VQ-VA World: Towards High-Quality Visual Question-Visual Answering

TL;DR

The paper tackles the scarcity of high-quality data for open-source Visual Question-Visual Answering (VQ-VA) by introducing VQ-VA World, a data-centric framework that builds a large, diverse training corpus through an agentic pipeline operating on web-interleaved documents. It also releases IntelligentBench, a human-curated benchmark to rigorously assess VQ-VA across world knowledge, design knowledge, and reasoning. Fine-tuning LightFusion on the ~1.8M VQ-VA World samples yields substantial gains (e.g., 53.06 on IntelligentBench) and narrows the gap to proprietary systems, while improving reasoning-based and standard image-editing capabilities. The authors provide full model checkpoints, data, and pipelines to stimulate open research and further progress in VQ-VA and multimodal instruction following.

Abstract

This paper studies Visual Question-Visual Answering (VQ-VA): generating an image, rather than text, in response to a visual question -- an ability that has recently emerged in proprietary systems such as NanoBanana and GPT-Image. To also bring this capability to open-source models, we introduce VQ-VA World, a data-centric framework built around an agentic pipeline for large-scale, targeted data construction. Leveraging web-scale deployment, this pipeline crawls a massive amount of ~1.8M high-quality, interleaved image-text samples for model training. For evaluation, we further release IntelligentBench, a human-curated benchmark that systematically assesses VQ-VA along the aspects of world knowledge, design knowledge, and reasoning. Training with VQ-VA World data yields strong empirical gains: it helps LightFusion attain 53.06 on IntelligentBench, substantially surpassing the best prior open-source baselines (i.e., 7.78 from vanilla LightFusion; 1.94 from UniWorld-V1), and significantly narrowing the gap toward leading proprietary systems (e.g., 81.67 from NanoBanana; 82.64 from GPT-Image). By releasing the full suite of model weights, datasets, and pipelines, we hope to stimulate future research on VQ-VA.

Figures (38)

• Figure 1: Examples of Visual Question-Visual Answering (VQ-VA), highlighting the substantial gap between existing closed-source models and open-weight models. The rightmost column further shows that a model trained with VQ-VA World dataset significantly improves its VQ-VA performance.
• Figure 2: Illustration of the VQ-VA World framework for creating VQ-VA data. The framework consists of two stages: (1) preprocessing, which classifies and filters web-interleaved documents, and (2) an agentic pipeline that generates VQ-VA samples from the filtered documents. The agentic pipeline contains five sub-modules: retriever, filter, instruction generator, rewriter, and reasoner.
• Figure 3: Illustration of the three question types in IntelligentBench. Each type is shown with two examples, and each example contains a question image, question text, and the answer image.
• Figure 4: Alignment between VLM and human scores. We compare Gemini-2.5-Flash vs. human experts, GPT-4o vs. human experts, and agreement among human experts. We report the Accuracy and Spearman Rank Correlation Coefficient (SRCC) for comprehensive comparison.
• Figure 5: Comprehensive visualization of model performance on IntelligentBench (Subset Design, part 1/9).