---

Pairwise Comparisons: Foreground Object Consistency

---

Example 1

Vbench metric (VB-SC) fails to capture the finegrained distortions near the cat's face. In addition, VB-SC tends to favour videos with lesser camera motion. Our Tracker-FG captures fine-grained long term dependencies more effectively.

---

Example 2

Vbench metric (VB-SC) fails when there are multiple subjects in the scene, likely assigning every subject instance into a single feature. Tracker-FG tracks each subject individually and computes consistency.

---

Example 3

Vbench metric (VB-SC) is highly sensitive to camera motion as evident in its score (0.55 compared to values close to 0.95).

---

Example 4

The bias of camera motion on Vbench metric (VB-SC) is evident in its scores, whereas Tracker-FG purely focuses on object consistency.

---

Cases where primary objects are not generated from the same prompt

Example 5

The model in Video 2 is incapable of generating the primary object, thus we consider this video to be low quality. Object detection does not work for this scene. Therefore, we use object detection as the pairwise metric.

---