KT
KathaTrace
Semantic Trajectory Gap Demo
VLM JUDGE · TEXT-ONLY · IMAGE-ONLY · TEXT+IMAGE

Stories that look whole, until you trace the gap.

KathaTrace decomposes a storyboard into scene-to-scene transitions, then asks the same question three ways — from text alone, from images alone, and from both — to measure exactly how much narrative meaning a generator loses on the way from page to picture.

Browse KathaBench-25K
5,000
narratives
20,000
transitions
28,712
recoverability Qs
0.845
fleiss' kappa
23.5
mean STG
Step 01

Tell us a story

Pick a curated narrative from KathaBench-25K, or imagine pasting your own fable. A short moral story works best — KathaTrace needs at least one scene-to-scene transition to trace.

AE The Cat and the Birds
AE The Peacock and the Crane
AE The Wolf and the Lamb
Story
Meaning
Beware deceptive appearances.

Step 02

Live storyboard render

Press "Compose Storyboard" above — KathaTrace traces each transition while the scenes assemble, then reveals the rendered storyboard for the selected story.

Step 03

Switch evidence conditions

The same five transitions, re-scored under each evidence condition. Watch image-only recoverability fall well below text-only — that drop is the Semantic Trajectory Gap.

Causal link
94%
Intentional link
91%
Action link
96%
Consequence link
89%
23.5
Semantic Trajectory Gap for this storyboard — text-only minus image-only recoverability, averaged across all four transition dimensions.