FlexTok: Resampling Images into 1D Token Sequences of Flexible Length

A cotton-candy dawn over a lab bench, with a dial that tells the story long or short.

flextok.epfl.chFlexTok: Resampling Images into 1D Token Sequences of Flexible Length
Screenshot of https://flextok.epfl.ch/, © the authors. Shown for reference. Request removal
Venue
ICML 2025
Year
2025
First author
Roman Bachmann
Field
image tokenization
Template
Nerfies (Bulma 0.9.1) with custom glass nav, slider summary and related-project cards
Theme
Light

Checked against the rendered page. Google Sans loads only 400 so the 600 title is synthesized bold; American Typewriter is macOS-only and renders as Times New Roman on Windows.

Why it is here

Pastel gradient header with sticky dropdown navigation. The token-count teaser grid comes straight after it, followed by an interactive TL;DR-to-detailed slider for the summary and a row of three overview cards. Strong hierarchy and an interaction idea others could reuse.

Overview

Color Palette

Typography

Google Sans
FlexTok: Resampling Images into 1D Token Sequences of Flexible Length Abstract · Method · Results · BibTeX — 0123456789
BlinkMacSystemFont / Segoe UI system stack
FlexTok: Resampling Images into 1D Token Sequences of Flexible Length Abstract · Method · Results · BibTeX — 0123456789
American Typewriter
FlexTok: Resampling Images into 1D Token Sequences of Flexible Length Abstract · Method · Results · BibTeX — 0123456789

Spacing & Shape

Page Modules

Guidelines

More like this

ACL 2025
Scaling Text-Rich Image Understanding via Code-Guided Synthetic Multimodal Data Generation
A university press catalogue: small-caps serif title, a colour wheel of specimens and an earthy bar chart.
ACL 2025
DICE-BENCH: Evaluating the Tool-Use Capabilities of Large Language Models in Multi-Round, Multi-Party Dialogues
A Nerfies page with a game-box cover: a charcoal dice banner stamps the identity, then pastel brand-tinted pills and a grey carousel strip carry you into a quiet two-column method read.
ACL 2025
TUNA: Comprehensive Fine-grained Temporal Understanding Evaluation on Dense Dynamic Videos
A benchmark that leads with its scoreboard: a fish-logo small-caps wordmark, a gradient venue line, then a full-width grey-headed leaderboard with zebra rows and blue model links.
CHI 2026
Gesturing Toward Abstraction: Multimodal Convention Formation in Collaborative Physical Tasks
A seminar handout on clean paper: a big soft headline, six charcoal tabs, and a teal-and-magenta comic strip of hands explaining blocks.
CHI 2026
Making Multimodal LLMs Reliable Chart Data Extractors: A Benchmark and Training Framework
A lab notebook opened to its best page: a framed video like a monitor bezel, emoji in the headline, and every figure pinned with a soft shadow.
CHI 2026
What If Moderation Didn't Mean Suppression? A Case for Personalized Content Transformation
A quiet reading room with a purple ribbon on the lectern: slate sans headlines, Garamond prose, and one warm amber note pinned to the wall.