The generator knows more than the perceiver: vision models borrow back their own imagination

The generator sees better

Unified multimodal models can both see and draw, and the two halves have mostly been treated as roommates. In a preprint posted this week, the authors of GAFU show the generative half is the better observer: generation-side representations are more sensitive to fine-grained semantic changes than understanding-side representations. Their method compresses generation-side visual representations into compact G tokens, uses pairwise preference learning to encourage effective utilization, and — the practical part — enables task-conditioned retrieval by the understanding pathway without requiring explicit image generation at inference. m:…3893

The gains are large for a perception method: 20.00% relative improvement on VStar-Bench, 17.94% on OmniSpatial, and 8.62% on VSR-Bench, consistently beating SFT baselines across the BAGEL, Janus-Pro, and Lance architectures. One caveat the authors note: the advantages depend on correct spatial organization. This suggests a concrete architecture bet for developers building unified models — train the generator well, because its representations may be your best perception features. m:…3893

The reason it works is representational, not architectural: generation-side representations capture task-relevant changes that understanding-side representations miss, and GAFU consistently outperforms SFT baselines across diverse visual understanding tasks and unified-multimodal architectures. If your perception stack is a separate model bolted onto a generator, this paper is an argument that the generator was the perception model all along. m:…3893

Two roads through the model

A second preprint this week peers inside vision-language models and finds two distinct pathways for solving visual tasks: a direct pathway and a text-mediated pathway. Pathway selection is task-dependent and can be modulated by data distribution and prompt design — and the models show flexible fallback behavior, relying on the text-mediated pathway when the direct pathway is ablated. m:…2811

For developers, this is a debugging lens: when a VLM fails a visual task, the question may not be what it saw but which pathway it took. Prompt design and data distribution are not just input choices — they are routing decisions. This is the kind of mechanistic finding that turns into prompt-engineering best practice within months. m:…2811

One model, objects to pixels

The third preprint this week is the product-shaped one. WeVisOne-VL unifies object-level and pixel-level visual perception via structured token prediction, integrating a lightweight pixel decoder through a special token, and is trained with a three-stage curriculum SFT followed by mixed-task RL. It handles detection, localization, counting, OCR, segmentation, and matting in one model, and maintains robust rejection for object-absent queries — the hallucination guard that perception stacks need. m:…5945

Tencent released WeVisOne-VL to the open-source community, and the authors report leading performance on over 30 benchmarks. Put the three papers together and the vision stack's direction is clear: perception is converging into single models that borrow the generator's representations, route through inspectable pathways, and cover the whole object-to-pixel range. For teams still maintaining separate detection, segmentation, and OCR pipelines, the consolidation question just got harder to defer. m:…5945

The builder's takeaway

This suggests a near-term consolidation: one model that generates, perceives, and routes — with generation-side representations as the perception backbone, pathway-aware prompting as the debugging tool, and object-to-pixel coverage as the product surface. The separate-detector era is ending because the generator, trained to imagine fine-grained visual changes, turns out to be the most sensitive observer in the stack. m:…3893 m:…2811