diff --git a/comfy/model_detection.py b/comfy/model_detection.py
index 0f7750f29..a9c359226 100644
--- a/comfy/model_detection.py
+++ b/comfy/model_detection.py
@@ -668,7 +668,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["text_model_embed_dim"] = state_dict['{}text_proj.weight'.format(key_prefix)].shape[1]
         dit_config["use_bbox"] = '{}bbox_proj.weight'.format(key_prefix) in state_dict_keys
         dit_config["bias"] = '{}text_proj.bias'.format(key_prefix) in state_dict_keys
-        dit_config["rope_theta"] = 10000
+        dit_config["rope_theta"] = 10000  # not stored in the state dict; upstream's fixed constant
         return dit_config
 
     if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
diff --git a/comfy/sd.py b/comfy/sd.py
index 74e388553..4ab0d34bd 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -490,7 +490,6 @@ class VAE:
         self.disable_offload = False
         self.not_video = False
         self.size = None
-        self.cube3d = False
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
@@ -781,7 +780,6 @@ class VAE:
 
             # Roblox Cube3D shape tokenizer (OneDAutoEncoder, decode-only)
             elif "bottleneck.block.codebook.weight" in sd:
-                self.cube3d = True
                 self.latent_dim = 1
                 # The VQ bottleneck (get_codebook/lookup_codebook) reads raw parameters
                 # outside any hooked forward, so the streaming-offload cast hooks can't
@@ -809,6 +807,9 @@ class VAE:
                 self.process_input = lambda image: image
                 # shape is the token-ID latent (B, 1, num_tokens); size by num_tokens.
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[-1] * 768) * model_management.dtype_size(dtype)
+                # fp32-only (unlike most VAEs that allow fp16/bf16): the VQ codebook lookup
+                # and occupancy-grid query must run in fp32 to match upstream and keep the
+                # isosurface stable.
                 self.working_dtypes = [torch.float32]
 
             elif "vocoder.backbone.channel_layers.0.0.bias" in sd: #Ace Step Audio
diff --git a/comfy/supported_models.py b/comfy/supported_models.py
index 2c9d86328..4cc27de25 100644
--- a/comfy/supported_models.py
+++ b/comfy/supported_models.py
@@ -1572,6 +1572,8 @@ class Cube3D(supported_models_base.BASE):
         return model_base.Cube3D(self, device=device)
 
     def clip_target(self, state_dict={}):
+        # No bundled text encoder: the cube checkpoint is GPT-only. The graph wires a
+        # standard CLIPLoader(clip-l)/CLIPTextEncode, so there is no clip_target to build.
         return None
 
 class TripoSplat(supported_models_base.BASE):