Global Native-Resolution Encoding (GNE)
Paradigm for processing high-resolution images in MLLMs where the entire image is encoded in a single global ViT forward pass at native resolution — as opposed to Slice-Based Encoding (SBE) which tiles the image into fixed-size crops.
Why it matters
High-resolution inputs are necessary for tasks requiring fine detail: OCR, dense chart reading, small-object detection, spatially-distributed reasoning. GNE preserves cross-region context that SBE breaks at tile boundaries.
Key examples
- moonvit — Kimi-VL’s encoder; the primary production GNE system
- ViT-UHD — extends GNE with more aggressive token compression (surpasses MoonViT on TTFT)
Core challenge
Quadratic attention cost in the number of patch tokens. An image at 1920×1080, patch size 14 → ~9800 tokens → ~96M attention operations. Mitigated by:
- FlashAttention varlen kernels
- Post-encoder token compression (pixel shuffle, pooling, learned merging)