WeVisOne-VL: Unified Visual Perception from Objects to Pixels

The work’s strongest asset is its systematic, closed-loop data engine (WDE), which uses model errors to drive acquisition, annotation correction, and hard-example mining across more than 100M images. This iterative curation, combined with a deliberately lightweight pixel decoder, yields strong pixel-level results for segmentation and matting without chaining a separate large model such as SAM. The model also shows unusually robust rejection behavior, surpassing prior VLMs by large margins on benchmarks that include absent-target queries.

Two flaws stand out. First, the reinforcement-learning stage optimizes only box-level rewards for segmentation and matting, meaning mask and alpha-matte quality are never directly reinforced by the RL objective. Second, the reported gains are inseparable from a massive, multi-stage proprietary data pipeline that relies on in-house detectors, iterative model feedback, and undisclosed hard-negative distributions; the paper offers no ablation that isolates architecture from data scale. The authors themselves acknowledge trade-offs: the curriculum sacrifices general-purpose data and omits chain-of-thought reasoning.

The empirical breadth is impressive—eight tables cover detection, referring, counting, pointing, OCR, segmentation, and matting—and WeVisOne-VL-4B frequently beats 7B–9B rivals. Yet comparisons are uneven. Detection results rely on F1 rather than AP, which the authors justify because generative VLMs lack calibrated confidence scores, but this makes comparisons with modern specialist detectors incomplete. Internal benchmarks such as WeRef-Bench and WeRefMat lack independent availability, so the strong rejection and matting claims cannot be fully verified externally.

Code and model weights are public, and the paper sketches training stages and loss functions, including the segmentation warm-up loss $\mathcal{L}_{\mathrm{seg}}=20\mathcal{L}_{\mathrm{focal}}+\mathcal{L}_{\mathrm{dice}}$ and the joint loss $\mathcal{L}_{\mathrm{joint}}=\mathcal{L}_{\mathrm{NTP}}+10\mathcal{L}_{\mathrm{seg}}$. However, reproducing the full results is blocked by data opacity: the 100M-image corpus is generated via an internal engine combining SAM3, Rex-Omni, an in-house detector, and iterative feedback loops that depend on unpublished model versions. Key hyperparameters and RL reward details are relegated to appendices, and the curated matting annotator itself requires extra high-quality supervision not readily accessible.