Tübingen Paper: Image Tokenizers Should Be Trained Between Images

University of Tübingen researchers propose training image tokenizers on the space between image encodings, addressing decoder artifacts and cutting diffusion FID to 8.48 with fewer artifacts and faster convergence.

2026-10-15 ~ 2026-10-15 · 2 related posts

• InterpTok: Training Image Tokenizers Between Samples Cuts Diffusion FID to 8.48 — CGTuebingen · 2026-10-15
• New paper: Image Tokenizers Should Be Trained Between Images cuts artifacts, speeds training — ostrisai · 2026-10-15