DGS-Net decomposes gradients into harmful and beneficial directions, projects task updates onto the orthogonal complement of harmful ones, and aligns with distilled signals from frozen CLIP to fine-tune for AI-image detection while preserving priors.
UnivFD demonstrates that CLIP can effectively extract artifacts from images
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection
DGS-Net decomposes gradients into harmful and beneficial directions, projects task updates onto the orthogonal complement of harmful ones, and aligns with distilled signals from frozen CLIP to fine-tune for AI-image detection while preserving priors.