The paper addresses Text-and-Image-to-Image (TI2I) flow matching, a class of generative models used for controllable image editing. The authors position Classifier-Free Guidance (CFG) as a standard ingredient that is widely credited for producing high-quality outputs.
The title signals a critical reexamination of CFG in this setting. Rather than taking the guidance mechanism for granted, the work appears to ask whether CFG is optimally applied in TI2I flow matching and how it might be reconsidered.
The available abstract is brief; it sets out the study's scope and motivation but does not yet reveal the proposed changes or experimental outcomes. Readers will need the full paper to see what the rethinking entails.