Haina Zhu

Full-Context Rendering Reduces Music Codec Exposure Bias

FullDiT conditions a diffusion transformer on imperfect eight-stream codec plans, lyrics, and captions, improving ViSQOL by 0.77 under synthetic corruption.