I am encountering significant temporal flickering and flashing artifacts when generating videos in ComfyUI with SageAttention enabled.
The issue manifests as inconsistent details and sudden light changes across frames, which severely breaks the video's temporal consistency. I have confirmed that this issue is directly caused by SageAttention, as the flickering completely disappears when I switch back to the default sdpa attention mode (with all other workflow parameters remaining unchanged).
️ Environment & Setup
OS: [ Windows 11]
GPU: [ NVIDIA RTX 5070Ti]
CUDA Version: [13.0]
PyTorch Version: [2.9.1+cu130]
SageAttention Version: [3.1.0]
ComfyUI Custom Node: [ComfyUI-KJNodes (latest commit)]
Base Model: [wan2.2]
Request
Could the developers please investigate the compatibility of SageAttention3 with video generation models? It seems that the current acceleration mechanism might be compromising the temporal attention layers.
I am encountering significant temporal flickering and flashing artifacts when generating videos in ComfyUI with SageAttention enabled.
The issue manifests as inconsistent details and sudden light changes across frames, which severely breaks the video's temporal consistency. I have confirmed that this issue is directly caused by SageAttention, as the flickering completely disappears when I switch back to the default sdpa attention mode (with all other workflow parameters remaining unchanged).
️ Environment & Setup
OS: [ Windows 11]
GPU: [ NVIDIA RTX 5070Ti]
CUDA Version: [13.0]
PyTorch Version: [2.9.1+cu130]
SageAttention Version: [3.1.0]
ComfyUI Custom Node: [ComfyUI-KJNodes (latest commit)]
Base Model: [wan2.2]
Request
Could the developers please investigate the compatibility of SageAttention3 with video generation models? It seems that the current acceleration mechanism might be compromising the temporal attention layers.