HEaD+ detects object hallucinations early in diffusion generation via cross-attention maps, text, and a Predicted Final Image, raising complete image rates by 6-8% for four-object prompts and reducing time by up to 32%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Hallucination Early Detection in Diffusion Models
HEaD+ detects object hallucinations early in diffusion generation via cross-attention maps, text, and a Predicted Final Image, raising complete image rates by 6-8% for four-object prompts and reducing time by up to 32%.