InterpTok: Training Image Tokenizers Between Samples Cuts Diffusion FID to 8.48
CGTuebingen · hf · 2026-10-15
Image Tokenizers Should Be Trained Between Images (Univ. of Tübingen)
Latent diffusion models generate via an image tokenizer's codes; when generated codes fall outside the decoder's expected distribution, images show artifacts. Prior work adds noise during decoder training, but generated codes resemble mixtures of training codes more than a code plus noise.
InterpTok
- Fine-tunes MAETok to decode mixtures of image codes into natural images
- Since no target exists between two samples, a discriminator judges whether each decoded mixture looks as natural as its two sources
- A mixing-ratio loss preserves the blend ratio, preventing trivial collapse
Results
- Keeps MAETok's reconstruction quality; interpolation quality improves >3x, best in a recent benchmark
- Matched diffusion transformer's guidance-free FID: 11.61 → 8.48, and 8.11 at 1/8 sampling steps
- Generator on InterpTok codes reaches MAETok's final result in half the training steps
More from Multimodal
- A Midjourney portrait from February 2023 shows how far image generation has come — LinusEkenstam · 2026-10-15
- Gemini Argon 4 demo wows AI community — Saboo_Shubham_ · 2026-10-15
- SayGM posts video tutorials: crayon dragon flight, one-sentence stop-motion, stills-to-clip — markjeffrey · 2026-10-15
- A miniature sea inside a violin: creator's 30-second Seedance 2.5 demo — LudovicCreator · 2026-10-15
- "The old machines got us this far": AI-generated retro-future megacity stuns — piotrbinkowski · 2026-10-15
- Same Prompt, Two Models: User Asks Which Vertical-GPU Edit Is Gemini vs ChatGPT — BayrischerVorkoster · 2026-10-15