WeVisOne-VL - Matting WeRefMat Referring MAD: leaderboard
WeVisOne-VL official release:greedy decoding,temperature0,max1960visual tokens/image;Qwen3-VL Instruct and Qwen3.5 Thinking baselines. Generative visual grounding/perception,not image generation. Classical detectors and specialist OCR systems are outside the VLM cohort;mean alpha error over pixels,source multiplies by1000;only release VLMs,not classical matting engines or composite Qwen plus SAM systems. MAD (×1000).
Metric: MAD (×1000). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | WeVisOne-VL-4B | 9.8 | |
| 2 | WeVisOne-VL-2B | 15.8 |
Interactive version: theaggregate.ai/benchmark?slug=wevisone-vl-matting-werefmat-referring-mad · How It Works · Data refreshed daily, snapshot 2026-10-24.