A new training paradigm optimizes few-step diffusion image editors end-to-end using VLM feedback for instruction adherence and content preservation plus DMD loss, matching supervised paired-data models without any pairs.
a blue truck
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Learning an Image Editing Model without Image Editing Pairs
A new training paradigm optimizes few-step diffusion image editors end-to-end using VLM feedback for instruction adherence and content preservation plus DMD loss, matching supervised paired-data models without any pairs.