DSC
A metric measuring the degree of overlap in image segmentation, ranging from 0 to 1, where values closer to 1 indicate better overlap between predictions and ground-truth annotations.
Timeline (last 90 days)
On the IH EPVS dataset, the high-resolution T1 single modality achieves a DSC of 0.703, whereas fusing the three modalities T1, T2 and FLAIR lowers the DSC to about 0.695.
Occluding FLAIR reduces the DSC by 70.2%, an impact comparable to occluding T1.
At each sampling step, DSC evaluates, based on the current state, how much intervention is needed, in which space the intervention should be applied, and when to stop the intervention.
DSC is designed as a plug-and-play plugin that can adapt to different safety protection modules without retraining the model.
DSC (Decoupled Safety Control) is a plug-and-play training-free safety guidance control algorithm.
In standard safety benchmark experiments, DSC improved safety without sacrificing the quality of normal generation or prompt fidelity.
As an inference-stage plugin, DSC does not introduce a significant computational burden, and inference efficiency remains at a high level.
In standard safety benchmark experiments, DSC was combined with representative latent-space and text-space safety protection schemes, and the results show that DSC can suppress harmful content generation more effectively.
DSC reformulates training-free safety guidance as an explicit state-dependent control process, introducing a closed-loop idea, where the intervention actually executed is determined by the current generation state.
All Facts (9)
On the IH EPVS dataset, the high-resolution T1 single modality achieves a DSC of 0.703, whereas fusing the three modalities T1, T2 and FLAIR lowers the DSC to about 0.695.
50%UnverifiedOccluding FLAIR reduces the DSC by 70.2%, an impact comparable to occluding T1.
50%UnverifiedAt each sampling step, DSC evaluates, based on the current state, how much intervention is needed, in which space the intervention should be applied, and when to stop the intervention.
50%UnverifiedDSC is designed as a plug-and-play plugin that can adapt to different safety protection modules without retraining the model.
50%UnverifiedDSC (Decoupled Safety Control) is a plug-and-play training-free safety guidance control algorithm.
50%UnverifiedIn standard safety benchmark experiments, DSC improved safety without sacrificing the quality of normal generation or prompt fidelity.
50%UnverifiedAs an inference-stage plugin, DSC does not introduce a significant computational burden, and inference efficiency remains at a high level.
50%UnverifiedIn standard safety benchmark experiments, DSC was combined with representative latent-space and text-space safety protection schemes, and the results show that DSC can suppress harmful content generation more effectively.
50%UnverifiedDSC reformulates training-free safety guidance as an explicit state-dependent control process, introducing a closed-loop idea, where the intervention actually executed is determined by the current generation state.
50%