Wan 2.2 has a left-to-right word-order bias; rewriting prompts doubles spatial accuracy

PrateekVJoshi · x · 2026-11-20

Prateek VJoshi found an interesting spatial bias while experimenting with the open video model Wan 2.2: prompts like "a mug to the right of a laptop" often produce reversed placement.

The model appears to have a strong left-to-right ordering bias — the first object mentioned in the prompt tends to appear on the left, regardless of the spatial instruction.

He devised a simple inference-time workaround: instead of stating positions explicitly, rewrite prompts so the leftmost object is named first. In his experiments, spatial placement accuracy roughly doubled just from reordering words.

Takeaway: much of spatial instruction following may hinge on something as simple as word order.

Original post →

More from Multimodal

Multimodal channel →