update

.gitignore

@@ -1,3 +1,4 @@
 checkpoints
-__pycache__
-.pyc
+__pycache__/
+*.pyc
+example_data/

InfiniDepth/utils/hf_demo_utils.py

@@ -58,7 +58,8 @@ class ModelCache:
 
 def _parse_image_size(size_text: str) -> tuple[int, int]:
     try:
-        h_text, w_text = size_text.lower().split("x")
+        normalized = size_text.lower().replace(" ", "")
+        h_text, w_text = normalized.split("x")
         return int(h_text), int(w_text)
     except Exception as exc:
         raise ValueError(f"Invalid image size format: {size_text}, expected like 768x1024") from exc
@@ -228,6 +229,8 @@ def run_single_image_demo(
         raise ValueError("upsample_ratio must be in [1, 8]")
     if max_points_preview < 1000:
         raise ValueError("max_points_preview must be at least 1000")
+    else:
+        output_size = input_size
 
     device = torch.device("cuda")
     image, org_h, org_w = _prepare_image_tensor(image_np, input_size, device)
@@ -238,7 +241,7 @@ def run_single_image_demo(
         org_w=org_w,
         h=h_in,
         w=w_in,
-        output_size=input_size,
+        output_size=output_size,
         upsample_ratio=upsample_ratio,
     )
 

app.py

@@ -33,10 +33,19 @@ except ImportError:
 
 MODEL_CACHE = ModelCache()
 OUTPUT_ROOT = Path(tempfile.gettempdir()) / "infinidepth_hf_demo"
+EXAMPLE_DATA_ROOT = Path(__file__).resolve().parent / "example_data"
+EXAMPLE_IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".webp"}
+EXAMPLE_DEPTH_EXTENSIONS = {".png", ".npy", ".npz", ".h5", ".hdf5", ".exr"}
 
 CUSTOM_CSS = """
-.gradio-container {
-    max-width: 1280px !important;
+#main-layout {
+    width: 100%;
+    gap: 16px;
+}
+
+#left-panel,
+#right-panel {
+    min-width: 0 !important;
 }
 
 #hero {
@@ -93,6 +102,97 @@ def _none_if_invalid(value: Optional[float]) -> Optional[float]:
     return None
 
 
+def _strip_rgb_suffix(stem: str) -> str:
+    suffixes = ("_rgb", "-rgb", ".rgb", "_image", "-image", ".image", "_color", "-color", ".color")
+    lowered = stem.lower()
+    for suffix in suffixes:
+        if lowered.endswith(suffix):
+            return lowered[: -len(suffix)].strip("._-")
+    return lowered
+
+
+def _find_paired_depth(rgb_path: Path) -> Optional[Path]:
+    parent = rgb_path.parent
+    stem = rgb_path.stem.lower()
+    base = _strip_rgb_suffix(stem)
+
+    candidate_stems = []
+    for key in (stem, base):
+        if not key:
+            continue
+        candidate_stems.extend(
+            [
+                f"{key}_depth",
+                f"{key}-depth",
+                f"{key}.depth",
+                key,
+            ]
+        )
+    candidate_stems.extend(["depth", "gt_depth", "depth_gt"])
+
+    seen = set()
+    ordered_candidate_stems = []
+    for item in candidate_stems:
+        norm = item.strip("._-")
+        if not norm or norm in seen:
+            continue
+        seen.add(norm)
+        ordered_candidate_stems.append(norm)
+
+    for candidate_stem in ordered_candidate_stems:
+        for ext in EXAMPLE_DEPTH_EXTENSIONS:
+            candidate = parent / f"{candidate_stem}{ext}"
+            if candidate.is_file() and candidate.resolve() != rgb_path.resolve():
+                return candidate
+
+    fallback = [
+        path
+        for path in sorted(parent.iterdir())
+        if path.is_file()
+        and path.suffix.lower() in EXAMPLE_DEPTH_EXTENSIONS
+        and path.resolve() != rgb_path.resolve()
+        and "depth" in path.stem.lower()
+    ]
+    if len(fallback) == 1:
+        return fallback[0]
+    return None
+
+
+def _collect_example_samples(limit: int = 24) -> list[dict[str, Optional[str]]]:
+    if not EXAMPLE_DATA_ROOT.exists():
+        return []
+
+    rows: list[dict[str, Optional[str]]] = []
+    for path in sorted(EXAMPLE_DATA_ROOT.rglob("*")):
+        if not path.is_file():
+            continue
+        if path.suffix.lower() not in EXAMPLE_IMAGE_EXTENSIONS:
+            continue
+        if "depth" in path.stem.lower():
+            continue
+
+        depth_path = _find_paired_depth(path)
+        rows.append(
+            {
+                "rgb": path.as_posix(),
+                "depth": depth_path.as_posix() if depth_path else None,
+            }
+        )
+        if len(rows) >= int(limit):
+            break
+    return rows
+
+
+def _build_examples_rows(samples: list[dict[str, Optional[str]]]) -> list[list[Optional[str]]]:
+    rows: list[list[Optional[str]]] = []
+    for sample in samples:
+        rgb_path = sample.get("rgb")
+        if not rgb_path:
+            continue
+        rows.append([rgb_path, sample.get("depth")])
+    return rows
+
+
 def _prepare_output_dir(request: Optional[gr.Request]) -> Path:
     session_hash = "local"
     if request is not None and getattr(request, "session_hash", None):
@@ -194,11 +294,11 @@ DESCRIPTION_MD = """
 """
 
 
-with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS) as demo:
+with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS, fill_width=True) as demo:
     gr.Markdown(DESCRIPTION_MD)
 
-    with gr.Row():
-        with gr.Column(scale=5):
+    with gr.Row(elem_id="main-layout"):
+        with gr.Column(elem_id="left-panel"):
             image_input = gr.Image(type="numpy", label="Input RGB Image")
             depth_input = gr.File(
                 label="Optional Depth Map (.png/.npy/.npz/.h5/.hdf5/.exr)",
@@ -227,11 +327,12 @@ with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS)
                     value=1,
                     step=1,
                     label="Super-resolution Ratio",
+                    visible=True,
                 )
                 max_points_preview = gr.Slider(
                     minimum=10000,
                     maximum=1000000,
-                    value=60000,
+                    value=500000,
                     step=5000,
                     label="Max Preview Points",
                 )
@@ -248,14 +349,14 @@ with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS)
                 "Use lower preview points for faster 3D interaction."
             )
 
-        with gr.Column(scale=7):
+        with gr.Column(elem_id="right-panel"):
             with gr.Tabs():
                 with gr.Tab("3D View"):
                     pcd_viewer = gr.Model3D(
                         label="Point Cloud Viewer",
                         display_mode="solid",
                         clear_color=[1, 1, 1, 1],
-                        height=560,
+                        height="60vh",
                     )
                 with gr.Tab("Depth"):
                     depth_output = gr.Image(type="numpy", label="Predicted Depth (Colorized)")
@@ -267,6 +368,7 @@ with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS)
                     )
             status = gr.Textbox(label="Status", interactive=False)
 
+
     run_button.click(
         fn=lambda: (None, None, [], "Running..."),
         outputs=[depth_output, pcd_viewer, files_output, status],
@@ -288,6 +390,21 @@ with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS)
         outputs=[depth_output, pcd_viewer, files_output, status],
     )
 
+    example_samples = _collect_example_samples()
+    if example_samples:
+        example_rows = _build_examples_rows(example_samples)
+        gr.Markdown("### Example Data")
+        gr.Markdown("Use template-style examples. RGB is always loaded; paired depth is loaded when available.")
+        gr.Examples(
+            examples=example_rows,
+            inputs=[image_input, depth_input],
+            label="Example Data",
+            cache_examples=False,
+            examples_per_page=10,
+        )
+    else:
+        gr.Markdown("### Example Data\nNo images found in `example_data/`.")
+
 demo = demo.queue()
 
 

requirements.txt

@@ -4,7 +4,7 @@
 torch==2.9.1
 torchvision==0.24.1
 torchaudio==2.9.1
-xformers==0.0.33.post1 
+xformers==0.0.33.post2
 hydra-colorlog
 hydra-core
 h5py