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

.gitattributes

@@ -42,3 +42,127 @@ examples/t2i_capybara.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i_cat_zh.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i_diorama.png filter=lfs diff=lfs merge=lfs -text
 examples/t2i_launch.png filter=lfs diff=lfs merge=lfs -text
+compare/img/case00/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case00/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case00/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case01/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case01/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case01/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case01/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case01/turbo8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/ref_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case02/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/base_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/ref_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case03/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case04/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case04/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case04/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case04/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case05/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case05/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case05/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case05/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/ref_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case06/turbo_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case07/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case07/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case07/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case07/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case08/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case08/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case08/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case08/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case09/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case09/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case09/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case09/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case10/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case10/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case10/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case10/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case11/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case11/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case11/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case11/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case12/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case12/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case12/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case12/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case13/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case13/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case13/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case14/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case14/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case14/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case15/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case15/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case15/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case15/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case16/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case16/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case16/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case16/turbo8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case17/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case17/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case17/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case17/turbo8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case18/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case18/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case18/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case18/turbo8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case19/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case19/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case19/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case19/turbo8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case19/turbo8_t.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in2.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in3.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in4.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in5.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/in6.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case20/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/in2.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/in4.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/in5.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case21/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in1.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in10.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in2.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in3.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in4.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in5.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in6.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in7.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in8.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/in9.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/ref.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case22/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case23/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case23/turbo.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case24/base.webp filter=lfs diff=lfs merge=lfs -text
+compare/img/case24/turbo.webp filter=lfs diff=lfs merge=lfs -text
+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

README.md

@@ -11,7 +11,7 @@ pinned: false
 license: other
 license_name: qwen-research
 license_link: ./LICENSE
-short_description: v0.2.1 preview - 6-step Qwen-Image-2.1, T2I + editing
+short_description: 6-step Qwen-Image-2.1, T2I + editing, vs-base comparison
 models:
   - Viggle/Qwen-Image-2.1-viggle-turbo
   - Qwen/Qwen-Image-2.1
@@ -28,10 +28,12 @@ tags:
 # `hf_oauth` is not needed: the app calls no user-scoped Hub API.
 ---
 
-# Viggle Turbo v0.2.1 (preview) — 6-step Qwen-Image-2.1
+# Viggle Turbo v0.2.1 — 6-step Qwen-Image-2.1
 
 A DMD-distilled student of **Qwen-Image-2.1** that generates and edits images in **6 sampling steps**
-with **no classifier-free guidance**, against the teacher's 40 steps.
+with **no classifier-free guidance**, against the teacher's 40 steps: about **5× faster** and very competitive
+with the base model in quality — on some prompts we prefer its output. The **Comparison** tab shows it side by side
+with the base model on the official Qwen examples.
 
 **v0.2 (2026-09-23):** much better sample diversity than v0.1 — intra-prompt diversity 0.93× the 40-step base
 model, up from 0.75× — and composition / prompt adherence much closer to the base model (see the model card
@@ -52,18 +54,28 @@ text-to-image, or fill one to three of them for editing / composition / style tr
 The **Examples** table below the form is pre-rendered: every row carries the output this exact
 app produced for it (fixed seed; 6 steps and prompt enhancement on unless the row's status line says
 otherwise — the launch-graphic row is 8 steps at the 3:2 2048² size with the 3k-character brief sent as
-written), so you can see what the model does without spending GPU time — click a row to load the
-prompt, the references, the result, and the steps / size / enhancement settings that produced it.
-`release/render_examples.py` regenerates the rows and `examples/manifest.json` whenever the
-weights change. The reference photos `woman1/woman2/cat/cat_window/bird.webp` and the
-three-reference prompt come from the
-[black-forest-labs/flux-klein-9b-kv](https://huggingface.co/spaces/black-forest-labs/flux-klein-9b-kv) Space.
+written, and the five `qwen_*` rows are sent as written at 2048² / 1536²), so you can see what the model
+does without spending GPU time — click a row to load the prompt, the references, the result, and the
+steps / size / enhancement settings that produced it. `release/render_examples.py` regenerates the rows
+and `examples/manifest.json` whenever the weights change. The reference photos
+`woman1/woman2/cat/cat_window/bird.webp` and the three-reference prompt come from the
+[black-forest-labs/flux-klein-9b-kv](https://huggingface.co/spaces/black-forest-labs/flux-klein-9b-kv) Space;
+the prompts and input images of the `qwen_*` rows come from the
+[Qwen/Qwen-Image-2.1](https://huggingface.co/spaces/Qwen/Qwen-Image-2.1) Space.
+
+The **Comparison** tab needs no GPU: 33 of the 37 examples of the Qwen/Qwen-Image-2.1 Space, pre-rendered with
+this model in 6 steps (and in 8 steps on the 5 dense-text examples) and with the base model in 40 steps — same
+prompt, input images and seed 42, prompt enhancement off, one sample each — shown in an image slider with the
+render times, plus Qwen's own API output where their Space ships one. Its images (`compare/`) are built from
+the renders by `release/build_compare_space.py` when `release/upload_space.sh` uploads the Space; the four
+examples left out are Chinese infographics / a subtitled storyboard whose short prompts leave the on-image text
+to prompt enhancement.
 
 **Built with Qwen.** Distilled from [Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1).
 
-> **Status: preview, work in progress.** v0.2.1 is a large step up from v0.1 but still falls short of the 40-step base
-> model on complicated image editing (multi-reference composition, face swaps, identity-preserving edits, instructions
-> with several constraints). We will keep updating this Space as the distillation improves.
+> **Known limits.** Complicated edits (multi-reference composition, face swaps, identity-preserving edits,
+> instructions with several constraints) can still fall short of the 40-step base model, and at 6 steps very
+> small dense text can print less cleanly (8 steps helps). We keep updating the weights as the distillation improves.
 
 ## What the app loads
 

app.py

@@ -28,6 +28,7 @@ from pathlib import Path
 
 import gradio as gr
 import torch
+import compare
 from PIL import Image
 from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline, QwenImage21Transformer2DModel
 from diffusers.pipelines.qwenimage21.pipeline_qwenimage21 import calculate_dimensions
@@ -218,67 +219,72 @@ def refresh_sizes(image_1, image_2, image_3, current):
 
 with gr.Blocks(title="Viggle Turbo v0.2.1 · Qwen-Image-2.1 6-step") as demo:
     gr.Markdown(
-        "# Viggle Turbo v0.2.1 (preview) — 6-step Qwen-Image-2.1\n"
-        "A DMD-distilled student of **Qwen-Image-2.1** that generates and edits in **6 sampling steps**, "
-        "with no classifier-free guidance. Leave the reference images empty for text-to-image; "
-        "add one to three of them to edit, compose or transfer style. The size menu switches to the "
-        "editing buckets as soon as a reference is attached.\n\n"
-        "**v0.2:** much better sample diversity than v0.1 (intra-prompt diversity 0.97× the 40-step base model, up from 0.75×) "
-        "and closer prompt adherence / composition to the base model. **2026-09-24:** same weights, now sampled on 6 steps "
-        "instead of 5 - the extra step splits the highest-noise segment again, which removes the last of the composition drift "
-        "(0% of prompts vs 4% at 5 steps) and lifts diversity from 0.93× to 0.97×. **v0.2.1 (2026-09-24):** the step-700 "
-        "checkpoint of the same run (v0.2 was step 600) - a little sharper, diversity 0.98×, still 0% drift. Still a preview: complicated edits "
-        "(multi-reference composition, face swaps, identity-preserving edits) remain weaker than the 40-step base model.\n\n"
-        f"Weights: **{STUDENT_TAG}** from "
-        f"[{STUDENT_REPO}](https://huggingface.co/{STUDENT_REPO})."
+        "# Viggle Turbo v0.2.1 — 6-step Qwen-Image-2.1\n"
+        "A DMD-distilled student of **Qwen-Image-2.1** that generates and edits in **6 sampling steps**, with no "
+        "classifier-free guidance: about **5× faster** than the 40-step base model and very competitive with it in quality, "
+        "and on some prompts we prefer its output. The **Comparison** tab has 33 examples of the official Qwen Space side by "
+        "side, turbo vs base, same seed. Complicated edits (multi-reference composition, face swaps, identity-preserving "
+        "edits) can still fall short of the base model.\n\n"
+        "Leave the reference images empty for text-to-image; add one to three of them to edit, compose or transfer style. "
+        "The size menu switches to the editing buckets as soon as a reference is attached. "
+        "**v0.2.1 (2026-09-24):** the step-700 checkpoint of the v0.2 run on the 6-step schedule — intra-prompt diversity "
+        "0.98× the base model, 0% composition drift; earlier versions and the numbers are on the model card. "
+        f"Weights: **{STUDENT_TAG}** from [{STUDENT_REPO}](https://huggingface.co/{STUDENT_REPO})."
     )
-    with gr.Row():
-        with gr.Column(scale=3):
-            prompt = gr.Textbox(label="Prompt", lines=3, placeholder="Describe the image, or the edit to apply to the references.")
-            with gr.Row():
-                image_1 = gr.Image(label="Reference 1 (optional)", type="pil", image_mode="RGB", height=200)
-                image_2 = gr.Image(label="Reference 2 (optional)", type="pil", image_mode="RGB", height=200)
-                image_3 = gr.Image(label="Reference 3 (optional)", type="pil", image_mode="RGB", height=200)
-            with gr.Row():
-                # allow_custom_value: gradio validates API calls against the *initial* choices (the text-to-image
-                # menu), which would reject every editing size sent through /generate; generate() snaps
-                # anything off-menu itself.
-                size_label = gr.Dropdown(label="Output size", choices=T2I_CHOICES, value=T2I_CHOICES[0], allow_custom_value=True)
-                steps = gr.Slider(label="Steps", minimum=3, maximum=8, step=1, value=STEPS,
-                                  info="6 is the validated schedule. Extra steps are always added at the high-noise end (7 also validated, 8 untested); fewer steps drop them again (4 = training nodes, 3 = uniform spacing, off-contract).")
-            with gr.Row():
-                seed = gr.Number(label="Seed", value=0, precision=0, interactive=False)
-                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-            enhance = gr.Checkbox(
-                label="Enhance prompt",
-                value=True,
-                info="Rewrite the prompt with the official Qwen-Image prompt-enhancement instructions "
-                "(runs on the built-in Qwen3-VL text encoder, +4–15 s). The text actually sent is shown under the result.",
+    with gr.Tab("Generate"):
+        with gr.Row():
+            with gr.Column(scale=3):
+                prompt = gr.Textbox(label="Prompt", lines=3, placeholder="Describe the image, or the edit to apply to the references.")
+                with gr.Row():
+                    image_1 = gr.Image(label="Reference 1 (optional)", type="pil", image_mode="RGB", height=200)
+                    image_2 = gr.Image(label="Reference 2 (optional)", type="pil", image_mode="RGB", height=200)
+                    image_3 = gr.Image(label="Reference 3 (optional)", type="pil", image_mode="RGB", height=200)
+                with gr.Row():
+                    # allow_custom_value: gradio validates API calls against the *initial* choices (the text-to-image
+                    # menu), which would reject every editing size sent through /generate; generate() snaps
+                    # anything off-menu itself.
+                    size_label = gr.Dropdown(label="Output size", choices=T2I_CHOICES, value=T2I_CHOICES[0], allow_custom_value=True)
+                    steps = gr.Slider(label="Steps", minimum=3, maximum=8, step=1, value=STEPS,
+                                      info="6 is the validated schedule. Extra steps are always added at the high-noise end (7 also validated, 8 untested); fewer steps drop them again (4 = training nodes, 3 = uniform spacing, off-contract).")
+                with gr.Row():
+                    seed = gr.Number(label="Seed", value=0, precision=0, interactive=False)
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                enhance = gr.Checkbox(
+                    label="Enhance prompt",
+                    value=True,
+                    info="Rewrite the prompt with the official Qwen-Image prompt-enhancement instructions "
+                    "(runs on the built-in Qwen3-VL text encoder, +4–15 s). The text actually sent is shown under the result.",
+                )
+                run = gr.Button("Generate", variant="primary")
+            with gr.Column(scale=4):
+                output_image = gr.Image(label="Result", type="pil", height=560)
+                info = gr.Markdown()
+                used_prompt = gr.Textbox(label="Prompt sent to the model", lines=4, interactive=False, show_copy_button=True)
+
+        def example_details(prompt, image_1, image_2, image_3, result):
+            """Runs on an example click (no GPU): fills the status line and the prompt stored with the row, and sets the
+            steps / size / enhancement controls to what produced it (rows differ; Run then reproduces the result). It also sets
+            the size menu's mode itself: loading an example does not fire the reference images' .input listeners."""
+            row = next(row for row in EXAMPLES if row["prompt"] == prompt)
+            choices = EDIT_CHOICES if any(row["refs"]) else T2I_CHOICES
+            return row["info"], row["used_prompt"], row["steps"], gr.update(choices=choices, value=row["size_label"]), row["enhance"]
+
+        if EXAMPLES:
+            gr.Examples(
+                examples=[
+                    [row["prompt"], *[str(EXAMPLES_DIR / ref) if ref else None for ref in row["refs"]], str(EXAMPLES_DIR / row["result"])]
+                    for row in EXAMPLES
+                ],
+                inputs=[prompt, image_1, image_2, image_3, output_image],
+                fn=example_details,
+                outputs=[info, used_prompt, steps, size_label, enhance],
+                run_on_click=True,
+                examples_per_page=12,
+                label="Examples · results pre-rendered by this model — click a row to load it (6 steps, prompt enhancement on, unless the status line says otherwise)",
             )
-            run = gr.Button("Generate", variant="primary")
-        with gr.Column(scale=4):
-            output_image = gr.Image(label="Result", type="pil", height=560)
-            info = gr.Markdown()
-            used_prompt = gr.Textbox(label="Prompt sent to the model", lines=4, interactive=False, show_copy_button=True)
-
-    def example_details(prompt, image_1, image_2, image_3, result):
-        """Runs on an example click (no GPU): fills the status line and the prompt stored with the row, and sets the
-        steps / size / enhancement controls to what produced it (rows differ; Run then reproduces the result)."""
-        row = next(row for row in EXAMPLES if row["prompt"] == prompt)
-        return row["info"], row["used_prompt"], row["steps"], row["size_label"], row["enhance"]
-
-    if EXAMPLES:
-        gr.Examples(
-            examples=[
-                [row["prompt"], *[str(EXAMPLES_DIR / ref) if ref else None for ref in row["refs"]], str(EXAMPLES_DIR / row["result"])]
-                for row in EXAMPLES
-            ],
-            inputs=[prompt, image_1, image_2, image_3, output_image],
-            fn=example_details,
-            outputs=[info, used_prompt, steps, size_label, enhance],
-            run_on_click=True,
-            label="Examples · results pre-rendered by this model — click a row to load it (6 steps, prompt enhancement on, unless the status line says otherwise)",
-        )
+
+    with gr.Tab("Comparison"):
+        compare.render()
 
     gr.Markdown(
         "---\n"
@@ -289,8 +295,10 @@ with gr.Blocks(title="Viggle Turbo v0.2.1 · Qwen-Image-2.1 6-step") as demo:
         "(**non-commercial: research or evaluation purposes only**). This demo inherits that restriction."
     )
 
+    # .input (user upload / clear) and not .change: .change also fires when an example loads its references, and that
+    # refresh then raced example_details and reset the size to the first menu entry
     for image_input in (image_1, image_2, image_3):
-        image_input.change(
+        image_input.input(
             refresh_sizes,
             inputs=[image_1, image_2, image_3, size_label],
             outputs=size_label,

compare/cases.json

@@ -0,0 +1 @@
+[{"id": "case00", "title": "Transparent anime bride", "title_zh": "透明婚纱角色立绘", "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。", "size": [1664, 2496], "inputs": [], "reference": {"src": "img/case00/ref.webp", "thumb": "img/case00/ref_t.webp", "w": 1664, "h": 2496, "alpha": true}, "turbo": {"src": "img/case00/turbo.webp", "thumb": "img/case00/turbo_t.webp", "w": 1664, "h": 2496, "alpha": true}, "turbo_s": 4.7, "base": {"src": "img/case00/base.webp", "thumb": "img/case00/base_t.webp", "w": 1664, "h": 2496, "alpha": true}, "base_s": 26.0}, {"id": "case01", "title": "Character style infographic", "title_zh": "信息图", "prompt": "将输入图片转化为日系杂志风的「人物风格信息图」海报，**严格保持输入人物的面容、五官与气质不变**。版面采用月白底与浅灰线网格，整体清冷、高级、通透。\n\n【版面与文字位置】\n- 顶部留出一条**独立的横向标题栏**（贯穿画面顶部的空白留白带），大标题“月白纱语”置于此标题栏内左中位置，右侧副标题“VOL.05 · 清冷艺术”，最右端椭圆徽标内“写真特辑”。\n- 中央人物立绘**整体适当下移并居中**，人物（含头顶发髻）**不得进入顶部标题栏、不得遮挡或压住任何文字**；人物头顶与标题之间保留明显空白。\n- 各信息卡片沿左右两栏与下方排布，文字均在各自卡片内，不被人物遮挡。\n- 排版精致、留白充足、有呼吸感。\n\n【字体】所有文字采用清晰、精致、美观的字体；大标题使用优雅的中文衬线/艺术字体（笔画利落、有设计感、精致高级），正文用清秀易读的黑体；字形端正、无错字无乱码。\n\n【中央人物】身着白色欧根纱一字肩荷叶边长裙、腰间黑色缎带，高扎发，气质清冷优雅；左侧标注“168cm”、“身姿修长 — 天鹅颈优越”。\n\n左上“个人档案”，逐行：“风格定位 清冷高级感艺术女性”、“适合色系 月白/雾灰/裸粉”、“避免色系 荧光色/高饱和撞色”、“身材优势 天鹅颈/肩颈线/纤细四肢”、“穿搭关键词 纱质/透视/垂坠/解构”。\n左中“配饰细节”，四项：“珍珠耳饰 优雅点睛”、“黑色缎带 束腰点缀”、“细链项链 锁骨修饰”、“银质戒指 精致细节”。\n左下“人物时间线”四段：“晨间护肤”、“精致妆发”、“棚拍写真”、“艺术观展”。其下“搭配建议”：“方案一 清冷极简 #F2ECE9”、“方案二 艺术编辑 #C9BBB2”。\n右上“属性卡片”三头像：“凝视 | 冷白皮妆 | 高扎发 时尚大片”、“侧颜 | 雾面裸妆 | 湿发感 艺术写真”、“慵懒 | 水光裸妆 | 低盘发 杂志封面”。\n右中“服饰细节”四项：“欧根纱质”、“一字露肩”、“荷叶边饰”、“波点透纱”。\n右中“服饰色卡”标题“月白冷调”，6色块依次：“#FFFFFF 月白”、“#F2ECE9 珍珠白”、“#E5DED9 雾灰白”、“#D8CEC7 浅灰驼”、“#BFB2A8 裸棕”、“#2E2C2D 缎带黑”。\n右下“性格标签”雷达五维：“高级感 92%”、“清冷感 88%”、“艺术感 90%”、“氛围感 85%”、“松弛感 80%”。\n右下“场景推荐”四条：“美术馆 · 静谧午后”、“摄影棚 · 时尚大片”、“高级料理 · 微醺夜”、“艺术展 · 光影漫步”。\n底部“风格关键词”词云，仅含：“高级感”、“清冷风”、“氛围感”、“纱感”、“艺术感”、“杂志风”、“松弛感”、“仙气”、“编辑风”。\n\n【文字要求】画面中所有文字必须是以上明确指定的中文，逐字准确、清晰可读、无错字无乱码；不要生成任何未指定的多余文字。", "size": [2048, 1152], "inputs": [{"src": "img/case01/in1.webp", "thumb": "img/case01/in1_t.webp", "w": 1344, "h": 913, "alpha": false}], "reference": {"src": "img/case01/ref.webp", "thumb": "img/case01/ref_t.webp", "w": 2720, "h": 1536, "alpha": false}, "turbo": {"src": "img/case01/turbo.webp", "thumb": "img/case01/turbo_t.webp", "w": 2048, "h": 1152, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case01/base.webp", "thumb": "img/case01/base_t.webp", "w": 2048, "h": 1152, "alpha": false}, "base_s": 14.3, "turbo8": {"src": "img/case01/turbo8.webp", "thumb": "img/case01/turbo8_t.webp", "w": 2048, "h": 1152, "alpha": false}, "turbo8_s": 3.8}, {"id": "case02", "title": "Six-panel storyboard", "title_zh": "分镜图", "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。", "size": [2048, 1152], "inputs": [{"src": "img/case02/in1.webp", "thumb": "img/case02/in1_t.webp", "w": 1920, "h": 1080, "alpha": false}], "reference": {"src": "img/case02/ref.webp", "thumb": "img/case02/ref_t.webp", "w": 2048, "h": 1152, "alpha": false}, "turbo": {"src": "img/case02/turbo.webp", "thumb": "img/case02/turbo_t.webp", "w": 2048, "h": 1152, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case02/base.webp", "thumb": "img/case02/base_t.webp", "w": 2048, "h": 1152, "alpha": false}, "base_s": 14.0}, {"id": "case03", "title": "360-degree panorama", "title_zh": "全景图", "prompt": "Generate a complete 360-degree equirectangular panorama from the input perspective image.\n\nThe scene is an open paved urban plaza beside a very tall, slender observation tower with a twisting white lattice structure and a narrow antenna at the top, rising into a partly cloudy blue sky. A woman with long straight dark brown hair is close to the camera, wearing a brown baseball cap with a light-colored embroidered emblem, a black short-sleeved T-shirt, and a dark green shoulder bag. She has a gentle closed-mouth smile and one arm extended toward the camera. Preserve her facial appearance, expression, cap, clothing, bag, pose, and relationship to the tower. The plaza has broad gray stone paving, trimmed hedges, mature green trees, a red-framed low structure, slim light poles, a few distant pedestrians, and partial urban buildings beyond the landscaping. Extend the scene into a coherent full 360-degree plaza with continuous paving, green planted borders, nearby paths and city surroundings behind the camera. Keep the tower present only once, retain its recognizable lattice geometry and its position relative to the woman, and include its full upward extent through the correct spherical projection. Preserve the reference image's natural photographic style, soft bright daylight, cloud patterns, believable skin tones, and consistent shadows.\n\nUse a true equirectangular projection covering the entire 360-degree horizontal and 180-degree vertical field of view, including the sky overhead and the ground below. Continue the environment behind the camera consistently, with seamless left and right edges and a single continuous scene. Keep the woman present only once. Output resolution: 2880 x 1440 pixels, with an exact 2:1 aspect ratio.", "size": [2176, 1088], "inputs": [{"src": "img/case03/in1.webp", "thumb": "img/case03/in1_t.webp", "w": 1080, "h": 1440, "alpha": false}], "reference": {"src": "img/case03/ref.webp", "thumb": "img/case03/ref_t.webp", "w": 2880, "h": 1440, "alpha": false}, "turbo": {"src": "img/case03/turbo.webp", "thumb": "img/case03/turbo_t.webp", "w": 2176, "h": 1088, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case03/base.webp", "thumb": "img/case03/base_t.webp", "w": 2176, "h": 1088, "alpha": false}, "base_s": 14.3}, {"id": "case04", "title": "Foundation product scene", "title_zh": "粉底液商品场景", "prompt": "一张真实日常的手机随手拍摄生活照，如同在家中窗边餐桌随手记录的画面。保持<image1>中粉底液瓶的全部外观完全不变，包括磨砂半透明的哑光白色方形瓶身及其圆润边角、金色金属按压泵头、瓶身正面竖排金色手写体\"INS BAHA\"标识、黑色无衬线三行文字\"NO WEAR LASTING FOUNDATION\"、文字下方的短横线、以及深蓝色中文\"印彩巴哈\"与\"透薄持妆粉底液\"字样，磨砂玻璃的细腻雾面纹理与泵头的金属光泽均与<image1>完全一致。移除原图中握持瓶身的手部与深色杂乱背景，画面中不出现任何人物。将瓶身竖直立在浅原木色木纹餐桌中部偏右位置，正面标签正对镜头，瓶底紧贴桌面并投下向右侧延伸的柔和接触阴影。采用日常随手记录的轻微俯拍视角，瓶身左侧紧邻一只白色哑光陶瓷茶杯，杯中盛着半杯浅琥珀色茶水，杯壁映出窗光高光。瓶身右侧一个白色小圆瓷盘盛着三颗带绿蒂的鲜红草莓。瓶身右后方一个透明玻璃小花瓶插着两三支带嫩绿叶片的尤加利枝，枝叶向画面右上方舒展。画面左下角露出一本合上的浅灰色布面笔记本的边角，被画框裁切只显局部。桌面上斜铺一条米白色棉麻桌旗，边缘带着随意褶皱，从瓶身下方延伸至画面右下。瓶身左后方一把浅木色餐椅的椅背局部入镜。画面左侧一扇白色木框窗户挂着半透明白色纱帘，窗外日光透过纱帘柔和洒入，窗台上一盆白色陶瓷小盆栽种着翡翠色多肉植物。瓶身正后方远处一面浅米色墙面前立着一个原木色置物架，架上几件白瓷与陶土器皿轮廓柔和。透过窗户可见完全虚化的室外绿色树影与一角浅蓝天空。近处木纹桌面的细密纹理与瓶身保持清晰锐利，中景椅背与窗框略微柔和，远处置物架和窗外树影呈现舒适散焦。整体光线为明亮的窗外自然日光从左侧照入，磨砂瓶身透出柔和半透明光感，金色泵头泛起温暖高光，色彩真实不夸张，构图带着不经意的随意感，呈现出真实可信的手机随手拍摄日常质感。", "size": [1312, 1792], "inputs": [{"src": "img/case04/in1.webp", "thumb": "img/case04/in1_t.webp", "w": 1248, "h": 1664, "alpha": false}], "reference": {"src": "img/case04/ref.webp", "thumb": "img/case04/ref_t.webp", "w": 1760, "h": 2368, "alpha": false}, "turbo": {"src": "img/case04/turbo.webp", "thumb": "img/case04/turbo_t.webp", "w": 1312, "h": 1792, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case04/base.webp", "thumb": "img/case04/base_t.webp", "w": 1312, "h": 1792, "alpha": false}, "base_s": 14.2}, {"id": "case05", "title": "Bracelet styling", "title_zh": "手链佩戴", "prompt": "在完整保留输入图手链全部细节的前提下，为这条手链生成一张模特佩戴场景图：手链以黑色多孔火山石圆珠为主体珠串，正面居中为金色双盘形连接饰件，两侧对称排列两颗带猫眼光泽的棕黄色虎眼石圆珠、两枚金色南瓜纹隔珠与两颗金边包镶的淡紫色紫水晶圆珠，其余珠位由黑色火山石圆珠与金色小隔珠串联，珠序、珠体光泽、包镶结构与金属件造型均与输入图完全一致。佩戴手链的是一位年轻中国女性模特，她身穿米白色棉麻长袖衬衫，袖口挽至小臂中部，黑色长发松松挽在脑后，神态安静松弛，坐在一间中式茶室的实木长桌旁，目光低垂落在自己搭在桌面的手腕上。她佩戴手链的左手手背朝上、手指自然松弛地搭在铺有浅米色亚麻桌旗的桌面上，手腕微微抬起使手链正面完整朝向镜头并处于画面清晰焦点，右手则自然垂放在桌沿之下、位于画面之外；桌面右前方摆着一只素白瓷茶杯与一把青瓷小茶壶，背景为虚化的木质格栅窗和窗外朦胧的绿植，营造出安静雅致的中式茶室氛围。光线采用午后从画面左侧窗格斜入的暖白色自然侧光，柔和而具有方向性，在火山石珠的哑光表面刻画出细密孔隙的微小阴影，在紫水晶与虎眼石珠面点亮通透的高光，令金色隔珠与连接饰��泛出温暖金属光泽，同时模特的手背、小臂、衣袖与桌面亚麻纹理也在同一侧光下呈现一致的明暗过渡与阴影方向，背景木色与绿植统一在暖调氛围中。构图为近景特写结合中景人物，焦点锁定手腕手链，模特面部与上半身居于画面右后方并带轻柔的景深虚化，整体色彩以暖木色、米白色衬托黑色珠串与金色配饰，画面清晰细腻，呈现真实自然的高品质产品佩戴场景摄影质感。", "size": [1536, 1536], "inputs": [{"src": "img/case05/in1.webp", "thumb": "img/case05/in1_t.webp", "w": 2048, "h": 2048, "alpha": false}], "reference": {"src": "img/case05/ref.webp", "thumb": "img/case05/ref_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo": {"src": "img/case05/turbo.webp", "thumb": "img/case05/turbo_t.webp", "w": 1536, "h": 1536, "alpha": false}, "turbo_s": 3.1, "base": {"src": "img/case05/base.webp", "thumb": "img/case05/base_t.webp", "w": 1536, "h": 1536, "alpha": false}, "base_s": 14.2}, {"id": "case06", "title": "Holiday cushion scene", "title_zh": "抱枕节日布置", "prompt": "完全保留<image1>中的抱枕原貌，包括其方形造型、深海军蓝黑色背景、中央金色生命之树图案（带有盘曲的树干和伸展的根系）、蓝金相间的树叶和花朵、栖息的小鸟和蝴蝶，以及环绕边缘的金色藤蔓和散落的金色圆点。将抱枕布置成包装好的节日礼物，放置在舒适高档的美式圣诞客厅中，将其竖立并略微倾斜朝向镜头，放在深翠绿色天鹅绒翼背扶手椅上，使完整的生命之树图案面向观众，抱枕周围系着光泽红色缎带，蝴蝶结上悬挂着牛皮纸礼品标签，上面用优雅的黑色手写体写着\"Merry Christmas\"，抱枕沉入天鹅绒中带有柔和的接触阴影。在画面左下角的前景中，一个用深红色包装纸和金丝带包裹的礼物盒的一角探入画面，旁边散落着几个棕色松果和一枝带有鲜红浆果的光泽冬青 resting on the floor。扶手椅右侧搭着一条粗针脚奶油色流苏针织毯，旁边的圆形胡桃木边桌上放着白色陶瓷杯的热可可，上面铺着烤棉花糖，黄铜烛台里的琥珀色玻璃蜡烛散发着温暖的光圈，还有一枝新鲜桉树叶。椅子后方左侧矗立着一棵高大的圣诞树，缠绕着暖白色小灯，装饰着红色和金色玻璃球，顶部闪耀着金色星星，树下整齐堆放着用花纹纸包装的礼物，放在铺在蜂蜜色硬木地板上的天然黄麻编织地毯上。右侧是质朴的灰色石砌壁炉，炉火噼啪作响，木制壁炉架上装饰着茂密的常绿花环，点缀着松果和红色浆果，黄铜挂钩上挂着三件粗针织袜子（奶油色、红色和森林绿），壁炉架上还立着一个小型黑色黑板标志，上面用白色字体写着\"Happy Holidays\"。壁炉架上方，华丽的金框镜子柔和地反射着闪烁的树灯。远处的墙壁上，高大的白色窗框窗户展现出轻柔飘落的雪花和覆盖着雪的松树，背景是温暖的奶油色墙壁。整体光线将壁炉和树灯的金色光芒与雪窗外凉爽柔和的日光融合，营造出温暖、节日、高端的节日氛围。", "size": [1536, 1536], "inputs": [{"src": "img/case06/in1.webp", "thumb": "img/case06/in1_t.webp", "w": 2048, "h": 2048, "alpha": false}], "reference": {"src": "img/case06/ref.webp", "thumb": "img/case06/ref_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo": {"src": "img/case06/turbo.webp", "thumb": "img/case06/turbo_t.webp", "w": 1536, "h": 1536, "alpha": false}, "turbo_s": 3.1, "base": {"src": "img/case06/base.webp", "thumb": "img/case06/base_t.webp", "w": 1536, "h": 1536, "alpha": false}, "base_s": 14.0}, {"id": "case07", "title": "Necklace and bracelet styling", "title_zh": "项链与手链展示", "prompt": "保持<image1>中14K镀金蛇骨链项链和配套手链的全部外观完全不变，包括金色扁平蛇骨链节紧密排列的鳞片状纹理、光滑镜面反光表面、龙虾扣搭扣上刻有的\"14K\"字样、项链与手链的链宽和长度比例。将这套首饰佩戴在一位欧美年轻女性模特身上，模特约22至25岁，浅棕色自然微卷长发披散至锁骨下方，小麦色健康肤色，五官立体轮廓分明，淡妆自然清透带裸色唇彩，身穿一件简约白色圆领棉质T恤，领口微低恰好展示项链垂落在锁骨之间的弧形线条。项链贴合模特颈部自然下垂至胸前正中，金色蛇骨链面在光线下呈现流畅的液态金属光泽，链身随呼吸微微起伏。手链佩戴在模特右手腕上，右手自然抬起轻触耳侧发丝，手腕处手链紧贴皮肤呈椭圆形环绕，蛇骨链面反射暖色高光。模特微微侧头看向镜头右方约15度，嘴角带自然微笑，眼神柔和自信。场景为洛杉矶风格的城市天台露台，模特身后一面浅灰色清水混凝土矮墙横贯画面中部，墙面有细微气孔和浇筑纹理。矮墙顶面左侧摆放两盆赤陶色粗陶花盆，一盆种着灰绿色莲座状石���花多肉，另一盆种着柱状仙人掌带细密白刺。矮墙顶面右侧一个透明玻璃水瓶插着三枝干燥蒲苇穗，穗头蓬松呈奶油白色。矮墙后方可见城市天际线中三四栋米白色和浅棕色中高层公寓楼，楼体表面规则排列深色铝合金窗框。画面左侧一棵修剪整齐的橄榄树，灰褐色扭曲树干带纵向裂纹，银绿色叶片在逆光中呈半透明质感。橄榄树旁一张浅灰色金属框架户外休闲椅，椅面搭一条米白色亚麻薄毯带流苏边。椅子旁一张圆形浅木色小边桌，桌面上一杯冰美式咖啡装在透明玻璃杯中杯壁凝结细密水珠，旁搁一部玫瑰金色手机屏幕朝下。画面右侧露台边缘黑色金属栏杆上悬挂一组暖白色圆球串灯，灯泡散发柔和暖光在混凝土墙面上投射圆形光斑。栏杆外远处可见两三棵棕榈树树冠剪影在暮色中摇曳。露台地面为浅灰色水泥自流平，表面有细微发丝裂纹和浅色使用痕迹。模特脚边地面一双白色帆布低帮鞋局部露出鞋头。画面左上方天空呈现日落时分渐变色彩，从地平线暖橙金色过渡到上方淡紫蓝色，几缕薄云被染成粉橙色和淡金色。模特左肩后方混凝土墙面上挂一幅小型黑色细金属框抽象线条装饰画，画面为白底单线连续人脸轮廓。装饰画下方一个浅橡木色悬浮搁板上摆放一瓶琥珀色玻璃香薰蜡烛带牛皮纸标签和一小束干燥尤加利圆叶枝。搁板右端一个白色陶瓷小花瓶插着一枝干燥棉花枝。整体光线为日落前黄金时段的暖色自然光从模特右后方约45度角照射，在模特面部左侧和颈部形成柔和的伦勃朗式三角光影，金色蛇骨链表面反射出温暖的橙金色连续高光带，白色T恤面料在暖光下呈现奶油色调，混凝土墙面被夕阳染上一层淡暖橙色，营造出时尚自然且富有生活气息的TikTok美区珠宝穿搭内容氛围。", "size": [1536, 1536], "inputs": [{"src": "img/case07/in1.webp", "thumb": "img/case07/in1_t.webp", "w": 2048, "h": 2048, "alpha": false}], "reference": {"src": "img/case07/ref.webp", "thumb": "img/case07/ref_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo": {"src": "img/case07/turbo.webp", "thumb": "img/case07/turbo_t.webp", "w": 1536, "h": 1536, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case07/base.webp", "thumb": "img/case07/base_t.webp", "w": 1536, "h": 1536, "alpha": false}, "base_s": 14.6}, {"id": "case08", "title": "Traditional clothing portrait", "title_zh": "传统服饰人像", "prompt": "将输入照片中展示在无头裁缝人台上的传统俾路支女性服饰，转化为一幅高端广告人像，其中人台被一位身着同款服饰的真实俾路支年轻女性所取代，服装的每一个细节都与输入照片完全一致：奶油色底布，V型领口饰有窄幅几何刺绣，紧身胸衣部分布满密集的垂直手工刺绣柱，图案为金黄色、深栗色、海军蓝和白色的菱形与八角星纹样，从腰部延伸至下摆的大矩形前襟饰满同样的多彩几何刺绣，并以图案化镶边带为框，长袖采用水平条纹状的同款刺绣拼接，袖口布满完整图案，裙身两侧及下摆则装饰着由红色、海军蓝、金黄色和白色钩编玫瑰花贴花拼缀而成的图案，并在奶油色侧片上点缀红色刺绣贴片。这位俾路支年轻女性以优雅放松的姿态为广告摆出造型，身体四分之三侧向镜头，面部转向镜头，带着温柔自信的微笑；她留着光泽柔顺的深色长发，呈柔和波浪状披散在肩头与背部，肤色干净自然，眉形清晰，唇色为明艳的玫瑰色，身上仅佩戴一对小金耳环和一只细金手镯作为首饰。她的一只手臂自然放松地垂在身侧，另一只手在臀部高度轻轻提起裙摆边缘，使袖子和前襟的刺绣正对镜头，重心自然落在一条腿上，双脚穿着简约中性的平底凉鞋稳稳踏在地板上。场景是一个专为服装广告打造的干净专业摄影棚：无缝背景从中心的暖象牙色渐变至边缘的柔和陶土色，带有微妙的灰泥质感，浅暖灰色地面在脚部和裙摆下方承接柔和的接触阴影。一盏主导的柔和主光从左前上方打下，呈暖中性色调，将她面部、头发及刺绣包裹在均匀、迷人的光线中，右侧留有开阔柔和的阴影，使绣线色彩显得准确而鲜艳。构图采用中长焦镜头在视平线拍摄的全身竖版人像，头部保持自然成人比例，即站姿高度的七分之一，使从领口到下摆的整件服装在平衡居中的构图中清晰对焦，人物周围留有从容的呼吸空间，呈现出一幅精致、逼真的时尚广告照片。", "size": [1248, 1888], "inputs": [{"src": "img/case08/in1.webp", "thumb": "img/case08/in1_t.webp", "w": 1664, "h": 2528, "alpha": false}], "reference": {"src": "img/case08/ref.webp", "thumb": "img/case08/ref_t.webp", "w": 1664, "h": 2528, "alpha": false}, "turbo": {"src": "img/case08/turbo.webp", "thumb": "img/case08/turbo_t.webp", "w": 1248, "h": 1888, "alpha": false}, "turbo_s": 2.9, "base": {"src": "img/case08/base.webp", "thumb": "img/case08/base_t.webp", "w": 1248, "h": 1888, "alpha": false}, "base_s": 14.2}, {"id": "case09", "title": "Running shoe campaign", "title_zh": "跑鞋广告", "prompt": "一张为专业跑步品牌拍摄的高品质运动鞋广告编辑摄影作品。保持<image1>中的跑鞋完全不变——其白色工程网眼鞋面、侧面三条醒目的黑色斜条纹、带有小号黑色文字\"LIGHTSTRIKE PRO\"的厚实雕塑感白色中底、白色鞋带、鞋舌和鞋领——并将其作为一双配对鞋穿在一位伊朗男性模特脚上，同时完全移除<image1>中握住鞋子的手和纯灰色摄影棚背景，新画面中不出现任何手部和摄影棚元素。相机贴近地面低位拍摄，捕捉模特从膝盖以下在沥青公园跑道上自信迈步的姿态，模特穿着黑色锥形跑步短裤和白色短袜：前脚平踩地面，使鞋面条纹侧面朝向镜头，后脚拖后抬起脚跟以展现鞋子的雕塑感轮廓，两只鞋底紧压路面，在左下方投射出紧密的深色接触阴影。最靠近镜头处，纹理丰富的沥青路面呈现出细腻颗粒感、发丝般的裂纹，以及一条褪色的白色车道线沿对角线横穿画面下部，两片干燥的法国梧桐树叶静静躺在上面。沿着跑道左侧边缘，一道低矮风化的混凝土路缘石将沥青路面与一条露水湿润的绿色草地分隔开来，几根草叶在模特后脚下弯曲。在左侧中景处，一张带有黑色金属腿的木质公园长椅空置在一棵纤细的法国梧桐树干下方，右侧一根黑色铸铁路灯柱延伸至画面上方之外，其底部环绕着低矮的绿色灌木。更远处，第二排树干沿着弯曲的小路向远处延伸，树冠融合成柔和模糊的绿色植被，而在公园边缘之外，淡淡的城市天际线在苍白的天空下融入温暖的金色晨雾中。低角度的晨光从右侧斜照，在两只鞋子的白色网眼和中底上勾勒出清晰的高光，将沥青路面染成蜜糖色调，并在跑道上投下迈步双腿的一道修长柔和阴影，营造出充满活力却又宁静祥和的晨跑氛围。", "size": [1312, 1792], "inputs": [{"src": "img/case09/in1.webp", "thumb": "img/case09/in1_t.webp", "w": 1760, "h": 2368, "alpha": false}], "reference": {"src": "img/case09/ref.webp", "thumb": "img/case09/ref_t.webp", "w": 1760, "h": 2368, "alpha": false}, "turbo": {"src": "img/case09/turbo.webp", "thumb": "img/case09/turbo_t.webp", "w": 1312, "h": 1792, "alpha": false}, "turbo_s": 2.9, "base": {"src": "img/case09/base.webp", "thumb": "img/case09/base_t.webp", "w": 1312, "h": 1792, "alpha": false}, "base_s": 14.2}, {"id": "case10", "title": "Mangrove portrait", "title_zh": "红树林人像", "prompt": "帮我生成一张抓拍风格的海滨红树林生活照：以输入图片中这位年轻女性为身份基准，其面部五官、黑色长直发与耳畔的蓝色耳饰均沿用原图不变，围绕她重建整个画面为中国南方海岸的红树林湿地场景。画面中的人物换上一件纯色白圆领短袖T恤，下身穿着带侧袋的宽松卡其色工装裤，脚踩米白色帆布鞋，右手自然垂下、用手指捏住一顶宽檐草编帽的帽檐，帽身垂落在腿侧，左臂随步伐自然弯曲摆动，左脚向前迈踏在桥面上，呈现行走中的半步动态；她的上身与面部自然朝向镜头，目光落在镜头上，嘴角保留原本轻扬的抿唇浅笑，如同行走间被同伴忽然唤住而抓拍下来，几缕发丝随海风轻轻飘动。场景是一条伸向红树林深处的木质浮桥：桥面由带细小缝隙的风化灰褐色木板铺成，从画面前景底部向远处延伸并向左后方弧形转弯；桥两侧密布红树林的支柱根与呼吸根，拱形插入泥滩之中，浅滩上积着一层薄薄海水，映出天空与树影的倒影，湿润的泥面呈现深褐色质感；中景左右两侧的红树林树干交错而立，背景处桥身与树影逐渐虚化成深绿色的湿地纵深，画面上方被浓密的树冠遮蔽，桥面、帽身与衣物上均为素面而无任何可读文字。温暖的午后阳光透过树叶缝隙洒落，在桥面、叶片与人物的肩头、发丝上投下斑驳光斑，人物面部处于柔和明亮的散射光中，皮肤保留原有的自然质感，光斑与树影在桥面和水面上方向一致地铺陈，形成统一的暖调光影逻辑。构图为全身环境抓拍取景，人物位于画面中央偏右，占据画面高度的四分之三，双脚扎实踩在木板上并与桥面形成清晰的接触与投影；镜头以数米外的中距离平视拍摄，中长焦段将背景湿地渲染出柔和虚化而保留层次，整体色调温暖自然，呈现真实、松弛的抓拍生活照质感。", "size": [1344, 1760], "inputs": [{"src": "img/case10/in1.webp", "thumb": "img/case10/in1_t.webp", "w": 896, "h": 1184, "alpha": false}], "reference": {"src": "img/case10/ref.webp", "thumb": "img/case10/ref_t.webp", "w": 1792, "h": 2368, "alpha": false}, "turbo": {"src": "img/case10/turbo.webp", "thumb": "img/case10/turbo_t.webp", "w": 1344, "h": 1760, "alpha": false}, "turbo_s": 2.8, "base": {"src": "img/case10/base.webp", "thumb": "img/case10/base_t.webp", "w": 1344, "h": 1760, "alpha": false}, "base_s": 14.0}, {"id": "case11", "title": "Indoor candid portrait", "title_zh": "室内抓拍人像", "prompt": "对输入图片进行完整的场景重绘编辑，为同一位年轻女性生成一张室内近距离抓拍人像：完全保留她的面部特征、五官比例、黑色蓬松长卷发型，以及她所穿的费尔岛花纹针织毛衣与针织围巾不变。场景设定在一间温馨的中国城市公寓室内，人物坐在实木餐桌后方，原本正低头专注地拆桌上的快递纸箱，此刻突然抬头发现镜头，脸上绽出明朗的开怀笑容，嘴角明显上扬、面颊随笑意提起、眼睛弯成月牙状，带着惊喜而愉悦的神情；上半身仍保持向桌侧自然前倾的姿态，头部由低垂状态顺畅地抬起转向镜头，目光直视镜头，双手自然按在已敞开的纸箱两侧箱盖上。取景为近距离半身抓拍视角，桌沿与桌上物件构成前景，人物位于画面略偏中心一侧。桌面上快递纸箱的箱盖向外摊开，箱体上的快递面单朝向人物一侧，镜头方向只可见素色牛皮纸箱表面与几道透明胶带痕迹，纸箱旁散落着一把银色金属剪刀、一团揉皱的牛皮包装纸和两片气泡膜，共同构成真实的拆快递瞬间。人物身后是居家内景，可见米色布艺沙发与原木收纳柜，墙上一盏暖黄色壁灯亮着，背景整体在焦点偏软的效果下呈现轻微虚化。光线以CCD卡片机的机位正面直闪为主光，人物面部、手部与桌面物件被照得明亮均匀，在人物身后的背景墙上投下一层柔和阴影，室内暖色环境光作为弱补光，使人物与环境融于同一套抓拍光逻辑。焦点落在人物面部与桌上纸箱之间，全图带有轻微的跑焦柔感，画面边缘出现轻微虚化与暗角，并叠加CCD直出特有的细小噪点与怀旧色彩质感，整体呈现一张连贯、真实、充满生活气息的室内抓拍快照。", "size": [1344, 1760], "inputs": [{"src": "img/case11/in1.webp", "thumb": "img/case11/in1_t.webp", "w": 1760, "h": 2336, "alpha": false}], "reference": {"src": "img/case11/ref.webp", "thumb": "img/case11/ref_t.webp", "w": 1760, "h": 2336, "alpha": false}, "turbo": {"src": "img/case11/turbo.webp", "thumb": "img/case11/turbo_t.webp", "w": 1344, "h": 1760, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case11/base.webp", "thumb": "img/case11/base_t.webp", "w": 1344, "h": 1760, "alpha": false}, "base_s": 14.4}, {"id": "case12", "title": "Hairstyle editing", "title_zh": "发型编辑", "prompt": "将输入图片中年轻女性的发型由当前盘起的束发与脸颊旁垂落的碎发改为蓬松的大波浪长卷发，让长发自头顶自然披落至双肩与胸前，卷度呈大而清晰的波浪弧形、层次分明且发丝带有柔和光泽，原本散落在脸颊与颈侧的碎发顺势融入整体波浪卷的走向之中；新发型的受光与阴影顺应画面既有的柔和户外自然光，在卷曲的凹陷处形成自然的柔和暗部、在发脊处形成细腻高光。此为在原图上的局部编辑，保持原图的取景、构图、缩放级别与人物在画面中的大小位置不变，不作重新构图或整体重绘，画面其余部分均与输入图保持一致。", "size": [1280, 1824], "inputs": [{"src": "img/case12/in1.webp", "thumb": "img/case12/in1_t.webp", "w": 864, "h": 1216, "alpha": false}], "reference": {"src": "img/case12/ref.webp", "thumb": "img/case12/ref_t.webp", "w": 1728, "h": 2432, "alpha": false}, "turbo": {"src": "img/case12/turbo.webp", "thumb": "img/case12/turbo_t.webp", "w": 1280, "h": 1824, "alpha": false}, "turbo_s": 2.7, "base": {"src": "img/case12/base.webp", "thumb": "img/case12/base_t.webp", "w": 1280, "h": 1824, "alpha": false}, "base_s": 13.7}, {"id": "case13", "title": "Expression editing", "title_zh": "表情编辑", "prompt": "将输入图片中并肩站立的两位年轻女性的表情由抿嘴浅笑改为一同开口大笑：两人均张嘴露齿开怀大笑，笑肌提起、嘴角大幅上扬，双眼因笑意弯成明亮的月牙形，眉宇舒展，两人的笑容情绪相互呼应、程度一致；同时将动作调整为一起比耶的活泼状态，左侧女生把原本背在身后的一只手举起到脸颊外侧，食指与中指自然伸直张开成V字比耶手势、手指结构完整清晰，右侧女生已举起的比耶手调整为更俏皮用力的角度，掌中所握的白色梳子仍稳稳握持、握姿符合真实受力关系；再配以轻快的肢体语言，两人头部朝彼此方向歪斜成俏皮的靠近角度、肩膀放松且身体轻轻相靠，使整幅画面充满活泼灵动的互动感。此为在原图上的局部编辑，保持原图的取景、构图、缩放级别与两人在画面中的大小位置不变，新增抬起的手臂与手部、以及大笑表情的受光与阴影均遵循原图户外自然光的统一光照逻辑，画面其余部分与输入图保持一致、不作重绘或重新构图。", "size": [1344, 1760], "inputs": [{"src": "img/case13/in1.webp", "thumb": "img/case13/in1_t.webp", "w": 896, "h": 1184, "alpha": false}], "reference": {"src": "img/case13/ref.webp", "thumb": "img/case13/ref_t.webp", "w": 480, "h": 636, "alpha": false}, "turbo": {"src": "img/case13/turbo.webp", "thumb": "img/case13/turbo_t.webp", "w": 1344, "h": 1760, "alpha": false}, "turbo_s": 2.8, "base": {"src": "img/case13/base.webp", "thumb": "img/case13/base_t.webp", "w": 1344, "h": 1760, "alpha": false}, "base_s": 14.1}, {"id": "case14", "title": "Vintage photo restoration", "title_zh": "老照片修复", "prompt": "Significantly improve the resolution and overall clarity of this vintage photograph and restore it as a realistic, naturally colorized high-fidelity photograph. Remove age-related dust, scratches, spots, fading, and excessive grain; recover fine detail and balanced highlights and shadows without oversharpening, waxy skin, or painted textures. Restore the seated elderly man with unruly white hair and a mustache, holding a curved tobacco pipe in his mouth and writing on paper. Preserve his exact facial anatomy, wrinkles, gaze, head angle, pipe contact with the mouth, and both hands, including the precise grip of the pen and the hand resting near the paper. Clarify fine hair, eyebrows, mustache, natural skin texture, the ribbed knit sweater, white shirt collar, and curved pipe stem without altering his expression. Recover the patterned chair upholstery, dark shelves and books behind him, and the foreground paper while maintaining the original depth of field. Colorize with realistic skin tones, silver-white hair, a muted warm gray knitted sweater, ivory collar and paper, a dark brown wooden pipe with a black stem, subdued burgundy and gold-brown upholstery, and dark wooden bookshelves. Keep the original indoor light and tonal modeling; do not fabricate handwriting, add smoke, or change the chair pattern. Strictly preserve the original aspect ratio, framing, camera viewpoint, exact number and arrangement of people, identities, ages, poses, expressions, clothing designs, object placement, and occlusions. Preserve the original direction and character of the lighting. Use believable, restrained colors appropriate to the period and materials, with lifelike skin tones and no color bleeding; produce a full-color photograph, not monochrome or sepia. Where the original is unclear, retain natural softness rather than inventing anatomy, objects, symbols, or readable text. Do not crop, expand, beautify faces, modernize the scene, or alter existing printed marks.", "size": [1344, 1760], "inputs": [{"src": "img/case14/in1.webp", "thumb": "img/case14/in1_t.webp", "w": 678, "h": 904, "alpha": false}], "reference": {"src": "img/case14/ref.webp", "thumb": "img/case14/ref_t.webp", "w": 1728, "h": 2304, "alpha": false}, "turbo": {"src": "img/case14/turbo.webp", "thumb": "img/case14/turbo_t.webp", "w": 1344, "h": 1760, "alpha": false}, "turbo_s": 2.9, "base": {"src": "img/case14/base.webp", "thumb": "img/case14/base_t.webp", "w": 1344, "h": 1760, "alpha": false}, "base_s": 14.1}, {"id": "case15", "title": "Photo stylization", "title_zh": "风格化", "prompt": "Transform the supplied amusement-park photograph into the requested visual medium across the entire image. Preserve the original vertical composition, aspect ratio, low-angle viewpoint, Ferris wheel silhouette and hub position, recognizable spoke geometry, red and yellow gondolas, green support tower at left, partially cropped ride at right, blue parasols and kiosks, wooden boardwalk, and the arrangement and poses of the foreground visitors. Keep the large open sky and original foreground-to-background depth. Retain recognizable placement and wording of existing signs such as PACIFIC PARK, POPCORN, CHURRO, and PHOTOS where legible, without adding new slogans or captions. Render a luminous impressionist oil painting on fine canvas. Construct every object from confident layered brushstrokes, broken color, subtle impasto highlights, and painterly transitions. Use warm late-afternoon light, a softly brushed blue-green sky, creamy wheel structure, rich cadmium red and golden yellow gondolas, deep green supports, and cool violet-blue shadows in the busy foreground. Let directional strokes describe the wheel's curved rim and radial spokes without bending or melting its mechanical structure. Paint the visitors with economical expressive marks and the boardwalk with broad receding strokes. Keep the scene recognizable and compositionally faithful while making the entire surface visibly painted; no photographic patches, generic blur, digital plastic shading, decorative frame, or signature.", "size": [1248, 1888], "inputs": [{"src": "img/case15/in1.webp", "thumb": "img/case15/in1_t.webp", "w": 2376, "h": 3583, "alpha": false}], "reference": {"src": "img/case15/ref.webp", "thumb": "img/case15/ref_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo": {"src": "img/case15/turbo.webp", "thumb": "img/case15/turbo_t.webp", "w": 1248, "h": 1888, "alpha": false}, "turbo_s": 3.0, "base": {"src": "img/case15/base.webp", "thumb": "img/case15/base_t.webp", "w": 1248, "h": 1888, "alpha": false}, "base_s": 14.3}, {"id": "case16", "title": "Travel planner UI", "title_zh": "旅行规划应用界面", "prompt": "The image is a vertical smartphone screenshot of a modern travel planner app interface named \"Wanderlust\", designed with a clean, light-themed aesthetic, soft shadows, and rounded corners, fully occupying the screen. The top edge features a standard status bar with black text \"9:41\" on the left, and on the right, black cellular signal bars, a Wi-Fi icon, and a battery icon showing \"85%\". Below the status bar is a top navigation bar. On the left, a circular user avatar shows a smiling young woman with a light blue background. Next to it, bold black text reads \"Good morning,\nSarah\". On the far right, a dark gray rounded square icon contains a white gear symbol for settings.\n\nThe main hero section is a large, rounded-corner card with a soft drop shadow. The background is a vibrant, high-quality photograph of the Santorini coastline with white cubic buildings and bright blue domes against a clear sky. Overlaid on the bottom left of the image is a semi-transparent dark gradient. Text inside reads \"Santorini,\nGreece\" in large bold white font, followed by \"Oct 12 -\nOct 19\" in smaller white font. On the bottom right, a solid white rounded button contains black text \"View\nItinerary\". At the very top right of this card, a small white pill-shaped badge contains black text \"5 Days\nLeft\".\n\nBelow the hero card is a section titled \"Plan Your Trip\" in bold dark gray text. Underneath is a row of four evenly spaced, light gray rounded square cards with subtle borders. The first card contains a blue airplane icon and black text \"Flights\" below. The second card contains an orange bed icon and black text \"Hotels\" below. The third card contains a green camera icon and black text \"Activities\" below. The fourth card contains a purple train icon and black text \"Transfers\" below.\n\nThe next section is titled \"Suggested for You\" in bold black text, with a blue text link \"See All\" aligned to the right. Below is a vertical list of three destination recommendation cards, each with a white background and a soft shadow. The first card has a square thumbnail of the Eiffel Tower on the left. The right side has the title \"Paris\" in bold black, the subtitle \"Romantic\nGetaway\" in dark gray, and a bottom row containing a yellow star icon, black text \"4.8\", and a small light gray pill with black text \"Flights from\n$450\". The second card has a square thumbnail of the Colosseum on the left. The right side has the title \"Rome\" in bold black, the subtitle \"Historical\nTour\" in dark gray, and a bottom row containing a yellow star icon, black text \"4.7\", and a small light gray pill with black text \"Flights from\n$520\". The third card has a square thumbnail of a Swiss Alpine village on the left. The right side has the title \"Zermatt\" in bold black, the subtitle \"Mountain\nRetreat\" in dark gray, and a bottom row containing a yellow star icon, black text \"4.9\", and a small light gray pill with black text \"Flights from\n$600\".\n\nBelow the suggestions is a section titled \"Past Trips\" in bold black text. This section contains a vertical list of two past trip history cards. The first card features a rectangular thumbnail of the neon-lit Shibuya crossing on the left. The middle area displays the title \"Tokyo,\nJapan\" in bold black text and the dates \"Aug 05 -\nAug 12\" in dark gray text. The right side features a solid blue rounded button with white text \"Rebook\". The second card features a rectangular thumbnail of the red rocks of the Grand Canyon on the left. The middle area displays the title \"Arizona,\nUSA\" in bold black text and the dates \"May 10 -\nMay 15\" in dark gray text. The right side features a solid blue rounded button with white text \"Rebook\".\n\nThe bottom layer is a fixed bottom navigation bar with a pure white background and a subtle top shadow, spanning the full width. It contains four evenly spaced interactive elements. The first element is a solid blue home icon with blue text \"Home\" directly below. The second element is a gray magnifying glass icon with gray text \"Explore\" directly below. The third element is a gray calendar icon with gray text \"Bookings\" directly below. The fourth element is a gray person icon with gray text \"Profile\" directly below. The entire interface uses a consistent design language with uniform corner radii, a harmonious color palette of white, light gray, dark gray, and vibrant blue accents, and clear typographic hierarchy ensuring all text is perfectly readable.", "size": [1536, 2720], "inputs": [], "reference": {"src": "img/case16/ref.webp", "thumb": "img/case16/ref_t.webp", "w": 798, "h": 1420, "alpha": false}, "turbo": {"src": "img/case16/turbo.webp", "thumb": "img/case16/turbo_t.webp", "w": 1536, "h": 2720, "alpha": false}, "turbo_s": 4.9, "base": {"src": "img/case16/base.webp", "thumb": "img/case16/base_t.webp", "w": 1536, "h": 2720, "alpha": false}, "base_s": 26.9, "turbo8": {"src": "img/case16/turbo8.webp", "thumb": "img/case16/turbo8_t.webp", "w": 1536, "h": 2720, "alpha": false}, "turbo8_s": 6.3}, {"id": "case17", "title": "Mathematics exam paper", "title_zh": "数学试卷", "prompt": "这是一张2026年数学高考试卷的平面设计图，画面直接呈现试卷版面，全出血排版，试卷边缘即为画布边缘。试卷整体为纯白底色，文字和图形均为黑色印刷效果，排版严谨规范，采用标准的高考卷面布局。页面左侧边缘有一条垂直的虚线作为装订密封线。页面顶部居中位置印有黑色粗体大字“绝密★启用前”。其下方是大标题，居中印有“2026年普通高等学校招生全国统一考试 数学”。大标题下方是“注意事项”部分，使用黑色正楷体，分条列出考生须知，内容依次为：“1. 答卷前，考生务必将自己的 姓名、准考证号填写在答题卡上。 2. 回答选择题时，选出每小题答案后， 用铅笔把答题卡上对应题目的答案标号涂黑。 3. 回答非选择题时，将答案写在答题卡上。 写在本试卷上无效。 4. 考试结束后，将本试卷和答题卡一并交回。”注意事项下方是试卷正文部分，采用标准宋体印刷。第一大部分标题为“一、单项选择题：本题共8小题， 每小题5分，共40分。”紧接着是第1题：“1. 已知集合A={x|x²-3x+2=0}， B={x|x²-2x=0}，则A∩B=”，下方列出四个选项：“A. {0} B. {1} C. {2} D. {0, 2}”。第2题为：“2. 若复数z满足z(1+i)=2， 则|z|=”，下方选项为：“A. 1 B. √2 C. 2 D. 2√2”。中间穿插有填空题部分，标题为“二、填空题：本题共4小题， 每小题5分，共20分。”，其中第9题为：“9. 已知向量a=(1, 2)，b=(x, -1)， 若a⊥b，则x=______。”页面下半部分为解答题区域，标题为“三、解答题：本题共5小题，共70分。”第15题内容为：“15.（13分）在△ABC中，内角A, B, C 的对边分别为a, b, c，已知 acosC+ccosA=bsinB。 （1）求角B的大小； （2）若b=2，求△ABC面积的最大值。”题目右侧空白处配有一个简单的黑色线条绘制的三角形几何示意图，顶点分别标有字母“A”、“B”、“C”。整个卷面文字清晰，数学公式排版规范，行距适中，呈现出标准的高考试卷视觉效果。", "size": [1664, 2496], "inputs": [], "reference": {"src": "img/case17/ref.webp", "thumb": "img/case17/ref_t.webp", "w": 1060, "h": 1584, "alpha": false}, "turbo": {"src": "img/case17/turbo.webp", "thumb": "img/case17/turbo_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo_s": 4.8, "base": {"src": "img/case17/base.webp", "thumb": "img/case17/base_t.webp", "w": 1664, "h": 2496, "alpha": false}, "base_s": 26.5, "turbo8": {"src": "img/case17/turbo8.webp", "thumb": "img/case17/turbo8_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo8_s": 6.4}, {"id": "case18", "title": "Academic infographic", "title_zh": "学术信息图", "prompt": "这是一张白色背景的学术论文式多面板信息图，整体以黑色无衬线标题、浅蓝色圆角模块、灰色坐标轴和柔和色彩的数据可视化构成。页面左上区域标题为 \"(a) Overall Architecture of UniMind\"，展示 UniMind 的整体架构流程。该区域左侧有标题 \"Multimodal Inputs\"，下面按纵向排列四种输入标签 \"Image\"、\"Text\"、\"Audio\"、\"Depth\"。第一行是一个小型写实照片，内容为一只浅黄色狗正面张嘴坐着或站着，右侧箭头指向浅蓝色圆角框，框中文字为 \"Vision Encoder (EvT)\"；第二行是浅蓝色文本卡片，内部写着 \"A dog is playing in the park.\"，箭头指向浅蓝色圆角框 \"Text Encoder (RoBERTa)\"；第三行是浅绿色音频卡片，包含绿色波形、左下角播放三角形和右下角 \"3s\"，箭头指向浅绿色圆角框 \"Audio Encoder (AST)\"；第四行是浅紫色深度图卡片，画面为灰紫色人物剪影深度图，箭头指向浅紫色圆角框 \"Depth Encoder (ResNet)\"。四个编码器右侧各有一组小圆角方块 token，顶部标注 \"Modality Tokens\"，颜色分别为蓝色、浅蓝色、绿色和紫色，并由连线汇入中央大型浅蓝色圆角矩形。中央上方写有 \"× N Layers\"，矩形内部文字为 \"Cross-Modal Transformer (Shared)\"，下方有一条虚线弧形箭头指向 token 区域并标注 \"Cross-Modal Alignment\"。中央模块右侧以箭头连接到任务头区域，顶部写着 \"Task Heads\"，纵向四个圆角框分别写有 \"Image Understanding\"、\"Text Generation\"、\"Audio Captioning\"、\"Depth Estimation\"，颜色与不同任务类型相呼应。页面右上区域标题为 \"(b) Zero-shot Transfer across Tasks\"，是一张分组柱状图，右上角图例依次为灰色 \"Single-Modal SOTA\"、蓝色 \"Multi-Modal SOTA\"、红色 \"UniMind (Ours)\"；纵轴标题为 \"Performance (Zero-shot)\"，刻度标有 \"20\"、\"40\"、\"60\"、\"80\"、\"100\"，横轴五组任务标签依次为 \"VQA (v2)\"、\"Text→Image (Gen)\"、\"Audio→Text (Cap)\"、\"Depth (Est)\"、\"Retrieval (Image↔Text)\"。每组都有三根柱，红色 UniMind 柱最高，并在上方以红字标注提升值 \" +7.6\"、\"+10.2\"、\"+8.9\"、\"+6.1\"、\"+9.3\"。页面左下区域标题为 \"(c) Qualitative Results\"，是一个三列表格，表头为 \"Image\"、\"Text Prompt / Input\"、\"UniMind Output\"。第一行左侧是金门大桥夕阳照片，橙色天空与平静水面构成风景，中间提示语为 \"Describe the scene.\"，右侧浅蓝输出框写着 \"A beautiful sunset over the Golden Gate Bridge with orange sky and calm water.\"。第二行左侧是彩色音频频谱图，呈现紫蓝绿色竖向频谱纹理，中间文字为 \"What is happening in the audio?\"，右侧浅绿色输出框写着 \"People are cheering and clapping in a stadium.\"。第三行左侧是灰色深度图，画面中有多个人形剪影，中间文字为 \"Estimate depth and objects.\"，右侧是带检测框的深度估计结果，画面上覆盖多个半透明矩形框与标签，包括三处 \"person\" 和一处 \"tree\"，背景呈紫橙色深度渐变。第四行左侧是狗在草地上奔跑的写实照片，中间文字为 \"Generate a caption.\"，右侧浅橙色输出框写着 \"A happy dog running on the grass in a park.\"。页面下方中部标题为 \"(d) Efficiency Comparison\"，是一张折线图，右上角图例显示灰色 \"OpenFlamingo\"、蓝色 \"LLaVA\"、红色 \"UniMind (Ours)\"，纵轴为 \"Performance (Average)\"，刻度标有 \"50\"、\"60\"、\"70\"、\"80\"、\"90\"，横轴为 \"Training FLOPs\"，刻度标注 \"10^19\"、\"10^20\"、\"10^21\"。三条带圆点的折线随训练 FLOPs 增大而上升，红色 UniMind 曲线整体高于蓝色和灰色曲线，图中红色注释写着 \"Better Performance with Fewer FLOPs\"，旁边有红色向上箭头。页面右下上半部分标题为 \"(e) Ablation Study\"，是一张浅灰网格表格，表头依次为 \"Model Variant\"、\"VQA\"、\"Cap\"、\"Depth\"、\"Avg\"。表格行内容为 \"Vision Only\"、\"62.4\"、\"-\"、\"-\"、\"62.4\"；\"+ Text\"、\"68.1\"、\"57.3\"、\"-\"、\"62.7\"；\"+ Audio\"、\"70.2\"、\"64.5\"、\"-\"、\"67.4\"；\"+ Depth\"、\"71.5\"、\"65.8\"、\"60.1\"、\"69.1\"；\"w/o Alignment\"、\"66.3\"、\"59.1\"、\"55.2\"、\"60.2\"；最后一行以红色强调 \"UniMind (Full)\"、\"74.0\"、\"67.8\"、\"64.3\"、\"68.7\"。页面右下下半部分标题为 \"(f) Cross-Modal Alignment Visualization\"，是一张二维散点嵌入可视化图，左侧标注 \"Aligned Semantics\" 并有黑色弧形箭头指向点云区域。点云由多个颜色簇组成，周围有彩色虚线椭圆轮廓，右侧图例标注红色点为 \"Image\"、蓝色点为 \"Text\"、绿色点为 \"Audio\"、紫色点为 \"Depth\"，散点簇之间部分重叠，表达跨模态语义对齐。底部是一整段论文图注，开头加粗样式为 \"Figure 1:\"，斜体标题为 \"UniMind: A Unified Multimodal Foundation Model.\"，以下为完整正文：\"(a) The overall architecture of UniMind, which employs modality-specific encoders to extract tokens from heterogeneous inputs and a shared cross-modal transformer to learn aligned representations for multiple tasks. (b) Zero-shot transfer performance on five representative tasks. UniMind consistently outperforms single-modal and previous multi-modal baselines. (c) Qualitative results on dense prediction, captioning, and generation across different modalities. (d) Performance-efficiency trade-off. UniMind achieves better performance with fewer training FLOPs compared to strong baselines. (e) Ablation study showing the contribution of each modality and the importance of cross-modal alignment. (f) Visualization of the learned cross-modal embeddings, where semantically related samples from different modalities are well aligned in the joint space.\"", "size": [2272, 1824], "inputs": [], "reference": {"src": "img/case18/ref.webp", "thumb": "img/case18/ref_t.webp", "w": 2288, "h": 1840, "alpha": false}, "turbo": {"src": "img/case18/turbo.webp", "thumb": "img/case18/turbo_t.webp", "w": 2272, "h": 1824, "alpha": false}, "turbo_s": 4.9, "base": {"src": "img/case18/base.webp", "thumb": "img/case18/base_t.webp", "w": 2272, "h": 1824, "alpha": false}, "base_s": 27.0, "turbo8": {"src": "img/case18/turbo8.webp", "thumb": "img/case18/turbo8_t.webp", "w": 2272, "h": 1824, "alpha": false}, "turbo8_s": 6.3}, {"id": "case19", "title": "Architecture presentation board", "title_zh": "建筑讲解图板", "prompt": "这是一张横向展开的建筑讲解图板，整体以米白羊皮纸质感背景、砂岩色与深棕色线稿为主，画面边缘和各分区使用细深棕色实线框与分割线组织版面。左上角是盾形城堡徽章标志，深棕色线框内有三座塔楼剪影和下方波浪纹，旁边的大标题为“奥拉维城堡”，下方英文副标题为“OLAVINLINNA CASTLE”。标题下方是一段说明文字，以下为完整正文：“奥拉维城堡建于1475年，是芬兰中世纪\n防御建筑的代表作。城堡位于萨翁林纳\n湖中的岩岛上，三面环水，地势险要，\n具有重要的历史与战略价值。”画面主体是一幅大型俯视透视建筑插画，约占页面上半部大部分宽度，表现奥拉维城堡坐落在湖中岩岛上，周围是淡灰蓝色水面和岩石岸线，城堡由多座圆形石塔、厚重环墙、内庭院、桥梁和入口组成，线稿细密，石材纹理、窗洞、箭孔和屋顶铜板分缝清晰可见。主视觉上有深棕色圆形编号点与虚线引导标注，桥梁入口处标注“1 主入口”，左侧内院标注“2 第一庭院”，中央内院标注“2 第二庭院”，中央高大的圆形主塔上方标注“4 圣奥拉夫塔\n（主塔）”，右侧圆塔标注“5 小塔楼”，右侧外墙区域标注“6 环墙”。主视觉右上方还有一个简洁罗盘，圆形细线外框内是深棕色北向指针，上方文字为“N”。右侧竖向信息栏由多个矩形框组成，顶部框标题为“基本信息”，内部每行左侧配有线性小图标，依次是钟表、城堡、建筑山墙、立方体材料图标和面积立方体图标，对应文字为“建造时间：1475年”、“建筑类型：水上城堡”、“建筑风格：中世纪防御建筑”、“建筑材料：当地花岗岩”、“建筑面积：约1,780㎡”。其下方框标题为“平面概览”，包含一幅小型平面图，使用深棕线条描绘三个圆塔、连墙、入口桥与内部建筑轮廓，下方有比例尺，刻度文字为“0”、“10”、“20”、“30m”，右下角还有一个小圆形方向符号。再下方框标题为“立面比例分析”，展示一幅小型立面比例示意，虚线水平标尺旁标有“1.0”、“2.2”、“1.0”，用以比较塔顶、塔身和基座等高度关系。页面中部偏下是一幅大型南立面建筑线稿，标题为“南立面图 1:200”，立面显示三座圆塔与连接墙体、入口桥、岩石基座和细小人群尺度，塔顶为锥形铜色屋面，塔身为浅砂岩色块石墙面。立面上方有水平尺寸标注“18.2m”、“24.6m”、“18.2m”，左侧有竖向高度标注“32.8m”、“21.6m\n（塔身）”、“11.2m\n（墙体）”。立面右侧用虚线引出编号说明，依次为“1 塔顶\n铜质尖顶，便于排水”、“2 射击孔\n圆形孔洞，用于防御”、“3 拱形窗\n竖向狭长窗，节奏均匀”、“4 墙体开窗节奏\n小窗与箭孔交替布置，\n形成稳定的律”、“5 墙体材料\n当地花岗岩砌筑，厚实\n坚固，适应水上环境”、“6 基座\n依岩而建，直接与基岩\n结合，增强稳定性”。右侧中下部是“局部放大详图”区域，采用三个横向分隔的细框卡片，每个卡片左侧是放大的构造线稿，右侧是引线文字。第一项标题为“1 塔顶构造 1:20”，图中放大圆塔顶部，标注“铜板屋面”、“木构基层”、“砖石檐口”。第二项标题为“2 拱形窗构造 1:20”，展示拱形窗节点，标注“花岗岩拱券”、“花岗岩侧壁”、“木质窗框”、“铁制格栅”。第三项标题为“3 墙体与箭孔构造 1:20”，展示块石墙体与狭长箭孔，标注“花岗岩块石”、“箭孔”、“内侧抹灰”。页面底部左侧是“入口人流分析”模块，左边为图例，深棕实线箭头代表“主要人流动线”，深棕虚线箭头代表“次要人流动线”，深棕圆点代表“停留节点”，实心箭头代表“人流方向”；右侧小插图以俯视方式表现桥梁、入口门楼、岛屿平台和众多微型人物剪影，箭头沿桥面与门楼前广场组织人流方向。底部中间是“材料图例”模块，使用五个圆形材质样本展示石材、砖、铜板、木材和铁件的纹理，文字分别为“花岗岩\n主要墙体材料”、“砖砌\n拱券及局部构造”、“铜板\n屋面覆盖材料”、“木材\n门窗及结构构件”、“铁件\n门窗格栅及构件”。底部右侧是“构造剖面示意 1:100”，展示塔楼剖面与连接墙体，塔顶木屋架、楼板、塔身空间、墙体厚度和岩石基础以剖切方式呈现，右侧编号引线标注为“1 塔顶结构”、“2 木质屋架”、“3 塔身空间”、“4 楼板结构”、“5 墙体厚度\n（3.5~4.5m）”、“6 基岩基础”。整体图面像专业建筑展示板，线条清晰、图例完整、编号系统统一，带有历史建筑手绘测绘图与信息可视化结合的风格。", "size": [2048, 2048], "inputs": [], "reference": {"src": "img/case19/ref.webp", "thumb": "img/case19/ref_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo": {"src": "img/case19/turbo.webp", "thumb": "img/case19/turbo_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo_s": 4.9, "base": {"src": "img/case19/base.webp", "thumb": "img/case19/base_t.webp", "w": 2048, "h": 2048, "alpha": false}, "base_s": 27.6, "turbo8": {"src": "img/case19/turbo8.webp", "thumb": "img/case19/turbo8_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo8_s": 6.9}, {"id": "case20", "title": "Group portrait (6 images)", "title_zh": "多人合照（6 图）", "prompt": "Using the six people in <image1>, <image2>, <image3>, <image4>, <image5> and <image6> as identity references, generate a brand-new vertical group portrait in the style of a 1980s sitcom promotional photo, with all six posed together in a vintage bar interior. Scene and background: on a warm beige plaster wall, a gilded carved eagle with spread wings is mounted high; below it hangs a horizontal framed panel of four black-and-white vintage photographs showing a ship hull under construction in a shipyard; at the left edge of the frame a red triangular pennant protrudes, bearing white lettering \"Boston RED SOX WORLD\" [partially occluded]; on the right runs a dark-brown wooden staircase with turned baluster spindles rising vertically; a Tiffany-style stained-glass pyramid pendant lamp hangs from a metal chain at upper center-right; the whole scene is lit with soft, warm indoor studio light and a muted film-like color tone, in an eye-level medium-full group composition. Arrangement and new states: the heavyset man from <image5> stands at the far left of the back row, now wearing a brown herringbone tweed suit jacket over a white shirt with a dark maroon tie, body angled toward the center with one hand resting on the seat back beside the woman in red; the woman from <image1> is at center-left, keeping the hairstyle from <image1> exactly unchanged — the same golden-blonde upswept style with soft volume on top and fringe bangs across the forehead, identical in length, parting, volume and shape, not restyled into loose or shoulder-length waves — wearing a bright red V-neck knit sweater and small earrings, her right hand (screen left) resting on the shoulder of the young man seated in front of her; the man from <image3> stands tallest at center back, now in a light-gray suit jacket, white shirt and red striped tie, smiling faintly; the man from <image2> stands at the far right of the back row, now in a dark navy zip jacket over a light-blue shirt, one hand resting on the shoulder of the man in the light-gray suit jacket; the young man from <image6> sits on the floor at front left, now in a pink-mauve finely striped polo shirt with a pale-yellow collar, one arm resting on his knee with a thin cord bracelet on that wrist and his other hand placed on his own shoulder; the woman from <image4> sits at front right, turned sideways toward the camera with her hands resting naturally in her lap. The six are arranged in two rows, some standing and some seated, with relaxed natural expressions and a warm vintage ensemble mood.", "size": [1248, 1888], "inputs": [{"src": "img/case20/in1.webp", "thumb": "img/case20/in1_t.webp", "w": 864, "h": 1216, "alpha": false}, {"src": "img/case20/in2.webp", "thumb": "img/case20/in2_t.webp", "w": 1024, "h": 1024, "alpha": false}, {"src": "img/case20/in3.webp", "thumb": "img/case20/in3_t.webp", "w": 832, "h": 1248, "alpha": false}, {"src": "img/case20/in4.webp", "thumb": "img/case20/in4_t.webp", "w": 832, "h": 1248, "alpha": false}, {"src": "img/case20/in5.webp", "thumb": "img/case20/in5_t.webp", "w": 832, "h": 1248, "alpha": false}, {"src": "img/case20/in6.webp", "thumb": "img/case20/in6_t.webp", "w": 1024, "h": 1024, "alpha": false}], "reference": {"src": "img/case20/ref.webp", "thumb": "img/case20/ref_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo": {"src": "img/case20/turbo.webp", "thumb": "img/case20/turbo_t.webp", "w": 1248, "h": 1888, "alpha": false}, "turbo_s": 6.0, "base": {"src": "img/case20/base.webp", "thumb": "img/case20/base_t.webp", "w": 1248, "h": 1888, "alpha": false}, "base_s": 25.8}, {"id": "case21", "title": "Outfit styling (5 images)", "title_zh": "多品穿戴（5 图）", "prompt": "让【图1】中的模特换上【图3】中的玛丽珍鞋，拿着【图4】中的手提包，并戴上【图5】中的绒毛帽。将【图2】中的羽绒服敞开穿在外面，露出原有的内搭上衣。保持模特姿势和背景不变。", "size": [1312, 1792], "inputs": [{"src": "img/case21/in1.webp", "thumb": "img/case21/in1_t.webp", "w": 896, "h": 1184, "alpha": false}, {"src": "img/case21/in2.webp", "thumb": "img/case21/in2_t.webp", "w": 832, "h": 832, "alpha": false}, {"src": "img/case21/in3.webp", "thumb": "img/case21/in3_t.webp", "w": 800, "h": 800, "alpha": false}, {"src": "img/case21/in4.webp", "thumb": "img/case21/in4_t.webp", "w": 1024, "h": 1024, "alpha": false}, {"src": "img/case21/in5.webp", "thumb": "img/case21/in5_t.webp", "w": 800, "h": 800, "alpha": false}], "reference": {"src": "img/case21/ref.webp", "thumb": "img/case21/ref_t.webp", "w": 1760, "h": 2368, "alpha": false}, "turbo": {"src": "img/case21/turbo.webp", "thumb": "img/case21/turbo_t.webp", "w": 1312, "h": 1792, "alpha": false}, "turbo_s": 5.2, "base": {"src": "img/case21/base.webp", "thumb": "img/case21/base_t.webp", "w": 1312, "h": 1792, "alpha": false}, "base_s": 23.3}, {"id": "case22", "title": "Room furnishing (10 images)", "title_zh": "多物布置（10 图）", "prompt": "Using the empty living room in <image1> as the scene and the furniture and decor in <image2>, <image3>, <image4>, <image5>, <image6>, <image7>, <image8>, <image9>, <image10> as references, generate a brand-new fully furnished interior image. The scene is the living room of <image1>: light wood plank floor, gray cement-textured walls, a dark gray door on the left wall, two narrow black-framed windows on the right wall, a dark gray skirting board along the wall bases, and multiple warm recessed downlights set into the ceiling in straight rows running the depth of the room; eye-level indoor viewpoint, natural perspective, square 1:1 framing, overall warm lighting. Layout: mount the black-framed wine shelving unit from the background of <image5> on the back wall, with several wine bottles and a small dried-flower vase on its shelves; in front of it place the dining table from <image5>, with a woven-leather dining chair behind it and three round stools lined up on the near side, two wine bottles and a wine glass on the tabletop; to the front-right of the dining table place the high-back armchair from <image4>, facing the room center; on the left side of the room place the rocking chair from <image3> with its ottoman in front; lay the rug from <image6> flat in the middle of the room and on it place the multi-tier storage coffee table from <image2>, its raised wooden top holding a dessert plate, a drink can, a vase of white flowers and a phone, its trays and lower shelves holding a camera, game controllers, a handheld console, a watch and snack drinks, with packaging texts “CRISPY BISCUITS”, “Classic Classique”, “纯牛奶”, “CHA GEILI”, “合味道”, “Coca-Cola” and “RIO” visible; along the right wall, occupying the right side of the frame, place the sofa from <image9>: a large curved kidney-shaped sofa in cream bouclé fabric with one continuous rounded backrest wrapping into low rolled ends, a single seamless seat cushion, no armrests and no visible legs, resting on a recessed dark-wood plinth base; on it place the cushion from <image8> plus a light gray pillow; behind the sofa set a low wooden console top along the wall, on which stand the arch-shaped stone table clock and the dried-flower vase from <image7>. In the near-left foreground corner of the frame, standing directly on the wood plank floor, place the potted rubber plant from <image10>: large glossy dark-green oval leaves on upright stems in a matte greige cylindrical ceramic pot with a matching round saucer, its height reaching roughly the seat level of the nearby seating. Only the objects themselves are taken from <image9> and <image10> — the brick-view window, sheer curtain, floor lamp, framed picture, round marble table and beige rug behind the sofa, and the window, fireplace and sofa behind the plant, do not appear. All furniture rests naturally on the floor with soft contact shadows, lit consistently by the room's downlights, with realistic proportions and occlusions.", "size": [1536, 1536], "inputs": [{"src": "img/case22/in1.webp", "thumb": "img/case22/in1_t.webp", "w": 1024, "h": 1024, "alpha": false}, {"src": "img/case22/in2.webp", "thumb": "img/case22/in2_t.webp", "w": 896, "h": 1184, "alpha": false}, {"src": "img/case22/in3.webp", "thumb": "img/case22/in3_t.webp", "w": 864, "h": 1216, "alpha": false}, {"src": "img/case22/in4.webp", "thumb": "img/case22/in4_t.webp", "w": 1024, "h": 1024, "alpha": false}, {"src": "img/case22/in5.webp", "thumb": "img/case22/in5_t.webp", "w": 1184, "h": 896, "alpha": false}, {"src": "img/case22/in6.webp", "thumb": "img/case22/in6_t.webp", "w": 1024, "h": 1024, "alpha": false}, {"src": "img/case22/in7.webp", "thumb": "img/case22/in7_t.webp", "w": 1184, "h": 896, "alpha": false}, {"src": "img/case22/in8.webp", "thumb": "img/case22/in8_t.webp", "w": 1184, "h": 896, "alpha": false}, {"src": "img/case22/in9.webp", "thumb": "img/case22/in9_t.webp", "w": 1184, "h": 896, "alpha": false}, {"src": "img/case22/in10.webp", "thumb": "img/case22/in10_t.webp", "w": 1184, "h": 896, "alpha": false}], "reference": {"src": "img/case22/ref.webp", "thumb": "img/case22/ref_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo": {"src": "img/case22/turbo.webp", "thumb": "img/case22/turbo_t.webp", "w": 1536, "h": 1536, "alpha": false}, "turbo_s": 9.9, "base": {"src": "img/case22/base.webp", "thumb": "img/case22/base_t.webp", "w": 1536, "h": 1536, "alpha": false}, "base_s": 36.6}, {"id": "case23", "title": "Local editing: annotations", "title_zh": "局部编辑：圈选", "prompt": "将蓝色框选区域内佩戴在抬起手腕上的棕色皮表带金属手表整体移除，并以与手臂肤色及明暗过渡一致的皮肤自然延续填补腕部区域；将红色框选区域内男生的蓬松金色头发改为黑色：黑色头发沿原有发丝走向与蓬松质感呈现完整的明暗层次��高光处为带微光泽的深灰色、阴影处为深黑色；将两处绿色框选区域内的衣服——画面左下左臂上的白色短袖袖口与右肩处的白色衣袖部分——换成灰色短袖亚麻睡衣：睡衣面料呈灰色，带有清晰的亚麻织纹与自然褶皱，袖口为宽松的短袖剪裁，边缘贴合无缝痕；蓝色、红色与绿色标注线不得渲染在图像中。", "size": [1568, 1504], "inputs": [{"src": "img/case23/in1.webp", "thumb": "img/case23/in1_t.webp", "w": 386, "h": 370, "alpha": false}], "reference": {"src": "img/case23/ref.webp", "thumb": "img/case23/ref_t.webp", "w": 382, "h": 370, "alpha": false}, "turbo": {"src": "img/case23/turbo.webp", "thumb": "img/case23/turbo_t.webp", "w": 1568, "h": 1504, "alpha": false}, "turbo_s": 2.8, "base": {"src": "img/case23/base.webp", "thumb": "img/case23/base_t.webp", "w": 1568, "h": 1504, "alpha": false}, "base_s": 14.0}, {"id": "case24", "title": "Local editing: mask", "title_zh": "局部编辑：掩码", "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。", "size": [1248, 1888], "inputs": [{"src": "img/case24/in1.webp", "thumb": "img/case24/in1_t.webp", "w": 246, "h": 370, "alpha": false}, {"src": "img/case24/in2.webp", "thumb": "img/case24/in2_t.webp", "w": 246, "h": 370, "alpha": false}], "reference": {"src": "img/case24/ref.webp", "thumb": "img/case24/ref_t.webp", "w": 245, "h": 370, "alpha": false}, "turbo": {"src": "img/case24/turbo.webp", "thumb": "img/case24/turbo_t.webp", "w": 1248, "h": 1888, "alpha": false}, "turbo_s": 3.3, "base": {"src": "img/case24/base.webp", "thumb": "img/case24/base_t.webp", "w": 1248, "h": 1888, "alpha": false}, "base_s": 16.1}, {"id": "case25", "title": "Local editing: painted region", "title_zh": "局部编辑：涂抹", "prompt": "在图中右侧白色掩码标注的区域添加一名水肺潜水员：潜水员身穿黑色湿衣，佩戴潜水面镜和背负式气瓶，身体略微倾斜地悬停在水中，面向镜头，调节器呼出的一串气泡向水面上升。", "size": [1888, 1248], "inputs": [{"src": "img/case25/in1.webp", "thumb": "img/case25/in1_t.webp", "w": 1536, "h": 1024, "alpha": false}], "reference": {"src": "img/case25/ref.webp", "thumb": "img/case25/ref_t.webp", "w": 1536, "h": 1024, "alpha": false}, "turbo": {"src": "img/case25/turbo.webp", "thumb": "img/case25/turbo_t.webp", "w": 1888, "h": 1248, "alpha": false}, "turbo_s": 2.8, "base": {"src": "img/case25/base.webp", "thumb": "img/case25/base_t.webp", "w": 1888, "h": 1248, "alpha": false}, "base_s": 13.9}, {"id": "en_1", "title": "Editorial portrait", "title_zh": "Editorial portrait", "prompt": "Create an editorial portrait of a botanist in a sunlit greenhouse, surrounded by ferns and delicate orchids. Natural skin texture, linen clothing, soft morning backlight, subtle film grain, medium-format photography, calm expression, no text or watermark.", "size": [1664, 2496], "inputs": [], "reference": null, "turbo": {"src": "img/en_1/turbo.webp", "thumb": "img/en_1/turbo_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo_s": 4.6, "base": {"src": "img/en_1/base.webp", "thumb": "img/en_1/base_t.webp", "w": 1664, "h": 2496, "alpha": false}, "base_s": 26.1}, {"id": "en_2", "title": "Typography poster", "title_zh": "Typography poster", "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.", "size": [1664, 2496], "inputs": [], "reference": null, "turbo": {"src": "img/en_2/turbo.webp", "thumb": "img/en_2/turbo_t.webp", "w": 1664, "h": 2496, "alpha": false}, "turbo_s": 4.6, "base": {"src": "img/en_2/base.webp", "thumb": "img/en_2/base_t.webp", "w": 1664, "h": 2496, "alpha": false}, "base_s": 25.9}, {"id": "en_3", "title": "Six-panel storyboard", "title_zh": "Six-panel storyboard", "prompt": "Create a six-panel cinematic storyboard about a small robot restoring an abandoned rooftop garden. Show: arrival at dawn, discovery of a dried seedling, repairing an irrigation pipe, planting new seeds, the first rain, and a lush garden at sunset. Keep the robot's round yellow body and blue eyes consistent in every panel. Clear panel borders, expressive visual storytelling, detailed environments, no captions.", "size": [2496, 1664], "inputs": [], "reference": null, "turbo": {"src": "img/en_3/turbo.webp", "thumb": "img/en_3/turbo_t.webp", "w": 2496, "h": 1664, "alpha": false}, "turbo_s": 4.7, "base": {"src": "img/en_3/base.webp", "thumb": "img/en_3/base_t.webp", "w": 2496, "h": 1664, "alpha": false}, "base_s": 26.1}, {"id": "en_4", "title": "Product photography", "title_zh": "Product photography", "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.", "size": [2496, 1664], "inputs": [], "reference": null, "turbo": {"src": "img/en_4/turbo.webp", "thumb": "img/en_4/turbo_t.webp", "w": 2496, "h": 1664, "alpha": false}, "turbo_s": 4.6, "base": {"src": "img/en_4/base.webp", "thumb": "img/en_4/base_t.webp", "w": 2496, "h": 1664, "alpha": false}, "base_s": 26.1}, {"id": "zh_1", "title": "Film character turnaround", "title_zh": "电影角色三视图", "prompt": "生成一张真人电影质感的单角色三视图设定参考图。干净灰色背景，左侧为一张较大的完整头肩半身细节图，必须是同一角色从头部到肩部/胸口的连续完整特写，清楚展示面部、发型、上身服装、配饰和材质细节；左侧不要拆成多个局部小图、不要拼贴多个细节框、不要只给眼睛/衣料/配饰等碎片特写。右侧展示同一角色正面、侧面、背面全身三视图，全身可见。角色设定如下：丹尼尔·斯通，纯白背景，无其他人物和场景；丹尼尔·斯通的单人全身立绘，正对镜头站立，表情自然；30岁青年男性，2020年代现代美国，欧洲裔白人外貌，浅肤色，身高约185cm，9头身，魁梧健壮体型，肌肉发达，2020年代棕色短寸头，发质粗硬，眼窝深陷，瞳孔呈灰蓝色，鼻梁高挺且鼻头宽大，唇形厚实且唇色苍白，方下颌轮廓分明，面部皮肤粗糙并带有污垢痕迹，穿着2020年代脏污灰色连帽卫衣配同色系工装裤，脚穿2020年代黑色防滑工装靴，双手佩戴破旧皮革手套，双手自然下垂；光照均匀，高画质，手部完美，无文字水印。 三视图中面部特征、身形比例、发型、服装、配饰、鞋履、姿态和关键视觉元素保持一致；表情自然中性，站姿稳定，比例统一，皮肤和布料金属等材质细节清楚，电影级写实光照，高画质，无文字、无logo、无水印。", "size": [2496, 1664], "inputs": [], "reference": null, "turbo": {"src": "img/zh_1/turbo.webp", "thumb": "img/zh_1/turbo_t.webp", "w": 2496, "h": 1664, "alpha": false}, "turbo_s": 4.7, "base": {"src": "img/zh_1/base.webp", "thumb": "img/zh_1/base_t.webp", "w": 2496, "h": 1664, "alpha": false}, "base_s": 27.1}, {"id": "zh_2", "title": "Shaun the Sheep birthday story", "title_zh": "小羊肖恩生日故事", "prompt": "3D粘土漫画，生成小羊肖恩\n镜头一：清晨的青苔底农场暖意融融，小羊肖恩发现农夫正在精心装饰巧克力生日蛋糕。镜头二：随即心生主意，独自带着小羊提米靠近厨房。镜头三：它悄悄避开熟睡的牧羊犬，带着提米溜进屋内。镜头四：不料提米意外打翻面粉、蹭损蛋糕，场面十分狼狈。镜头五：关键时刻肖恩灵机一动，伸出手指，在厚厚的白色粉末上流畅地画了起来。\n首先是一个巨大的、歪歪扭拙却充满爱意的爱心；接着是\"HAPPY BIRTHDAY\"的字样。随后，他指挥提米将散落的鲜红草莓摆放在爱心周围，又将从窗外采摘的野花（雏菊、矢车菊）插在蛋糕受损的边缘，巧妙地遮盖了瑕疵。原本狼藉的桌面，瞬间变成了一幅质朴而温馨的田园画作。肖恩退后一步，满意地拍了拍手上的面粉，脸上露出自豪的微笑。镜头六：农夫归来后，看见这份质朴又温馨的布置开怀大笑，这场小小的意外，最终变成了农场治愈又暖心的生日惊喜。", "size": [2496, 1664], "inputs": [], "reference": null, "turbo": {"src": "img/zh_2/turbo.webp", "thumb": "img/zh_2/turbo_t.webp", "w": 2496, "h": 1664, "alpha": false}, "turbo_s": 4.7, "base": {"src": "img/zh_2/base.webp", "thumb": "img/zh_2/base_t.webp", "w": 2496, "h": 1664, "alpha": false}, "base_s": 26.6}, {"id": "zh_7", "title": "Giant creature battle storyboard", "title_zh": "巨兽对决故事板", "prompt": "核心大纲 1. 场景：3DCG质感，废弃都市遗迹（黄昏），尘土漫天，绯红色夕阳，地面有巨大脚印。 2. 人物：炎狱巨蜥（熔岩鳞片、喷烈焰）、霜牙巨象（雪白、喷寒气）、幸存人类（远景蜷缩）。 3. 剧情脉络： （1）对峙（0-20秒）：两只巨兽在废墟对峙，人类蜷缩避险，氛围压迫。 （2）开战（21-50秒）：炎狱巨蜥喷烈焰，霜牙巨象喷寒气对抗，引发冲击，高楼坍塌。 （3）激战（51-90秒）：双方互相攻击，均受重伤，嘶吼震彻废墟。 （4）落幕（91-110秒）：两只巨兽重伤对峙，镜头拉远，定格巨兽与废墟，留下悬念", "size": [2048, 2048], "inputs": [], "reference": null, "turbo": {"src": "img/zh_7/turbo.webp", "thumb": "img/zh_7/turbo_t.webp", "w": 2048, "h": 2048, "alpha": false}, "turbo_s": 4.7, "base": {"src": "img/zh_7/base.webp", "thumb": "img/zh_7/base_t.webp", "w": 2048, "h": 2048, "alpha": false}, "base_s": 26.1}]