Vision-OPD uses 6.2K synthetic samples to boost fine-grained vision understanding

小红书技术REDtech · wechat · 2026-07-28

Xiaohongshu’s dots team introduces Vision-OPD, a vision-language post-training method designed to close the “region-to-global gap” in fine-grained visual understanding.

The paper argues that multimodal models often answer correctly when shown a cropped, zoomed-in region but fail on the full image. Vision-OPD turns that advantage into training signal by using an online self-distillation setup: a cropped-image teacher and a full-image student share the same model, while token-level divergence is minimized on the student’s own sampled trajectories.

Key results:

Only about 6.2K fully synthetic samples were used.
A 9B model reached 79.7 average accuracy across fine-grained benchmarks and ranked first overall.
A 4B version reached 77.1 and also beat larger closed and open models.
Performance on broader visual tasks stayed intact, suggesting the method improves fine detail perception without forgetting general vision skills.

Original post →

Vision-OPD uses 6.2K synthetic samples to boost fine-grained vision understanding

More from Multimodal