**\n- 6 格之间画风、光影氛围、人物一致。输出为一张扁平整图，6 格并排合并，不要输出多张分离文件。\n\n【角色】以输入参考图（含面部特写与正/侧/背三视全身立绘）为角色设计基准：现代年轻女性，棕色波浪长发带空气刘海，妆容清透，穿粉色方领修身长袖针织上衣、棕色灯芯绒过膝半裙配棕色皮带、棕色皮质机车靴。每一格里都是同一个人——面容、发型、服饰、身材比例保持一致，只改变姿势、表情、动作与所处场景；参考图仅定义外观，不要照搬其站姿。把角色自然地融入各自的真实场景中，透视比例正确、光影贴合。\n\n【逐格场景与内容】\n第1格：清晨阳光洒入的现代简约公寓，女孩站在落地窗边，双手捧着马克杯，望向窗外的城市天际线，神情恬静。\n第2格：白天繁华的都市街道人行道，两旁是时尚店铺橱窗和梧桐树，女孩背着单肩包边走边微微回头，浅浅微笑。\n第3格：文艺清新的咖啡馆靠窗卡座，女孩坐着低头看手机，嘴角带笑，窗外光线柔和，桌上放着一杯拿铁。\n第4格：暖色调的独立书店书架走廊，女孩踮脚从书架上抽出一本书，侧脸专注，光线温柔。\n第5格：傍晚金色晚霞下的城市人行天桥，女孩倚着栏杆回眸，背景是虚化的车流与高楼，暖光洒在脸上。\n第6格：夜晚霓虹灯光的街角，女孩仰头微笑看向前方，暖黄路灯与散景光斑环绕，都市夜景氛围。",
+ "qwen_perfume": "Photograph a translucent emerald perfume bottle on pale limestone beside a shallow pool. Rippling sunlight reflects through the glass onto the stone. A single olive branch frames the upper left corner. Luxury product photography, realistic refraction, crisp bottle edges, soft shadows, uncluttered composition, no logo or text.",
+ "qwen_mask_edit": "图中画圈标注的地方需要补上一名骑坐姿态的西部牛仔男子。他头戴棕色宽檐牛仔帽，蓄着浓密的络腮胡，脸侧向画面左方；上身穿棕色帆布夹克，里面搭配深蓝色牛仔衬衫，颈间系着浅棕色围巾；下身穿着带流苏的棕色皮质护腿，脚穿皮靴踩进马镫，一只手搭在鞍部附近，整体呈现自然的骑乘状态。",
+ "qwen_rgba_bride": "这是一张带有透明度的RGBA格式图像，呈现一位身着华丽婚纱的动漫风格女性角色。画面中无任何可识别的文字信息。角色为年轻女性，拥有白皙肌肤、红润脸颊与明亮红色眼眸，面带温柔微笑，表情甜美而幸福。她头戴精致银色皇冠与白色蕾丝头纱，金色长发自然垂落肩头，发丝柔顺且富有光泽。身穿多层荷叶边设计的粉白渐变抹胸婚纱，裙摆蓬松飘逸，层次丰富，材质轻盈如花瓣般展开；胸前手持一束由白色玫瑰与绿色叶片组成的捧花，花束饱满，绿叶点缀其间，增添清新感。双腿修长，穿着透明薄纱质感的过膝袜与饰有玫瑰图案的白色高跟鞋，姿态优雅站立，双手轻握捧花置于腹前。整体构图为全身正面视角，色彩柔和梦幻，以粉、白、绿为主色调，光影细腻，具有典型的日系二次元插画美学风格，适用于游戏立绘、视觉小说或数字艺术收藏等场景。 该图像具有alpha通道，背景是透明的。"
+}