Figure 3. Attention score visualization for various query points.
Left: Vanilla CLIP (A ) shows noisy, unfocused attention.
CLIP
Center: VFM-to-CLIP distillation without low-rank eigenscaling
shows partial object grouping with limited detail. Right: Incor-
porating our low-rank eigenscaling (Eq. (7)) captures object-level
context, improving grouping within a single object.