FaceAlignmentCropper

Issue / Note: PyTorch 2.6+ Security Blocks Weights Loading in FaceAlignmentCropper
Problem Description:
When running a workflow with the LivePortrait Load FaceAlignmentCropper node, it fails to initialize and turns red on the canvas (entering an UNKNOWN state or throwing an initialization error). The ComfyUI console outputs a critical PyTorch serialization error.

Error Log:

Plaintext
torch.serialization.WeightsOnlyLoadFailed: Weights only load failed. Re-running `torch.load` with `weights_only=False` to show an exception that happened instead has failed.
AttributeError: Can't get attribute 'reduce_graph_module' on

Root Cause:
Starting with PyTorch 2.6+, the default model loading mechanism strictly enforces security restrictions (weights_only=True). This blocks the execution of any hidden or non-standard code embedded within model weight files. The checkpoint used by the FaceAlignment backend was saved using an outdated method and contains a reference to an internal structure called reduce_graph_module. The updated PyTorch environment flags this structure as unsafe and completely blocks its import.

Solution / Workaround:
Switch to the LivePortrait Load InsightFaceCropper node. It utilizes modern models in the .onnx format, which run via a separate ONNX Runtime engine. This completely bypasses PyTorch's pickle-based security checks, prevents initialization errors, and provides significantly more precise facial tracking utilizing GPU acceleration (CUDA) RTX 4080.

The issue is in the model loading mechanism for PyTorch 2.6+. Examine the checkpoint loading code in the face-alignment library, likely around where torch.load is called. The error references 'reduce_graph_module'. Check if the model checkpoint format needs updating or if weights_only must be set to False. Testing requires a PyTorch 2.6+ environment and the failing checkpoint.

Vom Indexierungsmodell aus dem Issue-Text verfasst.