Left: vanilla LTX-2.3 — clean, but no camera control interface at all. Middle: the same model handed
the warp-history camera interface it was never trained on — it leaks the ghost (black patches) or slides the frame
instead of re-rendering. Right: the trained adapter — same tokens, now read as pure camera direction.
The middle column is why the LoRA exists. Same seed, prompt, and window in every column.
glassblower-shop
pan right · press 2.0–3.0s — He raises the glowing glass orb toward the viewer, turning it slowly.
Pan: base+warp slides the frame like a postcard — no re-rendering.
vanilla LTX (no camera interface)
base + warp interface (untrained)
trained v2 @2500
beekeeper-hives
dolly in · press 1.5–2.5s — She lifts her veil and holds a honey dipper out toward the viewer.
Dolly-in: base+warp leaks the ghost's disocclusion holes (black patches).
vanilla LTX (no camera interface)
base + warp interface (untrained)
trained v2 @2500
dragon-teacup
dolly in · press 2.5–3.5s — It pokes its head over the rim and puffs a small smoke ring at the viewer.
Dolly-in: same ghost bleed on the base arm.
vanilla LTX (no camera interface)
base + warp interface (untrained)
trained v2 @2500
yeti-teahouse
dolly in · press 2.5–3.5s — It slides a steaming cup across the counter and gives two slow blinks.
Dolly-in: ghost bleed again — trained arm reads the same tokens as pure camera direction.
vanilla LTX (no camera interface)
base + warp interface (untrained)
trained v2 @2500
sand-sculptor
orbit left · press 2.0–3.0s — She pats the tower smooth, then presents it to the viewer with both arms.
Orbit: biggest disocclusion of all — heavy masking on base, clean re-render when trained.