Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,170 @@
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+
+import gradio as gr
+
+CURRENT_FILE = Path(__file__).resolve()
+PROJECT_ROOT = CURRENT_FILE.parents[1]
+for candidate in (CURRENT_FILE.parent, CURRENT_FILE.parents[1]):
+    if (candidate / "infer").exists() and (candidate / "models").exists():
+        PROJECT_ROOT = candidate
+        break
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from demo.real_world_pipeline import (  # noqa: E402
+    DEFAULT_BBOX_MODEL,
+    DEFAULT_REAL_CONFIG_PATH,
+    DEFAULT_RUN_NAME,
+    DEFAULT_WORK_DIR,
+    run_real_world_pipeline,
+)
+
+DEFAULT_EXAMPLE_DIR = Path(
+    os.environ.get(
+        "SYNLAYERS_EXAMPLE_DIR",
+        "/project/llmsvgen/share/data/kmw_layered_dataset/real_world_inference/layers_real_test_1024",
+    )
+)
+
+
+def list_example_images(limit: int = 6) -> list[list[str]]:
+    if not DEFAULT_EXAMPLE_DIR.exists():
+        return []
+
+    candidates = []
+    for ext in ("*.png", "*.jpg", "*.jpeg", "*.webp"):
+        candidates.extend(DEFAULT_EXAMPLE_DIR.glob(ext))
+    candidates = sorted(candidates)[:limit]
+    return [[str(path)] for path in candidates]
+
+
+def build_gallery(result: dict) -> list[tuple[str, str]]:
+    gallery: list[tuple[str, str]] = []
+    if result.get("whole_image_rgba"):
+        gallery.append((result["whole_image_rgba"], "Whole RGBA"))
+    if result.get("background_rgba"):
+        gallery.append((result["background_rgba"], "Background RGBA"))
+    for idx, path in enumerate(result.get("layer_images", [])):
+        gallery.append((path, f"Layer {idx}"))
+    return gallery
+
+
+def run_demo(
+    image_path: str,
+    sample_name: str,
+    max_new_tokens: int,
+    seed_value: float,
+):
+    if not image_path:
+        raise gr.Error("Please upload an input image first.")
+
+    seed = int(seed_value) if seed_value >= 0 else None
+
+    try:
+        result = run_real_world_pipeline(
+            image_path=image_path,
+            sample_name=sample_name or None,
+            work_dir=DEFAULT_WORK_DIR,
+            bbox_model=DEFAULT_BBOX_MODEL,
+            config_path=DEFAULT_REAL_CONFIG_PATH,
+            max_new_tokens=int(max_new_tokens),
+            seed=seed,
+            run_name=DEFAULT_RUN_NAME,
+        )
+    except Exception as exc:
+        raise gr.Error(str(exc)) from exc
+
+    return (
+        result["bbox_visualization"],
+        result["merged_image"],
+        result["bbox_record"].get("whole_caption", ""),
+        result["bbox_record"],
+        result["metadata"],
+        build_gallery(result),
+        result["archive_path"],
+        result["case_dir"],
+    )
+
+
+with gr.Blocks(title="SynLayers Real-World Demo") as demo:
+    gr.Markdown(
+        """
+        # SynLayers Real-World Decomposition
+
+        Upload a single image and run the full pipeline in one step:
+        1. fixed-prompt VLM whole-caption + bounding-box detection
+        2. SynLayers real-image layer decomposition with the `step_120000` checkpoint
+
+        The app uses `demo/infer` for the VLM stage and `infer/infer.py` + `infer/infer.yaml`
+        for the decomposition stage.
+        """
+    )
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            image_input = gr.Image(type="filepath", label="Input Image")
+            sample_name_input = gr.Textbox(
+                label="Optional Sample Name",
+                placeholder="Leave empty to use the uploaded filename",
+            )
+            max_new_tokens_input = gr.Slider(
+                minimum=128,
+                maximum=2048,
+                value=1024,
+                step=64,
+                label="VLM Max New Tokens",
+            )
+            seed_input = gr.Number(
+                value=42,
+                precision=0,
+                label="Seed (-1 keeps config default)",
+            )
+            run_button = gr.Button("Run Full Pipeline", variant="primary")
+
+        with gr.Column(scale=1):
+            bbox_vis_output = gr.Image(type="filepath", label="Detected Bounding Boxes")
+            merged_output = gr.Image(type="filepath", label="Merged Decomposition")
+
+    caption_output = gr.Textbox(label="Whole Caption", lines=6)
+    with gr.Row():
+        bbox_json_output = gr.JSON(label="BBox JSON")
+        meta_json_output = gr.JSON(label="Inference Metadata")
+    layer_gallery = gr.Gallery(label="Predicted Layers", columns=4, height="auto")
+    with gr.Row():
+        archive_output = gr.File(label="Download Result Bundle")
+        case_dir_output = gr.Textbox(label="Saved Case Directory")
+
+    examples = list_example_images()
+    if examples:
+        gr.Examples(examples=examples, inputs=[image_input], label="Example Images")
+
+    run_button.click(
+        fn=run_demo,
+        inputs=[
+            image_input,
+            sample_name_input,
+            max_new_tokens_input,
+            seed_input,
+        ],
+        outputs=[
+            bbox_vis_output,
+            merged_output,
+            caption_output,
+            bbox_json_output,
+            meta_json_output,
+            layer_gallery,
+            archive_output,
+            case_dir_output,
+        ],
+    )
+
+
+if __name__ == "__main__":
+    demo.queue().launch(
+        server_name="0.0.0.0",
+        server_port=int(os.environ.get("PORT", "7860")),
+    )