Fix hunyuan3d 2.1

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/custom_rasterizer/lib/custom_rasterizer_kernel/grid_neighbor.cpp

@@ -394,36 +394,36 @@ std::vector<std::vector<torch::Tensor>> build_hierarchy(std::vector<torch::Tenso
     std::vector<torch::Tensor> grid_evencorners(grids.size());
     std::vector<torch::Tensor> grid_oddcorners(grids.size());
 
-    texture_positions[0] = torch::zeros({seq2pos.size() / 3, 3}, float_options);
-    texture_positions[1] = torch::zeros({seq2pos.size() / 3}, float_options);
+    texture_positions[0] = torch::zeros({(int64_t)(seq2pos.size() / 3), 3}, float_options);
+    texture_positions[1] = torch::zeros({(int64_t)(seq2pos.size() / 3)}, float_options);
     float* positions_out_ptr = texture_positions[0].data_ptr<float>();
     memcpy(positions_out_ptr, seq2pos.data(), sizeof(float) * seq2pos.size());
     positions_out_ptr = texture_positions[1].data_ptr<float>();
-    for (int i = 0; i < grids[0].seq2grid.size(); ++i) {
+    for (int i = 0; i < (int)grids[0].seq2grid.size(); ++i) {
         positions_out_ptr[i] = (i < grids[0].num_origin_seq);
     }
 
-    for (int i = 0; i < grids.size(); ++i) {
-        grid_neighbors[i] = torch::zeros({grids[i].seq2grid.size(), 9}, int64_options);
-        long* nptr = grid_neighbors[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2neighbor.size(); ++j) {
+    for (int i = 0; i < (int)grids.size(); ++i) {
+        grid_neighbors[i] = torch::zeros({(int64_t)grids[i].seq2grid.size(), 9}, int64_options);
+        int64_t* nptr = grid_neighbors[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2neighbor.size(); ++j) {
             nptr[j] = grids[i].seq2neighbor[j];
         }
 
-        grid_evencorners[i] = torch::zeros({grids[i].seq2evencorner.size()}, int64_options);
-        grid_oddcorners[i] = torch::zeros({grids[i].seq2oddcorner.size()}, int64_options);
-        long* dptr = grid_evencorners[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2evencorner.size(); ++j) {
+        grid_evencorners[i] = torch::zeros({(int64_t)grids[i].seq2evencorner.size()}, int64_options);
+        grid_oddcorners[i] = torch::zeros({(int64_t)grids[i].seq2oddcorner.size()}, int64_options);
+        int64_t* dptr = grid_evencorners[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2evencorner.size(); ++j) {
             dptr[j] = grids[i].seq2evencorner[j];
         }
-        dptr = grid_oddcorners[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2oddcorner.size(); ++j) {
+        dptr = grid_oddcorners[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2oddcorner.size(); ++j) {
             dptr[j] = grids[i].seq2oddcorner[j];
         }            
-        if (i + 1 < grids.size()) {
-            grid_downsamples[i] = torch::zeros({grids[i].downsample_seq.size()}, int64_options);
-            long* dptr = grid_downsamples[i].data_ptr<long>();
-            for (int j = 0; j < grids[i].downsample_seq.size(); ++j) {
+        if (i + 1 < (int)grids.size()) {
+            grid_downsamples[i] = torch::zeros({(int64_t)grids[i].downsample_seq.size()}, int64_options);
+            int64_t* dptr = grid_downsamples[i].data_ptr<int64_t>();
+            for (int j = 0; j < (int)grids[i].downsample_seq.size(); ++j) {
                 dptr[j] = grids[i].downsample_seq[j];
             }
         }
@@ -534,38 +534,38 @@ std::vector<std::vector<torch::Tensor>> build_hierarchy_with_feat(
     std::vector<torch::Tensor> grid_evencorners(grids.size());
     std::vector<torch::Tensor> grid_oddcorners(grids.size());
 
-    texture_positions[0] = torch::zeros({seq2pos.size() / 3, 3}, float_options);
-    texture_positions[1] = torch::zeros({seq2pos.size() / 3}, float_options);
-    texture_feats[0] = torch::zeros({seq2feat.size() / feat_channel, feat_channel}, float_options);
+    texture_positions[0] = torch::zeros({(int64_t)(seq2pos.size() / 3), 3}, float_options);
+    texture_positions[1] = torch::zeros({(int64_t)(seq2pos.size() / 3)}, float_options);
+    texture_feats[0] = torch::zeros({(int64_t)(seq2feat.size() / feat_channel), (int64_t)feat_channel}, float_options);
     float* positions_out_ptr = texture_positions[0].data_ptr<float>();
     memcpy(positions_out_ptr, seq2pos.data(), sizeof(float) * seq2pos.size());
     positions_out_ptr = texture_positions[1].data_ptr<float>();
-    for (int i = 0; i < grids[0].seq2grid.size(); ++i) {
+    for (int i = 0; i < (int)grids[0].seq2grid.size(); ++i) {
         positions_out_ptr[i] = (i < grids[0].num_origin_seq);
     }
     float* feats_out_ptr = texture_feats[0].data_ptr<float>();
     memcpy(feats_out_ptr, seq2feat.data(), sizeof(float) * seq2feat.size());
 
-    for (int i = 0; i < grids.size(); ++i) {
-        grid_neighbors[i] = torch::zeros({grids[i].seq2grid.size(), 9}, int64_options);
-        long* nptr = grid_neighbors[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2neighbor.size(); ++j) {
+    for (int i = 0; i < (int)grids.size(); ++i) {
+        grid_neighbors[i] = torch::zeros({(int64_t)grids[i].seq2grid.size(), 9}, int64_options);
+        int64_t* nptr = grid_neighbors[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2neighbor.size(); ++j) {
             nptr[j] = grids[i].seq2neighbor[j];
         }
-        grid_evencorners[i] = torch::zeros({grids[i].seq2evencorner.size()}, int64_options);
-        grid_oddcorners[i] = torch::zeros({grids[i].seq2oddcorner.size()}, int64_options);
-        long* dptr = grid_evencorners[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2evencorner.size(); ++j) {
+        grid_evencorners[i] = torch::zeros({(int64_t)grids[i].seq2evencorner.size()}, int64_options);
+        grid_oddcorners[i] = torch::zeros({(int64_t)grids[i].seq2oddcorner.size()}, int64_options);
+        int64_t* dptr = grid_evencorners[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2evencorner.size(); ++j) {
             dptr[j] = grids[i].seq2evencorner[j];
         }
-        dptr = grid_oddcorners[i].data_ptr<long>();
-        for (int j = 0; j < grids[i].seq2oddcorner.size(); ++j) {
+        dptr = grid_oddcorners[i].data_ptr<int64_t>();
+        for (int j = 0; j < (int)grids[i].seq2oddcorner.size(); ++j) {
             dptr[j] = grids[i].seq2oddcorner[j];
         }
-        if (i + 1 < grids.size()) {
-            grid_downsamples[i] = torch::zeros({grids[i].downsample_seq.size()}, int64_options);
-            long* dptr = grid_downsamples[i].data_ptr<long>();
-            for (int j = 0; j < grids[i].downsample_seq.size(); ++j) {
+        if (i + 1 < (int)grids.size()) {
+            grid_downsamples[i] = torch::zeros({(int64_t)grids[i].downsample_seq.size()}, int64_options);
+            int64_t* dptr = grid_downsamples[i].data_ptr<int64_t>();
+            for (int j = 0; j < (int)grids[i].downsample_seq.size(); ++j) {
                 dptr[j] = grids[i].downsample_seq[j];
             }
         }

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/custom_rasterizer/lib/custom_rasterizer_kernel/rasterizer.cpp

@@ -100,24 +100,24 @@ std::vector<torch::Tensor> rasterize_image_cpu(torch::Tensor V, torch::Tensor F,
     auto INT64_options = torch::TensorOptions().dtype(torch::kInt64).requires_grad(false);
     auto findices = torch::zeros({height, width}, options);
     INT64 maxint = (INT64)MAXINT * (INT64)MAXINT + (MAXINT - 1);
-    auto z_min = torch::ones({height, width}, INT64_options) * (long)maxint;
+    auto z_min = torch::ones({height, width}, INT64_options) * (int64_t)maxint;
 
     if (!use_depth_prior) {
         for (int i = 0; i < num_faces; ++i) {
             rasterizeImagecoordsKernelCPU(V.data_ptr<float>(), F.data_ptr<int>(), 0,
-                (INT64*)z_min.data_ptr<long>(), occlusion_truncation, width, height, num_vertices, num_faces, i); 
+                (INT64*)z_min.data_ptr<int64_t>(), occlusion_truncation, width, height, num_vertices, num_faces, i); 
         }
     } else {
         for (int i = 0; i < num_faces; ++i)
             rasterizeImagecoordsKernelCPU(V.data_ptr<float>(), F.data_ptr<int>(), D.data_ptr<float>(),
-                (INT64*)z_min.data_ptr<long>(), occlusion_truncation, width, height, num_vertices, num_faces, i);
+                (INT64*)z_min.data_ptr<int64_t>(), occlusion_truncation, width, height, num_vertices, num_faces, i);
     }
 
     auto float_options = torch::TensorOptions().dtype(torch::kFloat32).requires_grad(false);
     auto barycentric = torch::zeros({height, width, 3}, float_options);
     for (int i = 0; i < width * height; ++i)
         barycentricFromImgcoordCPU(V.data_ptr<float>(), F.data_ptr<int>(),
-            findices.data_ptr<int>(), (INT64*)z_min.data_ptr<long>(), width, height, num_vertices, num_faces, barycentric.data_ptr<float>(), i);
+            findices.data_ptr<int>(), (INT64*)z_min.data_ptr<int64_t>(), width, height, num_vertices, num_faces, barycentric.data_ptr<float>(), i);
 
     return {findices, barycentric};
 }

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/custom_rasterizer/lib/custom_rasterizer_kernel/rasterizer_gpu.cu

@@ -108,20 +108,20 @@ std::vector<torch::Tensor> rasterize_image_gpu(torch::Tensor V, torch::Tensor F,
     auto INT64_options = torch::TensorOptions().dtype(torch::kInt64).device(torch::kCUDA, device_id).requires_grad(false);
     auto findices = torch::zeros({height, width}, options);
     INT64 maxint = (INT64)MAXINT * (INT64)MAXINT + (MAXINT - 1);
-    auto z_min = torch::ones({height, width}, INT64_options) * (long)maxint;
+    auto z_min = torch::ones({height, width}, INT64_options) * (int64_t)maxint;
 
     if (!use_depth_prior) {
         rasterizeImagecoordsKernelGPU<<<(num_faces+255)/256,256,0,at::cuda::getCurrentCUDAStream()>>>(V.data_ptr<float>(), F.data_ptr<int>(), 0,
-            (INT64*)z_min.data_ptr<long>(), occlusion_truncation, width, height, num_vertices, num_faces); 
+            (INT64*)z_min.data_ptr<int64_t>(), occlusion_truncation, width, height, num_vertices, num_faces); 
     } else {
         rasterizeImagecoordsKernelGPU<<<(num_faces+255)/256,256,0,at::cuda::getCurrentCUDAStream()>>>(V.data_ptr<float>(), F.data_ptr<int>(), D.data_ptr<float>(),
-            (INT64*)z_min.data_ptr<long>(), occlusion_truncation, width, height, num_vertices, num_faces); 
+            (INT64*)z_min.data_ptr<int64_t>(), occlusion_truncation, width, height, num_vertices, num_faces); 
     }
 
     auto float_options = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA, device_id).requires_grad(false);
     auto barycentric = torch::zeros({height, width, 3}, float_options);
     barycentricFromImgcoordGPU<<<(width * height + 255)/256, 256>>>(V.data_ptr<float>(), F.data_ptr<int>(),
-        findices.data_ptr<int>(), (INT64*)z_min.data_ptr<long>(), width, height, num_vertices, num_faces, barycentric.data_ptr<float>());
+        findices.data_ptr<int>(), (INT64*)z_min.data_ptr<int64_t>(), width, height, num_vertices, num_faces, barycentric.data_ptr<float>());
 
     return {findices, barycentric};
 }

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/hunyuanpaintpbr/unet/__init__.py

@@ -0,0 +1 @@
+# init

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/hunyuanpaintpbr/unet/modules.py

@@ -48,8 +48,10 @@ class Dino_v2(nn.Module):
 
     def __init__(self, dino_v2_path):
         super(Dino_v2, self).__init__()
+        print(f"[Dino_v2] Loading DINOv2 from: {dino_v2_path}...")
         self.dino_processor = AutoImageProcessor.from_pretrained(dino_v2_path)
         self.dino_v2 = AutoModel.from_pretrained(dino_v2_path)
+        print(f"[Dino_v2] DINOv2 loaded successfully")
 
         for param in self.parameters():
             param.requires_grad = False

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/textureGenPipeline.py

@@ -39,14 +39,14 @@
 
 
 class Hunyuan3DPaintConfig:
-    def __init__(self, max_num_view, resolution):
+    def __init__(self, max_num_view, resolution, multiview_pretrained_path=None, dino_ckpt_path=None, realesrgan_ckpt_path=None):
         self.device = "cuda"
 
         self.multiview_cfg_path = "cfgs/hunyuan-paint-pbr.yaml"
         self.custom_pipeline = "hunyuanpaintpbr"
-        self.multiview_pretrained_path = "tencent/Hunyuan3D-2.1"
-        self.dino_ckpt_path = "facebook/dinov2-giant"
-        self.realesrgan_ckpt_path = "/home/ubuntu/oldrigger/AutoRigger/auto_rigger/models/ComfyUI/custom_nodes/ComfyUI-3D-Pack/Checkpoints/Diffusers/tencent/Hunyuan3D-2.1/RealESRGAN_x4plus.pth"
+        self.multiview_pretrained_path = multiview_pretrained_path or "tencent/Hunyuan3D-2.1"
+        self.dino_ckpt_path = dino_ckpt_path or "facebook/dinov2-giant"
+        self.realesrgan_ckpt_path = realesrgan_ckpt_path or "RealESRGAN_x4plus.pth"
 
         self.raster_mode = "cr"
         self.bake_mode = "back_sample"

Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/utils/multiview_utils.py

@@ -34,17 +34,26 @@ def __init__(self, config) -> None:
         self.cfg = cfg
         self.mode = self.cfg.model.params.stable_diffusion_config.custom_pipeline[2:]
 
-        model_path = huggingface_hub.snapshot_download(
-            repo_id=config.multiview_pretrained_path,
-            allow_patterns=["hunyuan3d-paintpbr-v2-1/*"],
-        )
+        if os.path.isdir(config.multiview_pretrained_path):
+            print(f"[multiviewDiffusionNet] Using local model path: {config.multiview_pretrained_path}")
+            model_path = config.multiview_pretrained_path
+        else:
+            print(f"[multiviewDiffusionNet] Checking/Downloading model from HF: {config.multiview_pretrained_path}...")
+            model_path = huggingface_hub.snapshot_download(
+                repo_id=config.multiview_pretrained_path,
+                allow_patterns=["hunyuan3d-paintpbr-v2-1/*"],
+            )
+        print(f"[multiviewDiffusionNet] model_path: {model_path}")
 
         model_path = os.path.join(model_path, "hunyuan3d-paintpbr-v2-1")
+        print(f"[multiviewDiffusionNet] Loading DiffusionPipeline from {model_path}...")
+        print(f"[multiviewDiffusionNet] Custom pipeline: {custom_pipeline}")
         pipeline = DiffusionPipeline.from_pretrained(
             model_path,
             custom_pipeline=custom_pipeline, 
             torch_dtype=torch.float16
         )
+        print(f"[multiviewDiffusionNet] DiffusionPipeline loaded.")
 
         pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing")
         pipeline.set_progress_bar_config(disable=True)

nodes.py

@@ -5300,6 +5300,7 @@ def INPUT_TYPES(cls):
 
     @staticmethod 
     def _ensure_weights():
+        # 1. Main Model (Hunyuan3D-2.1)
         repo_id = f"{Load_Hunyuan3D_21_TexGen_Pipeline._REPO_ID_BASE}/{Load_Hunyuan3D_21_TexGen_Pipeline._REPO_NAME}"
         safe_repo_name = Load_Hunyuan3D_21_TexGen_Pipeline._REPO_NAME.replace(".", "_")
         base_dir = os.path.join(CKPT_DIFFUSERS_PATH, f"{Load_Hunyuan3D_21_TexGen_Pipeline._REPO_ID_BASE}/{safe_repo_name}")
@@ -5334,7 +5335,21 @@ def _ensure_weights():
         else:
             print(f"Hunyuan3D-2.1 TexGen weights already loaded")
 
-        return base_dir
+        # 2. DINOv2
+        dino_repo = "facebook/dinov2-giant"
+        dino_dir = os.path.join(CKPT_ROOT_PATH, "facebookresearch/dinov2-giant")
+        if not os.path.exists(os.path.join(dino_dir, "config.json")):
+            print(f"Loading DINOv2 folder: {dino_repo}")
+            snapshot_download(
+                repo_id=dino_repo,
+                local_dir=dino_dir,
+                resume_download=True,
+                ignore_patterns=HF_DOWNLOAD_IGNORE
+            )
+        else:
+            print(f"DINOv2 weights already loaded")
+
+        return base_dir, dino_dir
 
     @staticmethod
     def _ensure_realesrgan():
@@ -5368,12 +5383,17 @@ def load(self, max_num_view, resolution, enable_mmgp):
             print(f"[TexGen-Loader] Using cached pipeline {cache_key}")
             return (self._cache[cache_key],)
 
-        base_dir = self._ensure_weights()
+        base_dir, dino_dir = self._ensure_weights()
         realesrgan_path = self._ensure_realesrgan()
 
         # Configure pipeline
-        conf = Hunyuan3DPaintConfig_2_1(max_num_view=max_num_view, resolution=resolution)
-        conf.realesrgan_ckpt_path = realesrgan_path
+        conf = Hunyuan3DPaintConfig_2_1(
+            max_num_view=max_num_view, 
+            resolution=resolution,
+            multiview_pretrained_path=base_dir,
+            dino_ckpt_path=dino_dir,
+            realesrgan_ckpt_path=realesrgan_path
+        )
         conf.multiview_cfg_path = os.path.join(ROOT_PATH, "Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/cfgs/hunyuan-paint-pbr.yaml")
         conf.custom_pipeline = os.path.join(ROOT_PATH, "Gen_3D_Modules/Hunyuan3D_2_1/hy3dpaint/hunyuanpaintpbr")