convert to ZeroGPU

README.md

@@ -7,7 +7,6 @@ sdk: gradio
 sdk_version: 6.9.0
 python_version: "3.10"
 app_file: app.py
-hardware: zero-a10g
 preload_from_hub:
   - ritianyu/InfiniDepth infinidepth.ckpt,infinidepth_depthsensor.ckpt,infinidepth_gs.ckpt,infinidepth_depthsensor_gs.ckpt,moge2.pt,skyseg.onnx
 pinned: false
@@ -16,6 +15,3 @@ short_description: InfiniDepth Huggingface Demo
 ---
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
-For Hugging Face Spaces, `preload_from_hub` now preloads the three checkpoint files from
-`ritianyu/InfiniDepth` during build time. Set `INFINIDEPTH_CKPT_REPO=ritianyu/InfiniDepth`
-in the Space variables so runtime loading resolves the same cached files.

app.py

@@ -47,6 +47,22 @@ import numpy as np
 import torch
 from huggingface_hub import hf_hub_download
 
+try:
+    import spaces
+except ImportError:
+    class _SpacesFallback:
+        @staticmethod
+        def GPU(fn=None, **_kwargs):
+            if callable(fn):
+                return fn
+
+            def decorator(inner_fn):
+                return inner_fn
+
+            return decorator
+
+    spaces = _SpacesFallback()
+
 from InfiniDepth.gs import GSPixelAlignPredictor, export_ply
 from InfiniDepth.utils.gs_utils import (
     _build_sparse_uniform_gaussians,
@@ -87,10 +103,14 @@ INPUT_SIZE = (768, 1024)
 APP_NAME = "infinidepth-hf-demo"
 GS_TASK_CHOICE = "3DGS"
 TASK_CHOICES = ["Depth", GS_TASK_CHOICE]
-MODEL_CHOICES = ["InfiniDepth", "InfiniDepth_DepthSensor"]
+RGB_MODEL_TYPE = "InfiniDepth"
+DEPTH_SENSOR_MODEL_TYPE = "InfiniDepth_DepthSensor"
+MODEL_CHOICES = [RGB_MODEL_TYPE, DEPTH_SENSOR_MODEL_TYPE]
 OUTPUT_MODE_CHOICES = ["upsample", "original", "specific"]
 GS_SAMPLE_POINT_NUM = 2000000
 GS_COORD_DETERMINISTIC_SAMPLING = True
+DEPTH_GPU_DURATION_SECONDS = 180
+GS_GPU_DURATION_SECONDS = 240
 
 LOCAL_DEPTH_MODEL_PATHS = {
     "InfiniDepth": APP_ROOT / "checkpoints/depth/infinidepth.ckpt",
@@ -180,7 +200,10 @@ CSS = """
 
 def _ensure_cuda() -> None:
     if not torch.cuda.is_available():
-        raise gr.Error("CUDA GPU is required for InfiniDepth inference in this demo.")
+        raise gr.Error(
+            "No CUDA device is available for this request. On Hugging Face ZeroGPU, "
+            "GPU access is only attached while the decorated inference call is running."
+        )
 
 
 def _resolve_repo_asset(local_path: Path, filename: str) -> str:
@@ -303,6 +326,31 @@ def _format_gs_status(
     )
 
 
+def _model_availability_note(depth_path: Optional[str], model_type: str, *, auto_switched: bool = False) -> str:
+    if depth_path:
+        if auto_switched and model_type == DEPTH_SENSOR_MODEL_TYPE:
+            return (
+                "Depth file loaded. Switched model to `InfiniDepth_DepthSensor`. "
+                "You can still switch back to `InfiniDepth` for RGB-only inference."
+            )
+        return (
+            "Depth file loaded. `InfiniDepth_DepthSensor` is available. "
+            "You can also keep `InfiniDepth` for RGB-only inference."
+        )
+
+    if auto_switched:
+        return (
+            "No input depth loaded. Switched model back to `InfiniDepth`. "
+            "Upload a depth file to enable `InfiniDepth_DepthSensor`."
+        )
+
+    return "No input depth loaded. `InfiniDepth` will be used until you upload a depth file."
+
+
+def _compose_depth_info_message(base_message: str, note: str) -> str:
+    return f"{base_message}\n\n{note}" if note else base_message
+
+
 def _load_example_image(example_name: str) -> tuple[np.ndarray, Optional[str], Optional[np.ndarray], str, str]:
     if not example_name:
         raise gr.Error("Select an example case first.")
@@ -315,13 +363,19 @@ def _load_example_image(example_name: str) -> tuple[np.ndarray, Optional[str], O
     image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
     depth_path = case.depth_path
     preview = None
-    detail = f"Loaded example `{case.name}`."
-    model_type = "InfiniDepth"
+    detail = _compose_depth_info_message(
+        f"Loaded example `{case.name}`.",
+        _model_availability_note(None, RGB_MODEL_TYPE),
+    )
+    model_type = RGB_MODEL_TYPE
 
     if depth_path is not None:
         preview, depth_msg = preview_depth_file(depth_path)
-        detail = f"Loaded example `{case.name}` with paired depth. {depth_msg}"
-        model_type = "InfiniDepth_DepthSensor"
+        model_type = DEPTH_SENSOR_MODEL_TYPE
+        detail = _compose_depth_info_message(
+            f"Loaded example `{case.name}` with paired depth. {depth_msg}",
+            _model_availability_note(depth_path, model_type, auto_switched=True),
+        )
 
     return image_rgb, depth_path, preview, model_type, detail
 
@@ -351,12 +405,34 @@ def _primary_view_for_task(task_type: str):
     return gr.update(selected=selected_tab)
 
 
-def _update_depth_preview(depth_path: Optional[str]) -> tuple[Optional[np.ndarray], str]:
+def _reset_uploaded_image_state(
+    _image: Optional[np.ndarray],
+    depth_path: Optional[str],
+) -> tuple[None, None, str, str]:
+    note = (
+        "Image updated. Cleared the previous depth file. Upload a new depth file to enable "
+        "`InfiniDepth_DepthSensor`."
+        if depth_path
+        else "Image updated. Upload a depth file to enable `InfiniDepth_DepthSensor`."
+    )
+    return None, None, RGB_MODEL_TYPE, note
+
+
+def _update_depth_preview(depth_path: Optional[str], model_type: str) -> tuple[Optional[np.ndarray], str, str]:
     try:
-        return preview_depth_file(depth_path)
+        preview, depth_msg = preview_depth_file(depth_path)
     except Exception as exc:
         raise gr.Error(f"Failed to preview depth file: {exc}") from exc
 
+    if depth_path:
+        next_model = DEPTH_SENSOR_MODEL_TYPE
+        note = _model_availability_note(depth_path, next_model, auto_switched=(model_type != next_model))
+    else:
+        next_model = RGB_MODEL_TYPE
+        note = _model_availability_note(depth_path, next_model, auto_switched=(model_type != next_model))
+
+    return preview, next_model, _compose_depth_info_message(depth_msg, note)
+
 
 def _settings_visibility(task_type: str, output_resolution_mode: str):
     is_depth = task_type == "Depth"
@@ -375,6 +451,7 @@ def _normalize_filtered_gaussians(filtered_result):
     return filtered_result
 
 
+@spaces.GPU(duration=DEPTH_GPU_DURATION_SECONDS)
 @torch.no_grad()
 def _run_depth_inference(
     image: np.ndarray,
@@ -395,7 +472,7 @@ def _run_depth_inference(
     _ensure_cuda()
     if image is None:
         raise gr.Error("Upload an image or load an example before running inference.")
-    if model_type == "InfiniDepth_DepthSensor" and not depth_file:
+    if model_type == DEPTH_SENSOR_MODEL_TYPE and not depth_file:
         raise gr.Error("InfiniDepth_DepthSensor requires an input depth file.")
 
     skyseg_path = _resolve_skyseg_path() if enable_skyseg_model else None
@@ -522,6 +599,7 @@ def _run_depth_inference(
     )
 
 
+@spaces.GPU(duration=GS_GPU_DURATION_SECONDS)
 @torch.no_grad()
 def _run_gs_inference(
     image: np.ndarray,
@@ -537,7 +615,7 @@ def _run_gs_inference(
     _ensure_cuda()
     if image is None:
         raise gr.Error("Upload an image or load an example before running inference.")
-    if model_type == "InfiniDepth_DepthSensor" and not depth_file:
+    if model_type == DEPTH_SENSOR_MODEL_TYPE and not depth_file:
         raise gr.Error("InfiniDepth_DepthSensor requires an input depth file for GS inference.")
 
     image_rgb, image_tensor, (org_h, org_w) = _prepare_image_tensors(image)
@@ -730,7 +808,7 @@ def _clear_outputs():
 with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
     gr.Markdown("# InfiniDepth Demo")
     gr.Markdown(
-        "Switch between depth inference and GS inference, choose `InfiniDepth` or `InfiniDepth_DepthSensor`, preview the generated result directly in the demo, and download exported assets."
+        "Switch between depth inference and GS inference. `InfiniDepth` works with RGB-only inputs, while `InfiniDepth_DepthSensor` is enabled only after you upload a depth file or load an example with paired depth."
     )
 
     selected_example_name = gr.State(DEFAULT_EXAMPLE_NAME)
@@ -738,7 +816,7 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
     with gr.Row(elem_id="top-workspace"):
         with gr.Column(scale=4, min_width=320, elem_id="controls-column"):
             task_type = gr.Radio(label="Inference Task", choices=TASK_CHOICES, value="Depth")
-            model_type = gr.Radio(label="Model Type", choices=MODEL_CHOICES, value="InfiniDepth")
+            model_type = gr.Radio(label="Model Type", choices=MODEL_CHOICES, value=RGB_MODEL_TYPE)
 
             gr.Markdown("### Example Data")
             example_gallery = gr.Gallery(
@@ -793,7 +871,7 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
                 height=240,
                 elem_id="input-depth-preview",
             )
-            depth_info = gr.Markdown("No input depth loaded.")
+            depth_info = gr.Markdown("No input depth loaded. `InfiniDepth` will be used until you upload a depth file.")
             submit_btn = gr.Button("Run Inference", variant="primary")
 
         with gr.Column(scale=8, min_width=640, elem_id="outputs-column"):
@@ -859,10 +937,16 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
         outputs=[selected_example_name, example_selection],
     )
 
+    input_image.input(
+        fn=_reset_uploaded_image_state,
+        inputs=[input_image, input_depth_file],
+        outputs=[input_depth_file, input_depth_preview, model_type, depth_info],
+    )
+
     input_depth_file.change(
         fn=_update_depth_preview,
-        inputs=[input_depth_file],
-        outputs=[input_depth_preview, depth_info],
+        inputs=[input_depth_file, model_type],
+        outputs=[input_depth_preview, model_type, depth_info],
     )
 
     load_example_btn.click(

requirements.txt

@@ -1,6 +1,8 @@
-# Runtime dependencies for the Hugging Face Space demo.
-# This app currently requires Python 3.10/3.11/3.12 because it depends on
-# `open3d` and `numpy==1.26.4`, which do not provide wheels for Python 3.13.
+# Runtime dependencies for the Hugging Face ZeroGPU Space demo.
+# Keep Python on 3.10 in README.md so the pinned torch build and open3d wheels
+# remain compatible with the deployment environment.
+
+spaces
 
 torch==2.9.1
 torchvision==0.24.1