How modern MLLMs handle native resolution: heavy ViT fine-tuning plus 2D RoPE
gabriberton · x · 2026-10-19
The thread explains why native-resolution image input was hard: ViTs were pretrained on small fixed squares (224-448px) while native images can be 1k×1k or larger.
The bitter-lesson solution most labs adopted: feed the whole image at native resolution up to a token cap (e.g. 16k tokens in Qwen3-VL), with the ViT undergoing extensive fine-tuning inside the MLLM, usually adding 2D RoPE. Examples: Qwen3-VL, Kimi-VL, Gemma 4.
Related event: Four Years of MLLM Visual Encoding: From Cropping to Native Multimodal(8 posts)→
More from Multimodal
- Flux3 Image Klein Distilled hands-on: 5.6B open model with 10-reference editing runs on 16GB — QuiveringLichenjr · 2026-10-19
- Qwen Image 2.1's built-in prompt enhancer is unusable: 12k tokens still not enough — dtdisapointingresult · 2026-10-19
- MOSS-Image first impressions: 'this is bad' — ranked below SDXL on quality-speed ratio — Electronic_Award1138 · 2026-10-19
- 6 open-source image models compared on one prompt: Krea2, Qwen2.1, Flux3 Klein, Flux 2 Klein, ZIT, MOSS — cradledust · 2026-10-19
- Moss speed tests: cutting steps from 50 to 20 slashes gen time from 27s to 11s — Winter_unmuted · 2026-10-19
- Ludovic Creator shares stunning Amazon warrior AI artwork in community chain — LudovicCreator · 2026-10-19