A training-free consistency method for text-to-image DiT models that keeps characters and backgrounds stable across shots via point-tracking attention, adaptive token merge, and foreground/background masking.
Kingma and Max Welling
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
CharaConsist: Fine-Grained Consistent Character Generation
A training-free consistency method for text-to-image DiT models that keeps characters and backgrounds stable across shots via point-tracking attention, adaptive token merge, and foreground/background masking.