A training-free two-step pipeline, localize a bounding box then refine the answer from a high-resolution crop, improves MLLM accuracy on high-resolution image benchmarks.
How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Zoom-Refine: Boosting High-Resolution Multimodal Understanding via Localized Zoom and Self-Refinement
A training-free two-step pipeline, localize a bounding box then refine the answer from a high-resolution crop, improves MLLM accuracy on high-resolution image benchmarks.