Fixed the Unexpected floating ScalarType issue that occurred when used with wanvideowrapper #6

README.md

@@ -255,8 +255,9 @@ Interactive visual editor for creating point and bounding box prompts on images/
 
 ## Model Downloads
 
-Download SAM3 model weights from the official repository:
+Download SAM3 model weights from the repository:
 - [SAM3 Models](https://huggingface.co/facebook/sam3)
+- [SAM3 FP16 Model](https://huggingface.co/yolain/sam3-safetensors/blob/main/sam3-fp16.safetensors)
 
 Place the downloaded models in: `ComfyUI/models/sam3/`
 
@@ -291,6 +292,14 @@ Contributions are welcome! Please feel free to submit issues or pull requests.
 
 ## Changelog
 
+### v1.0.4
+
+- Fixed the `Unexpected floating ScalarType` error that occurred when used with wanvideowrapper.
+
+### v1.0.3
+
+- Changed `sam3GetObjectMask` node's `obj_id` to object index ID for easier correspondence with index IDs in visualization
+
 ### v1.0.2
 
 - Added `easy framesEditor` node for interactive image/video frame annotation

README_CN.md

@@ -255,8 +255,9 @@
 
 ## 模型下载
 
-从官方仓库下载 SAM3 模型权重：
+从仓库下载 SAM3 模型权重：
 - [SAM3 模型](https://huggingface.co/facebook/sam3)
+- [SAM3 FP16 模型](https://huggingface.co/yolain/sam3-safetensors/blob/main/sam3-fp16.safetensors)
 
 将下载的模型放置在：`ComfyUI/models/sam3/`
 
@@ -291,6 +292,14 @@
 
 ## 更新日志
 
+### v1.0.4
+
+- 修复和 `wanvideowrapper` 节点同时使用时存在 `Unexpected floating ScalarType` 报错的情况
+
+### v1.0.3
+
+- 将 `sam3GetObjectMask` 节点的 `obj_id` 重新定义为对象索引ID，方便与可视化中的对象索引ID对应
+
 ### v1.0.2
 
 - 增加 `easy framesEditor` 节点以进行交互式图像/视频帧标注

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-easy-sam3"
 description = "A ComfyUI custom node package for segment anything 3"
-version = "1.0.3"
+version = "1.0.4"
 license = {file = "LICENSE"}
 dependencies = ["timm>=1.0.17", "ftfy==6.1.1", "regex", "iopath>=0.1.10", "einops>=0.6.0", "decord>=0.6.0", "pycocotools>=2.0.10", "numpy>=1.26", "tqdm", "typing_extensions"]
 

sam3/model/decoder.py

@@ -71,8 +71,7 @@ class TransformerDecoderLayer(nn.Module):
         return tensor if pos is None else tensor + pos
 
     def forward_ffn(self, tgt):
-        with torch.amp.autocast(device_type="cuda", enabled=False):
-            tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
         tgt = tgt + self.dropout4(tgt2)
         tgt = self.norm3(tgt)
         return tgt

sam3/model/sam3_tracking_predictor.py

@@ -46,8 +46,7 @@ class Sam3TrackerPredictor(Sam3TrackerBase):
         self.max_point_num_in_prompt_enc = max_point_num_in_prompt_enc
         self.non_overlap_masks_for_output = non_overlap_masks_for_output
 
-        self.bf16_context = torch.autocast(device_type="cuda", dtype=torch.bfloat16)
-        self.bf16_context.__enter__()  # keep using for the entire model process
+        self.bf16_context = None
 
         self.iter_use_prev_mask_pred = True
         self.add_all_frames_to_correct_as_cond = True