Breaking01 / AI
Image caching lowers word accuracy in Z-Image generation
The researchers reported that GuideCache restored how closely images followed prompts by recalculating only the parts with the largest predicted errors.

Researchers reported on 20 Nov 2026 that word accuracy fell to 46% when they tested a cache on the Z-Image model.
The paper put word accuracy at 62% before caching. The researchers tested TaylorSeer, which reuses intermediate calculations, with 2.9 times fewer floating-point operations.
The paper reported that ImageReward and CLIP Score, both image evaluation metrics, barely changed.
The researchers proposed GuideCache, which recalculates only the parts with the largest predicted errors. They reported that Z-Image and Qwen-Image followed prompts as closely as they did with full computation, using 2.4 times fewer floating-point operations.
The paper reported that reducing the number of generation steps at the same computational cost preserved nearly all the text in the images.
The researchers did not disclose word accuracy by language in the abstract.