The multi-crop era of open MLLMs: AnyRes-style schemes with per-model hyperparameter tuning
gabriberton · x · 2026-10-19
Most open MLLMs in 2024-2025 used multiple crops per image (LLaVA-NeXT, InternVL, Gemma 3, Molmo), in forms like AnyRes, Pan&Scan, and overlapping crops — all broadly similar.
The cost: higher complexity and per-model hyperparameter tuning (crop count, resolution, overlap). The attached LLaVA-1.5 report figure illustrates the AnyRes dynamic high-resolution scheme with a 2×2 grid, encoding crops and a resized thumbnail before flattening into the LLM.
Related event: Four Years of MLLM Visual Encoding: From Cropping to Native Multimodal(8 posts)→
More from Multimodal
- MOSS-Image finetune Kromoss trends on Hugging Face under Apache-2.0 — lodestones · 2026-10-20
- LMArena Benchmarks Meta's Muse Wonder Against Claude, GPT and Gemini Flagships — arena · 2026-10-20
- AI music generation shows dramatically improved beat and chord sync — AIandDesign · 2026-10-20
- NVAI-open: free open-source local AI image tool fills selections on your own GPU — No-Tap6022 · 2026-10-20
- Suno highlights three Studio 2.0 features users may have missed — suno · 2026-10-20
- AI video of Ben Franklin giving honest art reviews goes viral on Reddit — aeeiee · 2026-10-20