InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks

Vision foundation models lagged language models in scale, capping multimodal capability.

Scales the VISION foundation model to 6B parameters and progressively aligns it with the LLM using web-scale image-text data, with dynamic resolution and tiled high-resolution input fed through a pixel-shuffle-style MLP projector. Intended as a drop-in open alternative to ViT-22B.

Venue CVPR 2024 (crossref container '2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition'). The InternVL lineage is where vision-token reduction is most systematically engineered - its own successor (pap record internvl-x) and many successors attack token count directly, which is a sign the base design's token cost was felt as a practical problem rather than an academic one. Dynamic tiling plus pixel shuffle is the same trade as LLaVA-NeXT anyres: resolution for accuracy, paid for in tokens. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: cvpr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.