VisCo reuses a pretrained VLM to compress visual tokens more effectively

Yupeng Zheng · hf · 2026-07-27

• VisCo tackles visual token compression for VLMs, aiming to reduce inference latency and memory overhead.
• Instead of adding an external compression module, it reuses the pretrained VLM itself as an intrinsic compressor.
• The method is a parameter-sharing autoencoder that compresses visual information into a small set of memory tokens and transfers hierarchical information from encoding to decoding.
• Experiments show stronger results than prior methods across all tested compression ratios, with especially large gains at aggressive compression levels.
• The approach remains stable even in the extreme single-token setting, and the learned memory tokens can also improve the base model when combined with original visual tokens.

Original post →

More from Research

• PRO-LONG scores 97.4% on ARC-AGI-3 with a log-file harness and 30-line prompt — srchvrs · 2026-07-27
• SceneActBench tests whether VLM agents can act in full 3D scenes — Yifei Zhao · 2026-07-27
• Tencent Hunyuan maps scaling laws for native multimodal pretraining from scratch — Tencent-Hunyuan · 2026-07-27
• A training-free method fixes revisit inconsistency in generative rendering — Wenchao Ma · 2026-07-27
• NVIDIA open-sources Molt, a PyTorch-native framework for agentic RL — nvidia · 2026-07-27
• Skill Self-Play uses co-evolving skills to push LLM capability — QwenBusinessUnit-Edu · 2026-07-27

Research channel →