MixTok hybrid tokenizer hits gFID 1.95 on ImageNet with just ~131 tokens per image
ZhejiangUniversity · hf · 2026-10-12
Zhejiang University, HKUST, and the University of Adelaide introduce MixTok, a hybrid visual tokenizer combining a compact 1D global representation with sparse 2D refinement tokens. A content-adaptive router ranks candidate locations and allocates the 2D token budget per image. On ImageNet 256x256 it achieves rFID 1.195 and gFID 1.95 at average budgets of roughly 128 and 131 tokens per image, balancing compactness with fine detail.
More from Multimodal
- Midjourney film-look style code --sref 3534013297 shared with prompts in ALT — umesh_ai · 2026-10-12
- Kling 4.0 Flash demo fuels speculation full version ships next week — Kyrannio · 2026-10-12
- Lux in Music repo launches, showcasing YuE2 music model in action — TheMoonMidas · 2026-10-12
- "QT Your Eye" AI art thread floods X with surreal golden eye images — LudovicCreator · 2026-10-12
- Sailor Moon vs Godzilla short made with MiniMax H3 video model — Certain_Potato_4509 · 2026-10-12
- Opus 5.5 directs Jimeng CLI to auto-generate storyboards for surprisingly polished short films — op7418 · 2026-10-12