增加资源文件

assets/config_simplified.yml

@@ -0,0 +1,37 @@
+network_module: hymotion/network/hymotion_mmdit.HunyuanMotionMMDiT
+network_module_args:
+  apply_rope_to_single_branch: false
+  ctxt_input_dim: 4096
+  dropout: 0.0
+  feat_dim: 1024
+  input_dim: 201
+  mask_mode: narrowband
+  mlp_ratio: 4.0
+  num_heads: 16
+  num_layers: 18
+  time_factor: 1000.0
+  vtxt_input_dim: 768
+train_pipeline: hymotion/pipeline/motion_diffusion.MotionFlowMatching
+train_pipeline_args:
+  enable_ctxt_null_feat: true
+  enable_special_game_feat: true
+  infer_noise_scheduler_cfg:
+    validation_steps: 50
+  losses_cfg:
+    recons:
+      name: SmoothL1Loss
+      weight: 1.0
+  noise_scheduler_cfg:
+    method: euler
+  output_mesh_fps: 30
+  random_generator_on_gpu: true
+  test_cfg:
+    mean_std_dir: ./stats/
+    text_guidance_scale: 5.0
+  text_encoder_cfg:
+    llm_type: qwen3
+    max_length_llm: 128
+  text_encoder_module: hymotion/network/text_encoders/text_encoder.HYTextModel
+  train_cfg:
+    cond_mask_prob: 0.1
+  train_frames: 360

assets/wooden_models/boy_Rigging_smplx_tex.fbx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e1a4fc5b121d5fa61a631ee22ba360ca128279d794d1ed75b2acb9486e71cc8
+size 16490768

hymotion/pipeline/motion_diffusion.py

@@ -36,7 +36,6 @@ def length_to_mask(lengths: Tensor, max_len: int) -> Tensor:
 
 
 def start_end_frame_to_mask(start_frame: Tensor, end_frame: Tensor, max_len: int) -> Tensor:
-    # 生成一个 (B, max_len) 的mask，只有在[start_frame, end_frame]区间内为True，其余为False
     assert (start_frame >= 0).all() and (end_frame >= 0).all(), f"start_frame={start_frame}, end_frame={end_frame}"
     lengths = end_frame - start_frame + 1
     assert lengths.max() <= max_len, f"lengths.max()={lengths.max()} > max_len={max_len}"
@@ -184,6 +183,7 @@ class MotionGeneration(torch.nn.Module):
         if not allow_empty_ckpt:
             if not os.path.exists(ckpt_name):
                 import warnings
+
                 warnings.warn(f"Checkpoint {ckpt_name} not found, skipping model loading")
             else:
                 checkpoint = torch.load(ckpt_name, map_location="cpu", weights_only=False)
@@ -222,30 +222,6 @@ class MotionGeneration(torch.nn.Module):
             should_apply_smooothing=should_apply_smooothing,
         )
 
-    def _forward_smpl_batch(
-        self,
-        root_rot6d: Tensor,  # (B, L, 1, 6)
-        body_rot6d: Tensor,  # (B, L, 21, 6)
-        transl: Tensor,  # (B, L, 3)
-        left_hand_pose: Optional[Tensor] = None,  # (B, L, 15, 6)
-        right_hand_pose: Optional[Tensor] = None,  # (B, L, 16, 6)
-    ) -> Tensor:
-        device = transl.device
-        bsz, L = transl.shape[:2]
-        k3d_all = []
-        tmp_betas = torch.zeros(1, 16, device=device)
-        for bs in range(bsz):
-            out = self.body_model(
-                body_rot6d[bs],
-                tmp_betas,
-                root_rot6d[bs],
-                transl[bs],
-                left_hand_pose=(left_hand_pose[bs] if left_hand_pose is not None else None),
-                right_hand_pose=(right_hand_pose[bs] if right_hand_pose is not None else None),
-            )
-            k3d_all.append(out.detach().cpu())
-        return torch.stack(k3d_all, dim=0)  # (B, L, J, 3)
-
     def _decode_o6dp(
         self,
         latent_denorm: torch.Tensor,
@@ -299,32 +275,22 @@ class MotionGeneration(torch.nn.Module):
             transl_smooth = transl_fixed
 
         if self.body_model is not None:
-            print(f'{self.__class__.__name__} rot6d_smooth shape: {rot6d_smooth.shape}, transl_smooth shape: {transl_smooth.shape}')
+            print(
+                f"{self.__class__.__name__} rot6d_smooth shape: {rot6d_smooth.shape}, transl_smooth shape: {transl_smooth.shape}"
+            )
             with torch.no_grad():
                 vertices_all = []
                 k3d_all = []
                 for bs in range(rot6d_smooth.shape[0]):
-                    out = self.body_model.forward(
-                        {
-                            'rot6d': rot6d_smooth[bs],
-                            'trans': transl_smooth[bs],
-                        }
-                    )
+                    out = self.body_model.forward({"rot6d": rot6d_smooth[bs], "trans": transl_smooth[bs]})
                     vertices_all.append(out["vertices"])
-                    k3d_all.append(out['keypoints3d'])
+                    k3d_all.append(out["keypoints3d"])
                 vertices = torch.stack(vertices_all, dim=0)
                 k3d = torch.stack(k3d_all, dim=0)
-            print(f'{self.__class__.__name__} vertices shape: {vertices.shape}, k3d shape: {k3d.shape}')
-            # k3d = self._forward_smpl_batch(
-            #     rot6d_smooth[:, :, 0:1, :].to(device),
-            #     rot6d_smooth[:, :, 1:22, :].to(device),
-            #     transl_smooth,
-            #     left_hand_pose=(rot6d_smooth[:, :, 22:37, :].to(device) if left_hand_pose is not None else None),
-            #     right_hand_pose=(rot6d_smooth[:, :, 37:52, :].to(device) if right_hand_pose is not None else None),
-            # )
+            print(f"{self.__class__.__name__} vertices shape: {vertices.shape}, k3d shape: {k3d.shape}")
             # align with the ground
             min_y = vertices[..., 1].amin(dim=(1, 2), keepdim=True)  # (B, 1, 1)
-            print(f'{self.__class__.__name__} min_y: {min_y}')
+            print(f"{self.__class__.__name__} min_y: {min_y}")
             k3d = k3d.clone()
             k3d[..., 1] -= min_y  # (B, L, J) - (B, 1, 1)
             transl_smooth = transl_smooth.clone()