update

embodied_gen/models/layout.py

@@ -43,7 +43,7 @@ __all__ = [
 ]
 
 
-DISTRACTOR_NUM = 3  # Maximum number of distractor objects allowed
+DISTRACTOR_NUM = 2  # Maximum number of distractor objects allowed
 LAYOUT_DISASSEMBLE_PROMPT = f"""
     You are an intelligent 3D scene planner. Given a natural language
     description of a robotic task, output a structured description of

embodied_gen/scripts/imageto3d.py

@@ -219,9 +219,10 @@ def entrypoint(**kwargs):
                     num_images=12,
                 )
 
-                geo_flag, geo_result = GEO_CHECKER(
-                    [color_path], text=asset_node
-                )
+                color_img = Image.open(color_path)
+                half_height = int(color_img.height * 2 / 3)
+                crop_img = color_img.crop((0, 0, color_img.width, half_height))
+                geo_flag, geo_result = GEO_CHECKER([crop_img], text=asset_node)
                 logger.warning(
                     f"{GEO_CHECKER.__class__.__name__}: {geo_result} for {seg_path}"
                 )

embodied_gen/scripts/render_gs.py

@@ -104,7 +104,7 @@ def load_gs_model(
     # Normalize vertices to [-1, 1], center to (0, 0, 0).
     _, scale, center = normalize_vertices_array(gs_model._means)
     scale, center = float(scale), center.tolist()
-    transpose = [*[-v for v in center], *pre_quat]
+    transpose = [*[v for v in center], *pre_quat]
     instance_pose = torch.tensor(transpose).to(gs_model.device)
     gs_model = gs_model.get_gaussians(instance_pose=instance_pose)
     gs_model.rescale(scale)

embodied_gen/validators/quality_checkers.py

@@ -249,7 +249,7 @@ class SemanticConsistChecker(BaseChecker):
                 fewer than four legs or if the legs are unevenly distributed, are not allowed. Do not assume
                 hidden legs unless they are clearly visible.)
             - Geometric completeness is required: the object must not have missing, truncated, or cropped parts.
-            - The image must contain exactly one object. Multiple distinct objects are not allowed.
+            - The image must contain exactly one object. Multiple distinct objects (e.g. multiple pens) are not allowed.
                 A single composite object (e.g., a chair with legs) is acceptable.
             - The object should be shown from a slightly angled (three-quarter) perspective,
                 not a flat, front-facing view showing only one surface.