Preprocess without torchvision (ZeroGPU base lacks it); show_error=True

app.py

@@ -185,4 +185,4 @@ def build_demo():
 
 
 if __name__ == "__main__":
-    build_demo().queue().launch()
+    build_demo().queue().launch(show_error=True)

pipeline.py

@@ -25,10 +25,8 @@ import torch
 from lingbot_map.models.gct_stream import GCTStream
 from lingbot_map.utils.pose_enc import pose_encoding_to_extri_intri
 # (w2c kept for the GLB path → closed_form_inverse_se3_general intentionally NOT imported)
-# NB: load_and_preprocess_images (utils.load_fn) is the only thing here that pulls
-# torchvision; it is imported lazily inside preprocess_paths() so this module stays
-# importable on boxes where torchvision is unavailable (e.g. a pyenv build missing
-# _lzma) for CPU-only export testing.
+# NB: preprocess_paths() reimplements upstream load_and_preprocess_images WITHOUT torchvision
+# (the ZeroGPU base image ships torch but not torchvision), so nothing here imports torchvision.
 
 # NB: import the exporter from the *submodule*, not ``lingbot_map.vis``. The package
 # __init__ imports the viser-based PointCloudViewer at load time, which would force a
@@ -137,15 +135,59 @@ def preprocess_paths(
     image_size: int = 518,
     patch_size: int = 14,
 ) -> torch.Tensor:
-    """Wrap ``load_and_preprocess_images`` (crop mode → width 518, EXIF-honored).
+    """Preprocess images to the model's canonical input — a **torchvision-free**, faithful
+    reimplementation of upstream ``load_and_preprocess_images`` (crop mode).
+
+    The ZeroGPU base image ships torch but NOT torchvision, and upstream's loader imports
+    ``torchvision.transforms`` only for ``ToTensor``. We replicate the exact logic (EXIF
+    transpose, alpha-on-white, width=518, height→multiple of patch_size, center-crop height,
+    pad mismatched shapes with white) and swap ToTensor for a numpy→torch equivalent.
 
     Returns a tensor of shape [S, 3, H, W] in range [0, 1].
     """
-    from lingbot_map.utils.load_fn import load_and_preprocess_images  # lazy (torchvision)
-
-    return load_and_preprocess_images(
-        list(paths), mode="crop", image_size=image_size, patch_size=patch_size
-    )
+    from PIL import Image, ImageOps
+
+    def to_tensor_rgb(pil_img):
+        # torchvision.transforms.ToTensor() equivalent for an 8-bit RGB image:
+        # HWC uint8 [0,255] -> CHW float32 [0,1]
+        arr = np.array(pil_img, dtype=np.uint8)                   # [H, W, 3] (copy → writable)
+        return torch.from_numpy(arr).permute(2, 0, 1).contiguous().float().div_(255.0)
+
+    target = image_size
+    imgs = []
+    for p in paths:
+        img = Image.open(p)
+        img = ImageOps.exif_transpose(img)                        # honor phone/camera EXIF
+        if img.mode == "RGBA":
+            bg = Image.new("RGBA", img.size, (255, 255, 255, 255))
+            img = Image.alpha_composite(bg, img)
+        img = img.convert("RGB")
+        w, h = img.size
+        new_w = target
+        new_h = round(h * (new_w / w) / patch_size) * patch_size
+        img = img.resize((new_w, new_h), Image.Resampling.BICUBIC)
+        t = to_tensor_rgb(img)                                    # [3, new_h, new_w]
+        if new_h > target:                                        # center-crop height
+            sy = (new_h - target) // 2
+            t = t[:, sy:sy + target, :]
+        imgs.append(t)
+
+    # If frames differ in shape (mixed aspect ratios), pad to the max with white (1.0).
+    shapes = {(t.shape[1], t.shape[2]) for t in imgs}
+    if len(shapes) > 1:
+        max_h = max(s[0] for s in shapes)
+        max_w = max(s[1] for s in shapes)
+        out = []
+        for t in imgs:
+            ph, pw = max_h - t.shape[1], max_w - t.shape[2]
+            if ph > 0 or pw > 0:
+                pt, pb = ph // 2, ph - ph // 2
+                pl, pr = pw // 2, pw - pw // 2
+                t = torch.nn.functional.pad(t, (pl, pr, pt, pb), mode="constant", value=1.0)
+            out.append(t)
+        imgs = out
+
+    return torch.stack(imgs)
 
 
 # =============================================================================