Book cover design (A Song of Ice and Fire).
Abstract
Visual generation is rapidly expanding into diverse domains, from text-to-image/video synthesis to multimodal interactive creation. However, prevailing monolithic models are fundamentally limited by an inability for cumulative learning and self-evolution—the “perpetual novice” problem. They lack mechanisms to structure experiences into reusable knowledge and consequently rely on brittle, “from-scratch” reasoning for each task, leading to poor compositional generalization and inefficient knowledge retention. To address this, we introduce SymbOmni, an agentic omni-model designed for cumulative evolution via Symbolic Concept Learning. Our architecture centers on the Symbolic Concept Box (CB)—an optimizable memory module that abstracts low-level operations into reusable Symbolic Workflow Instructions (SWIs). SymbOmni operates through an induction-transduction cycle: experiences are abstracted into symbolic concepts (induction), which are then strategically composed to solve novel tasks (transduction). This cycle is advanced by verbalized backpropagation with language-based optimization gradients to drive continuous self-improvement without gradient-based fine-tuning. Extensive evaluations demonstrate that SymbOmni achieves: (I) Superior Performance, significantly outperforming existing agent-based systems for iterative creation and leading closed-source models (e.g., Nano Banana, GPT-Image-1) in both image quality and task success rates; (II) Enhanced Efficiency, reducing token consumption by over 30% while maintaining competitive generation quality; (III) Continuous Learning, demonstrating genuine cumulative improvement in online learning scenarios on ComfyBench, establishing a new state-of-the-art.
Methodology
Illustration of the self-evolving and symbolic concept learning mechanism in SymbOmni. The framework transforms user instructions into solutions through an iterative process. Key to its evolution is the dual-feedback loop: successful outcomes are abstracted into symbolic experiences stored in the Memory Bank, while failures initiate a replay process that refines future planning. This creates a continuous learning cycle, allowing the agent to accumulate and leverage knowledge beyond a single task.
Text-to-Image Generation Examples
High-quality image of snowy mountains and a river.
High-quality image of a winding trail through a dense forest.
High-quality image of an underwater coral reef with colorful fish.
People sitting around a campfire under a starry sky.
Automated robots in a futuristic factory.
2× super-resolution image of a snowy mountain peak.
Video Generation Examples
Cartoon panda walking in a bamboo forest.
Skateboarder performing tricks in an urban skate park.
Dog happily playing with a ball in a park.
Image Editing Comparisons
Qualitative comparison against leading agent systems and closed-source models (Ours vs. ComfyMind, BAGEL, Nano Banana, GPT-Image-1) across style transfer, object removal, inpainting, repaint, image matting, reasoning editing, and object insertion.
Complex Reasoning Generation
BibTeX
@inproceedings{liu2026symbomni,
title = {SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning},
author = {Liu, Jinxiu and Li, Jianru and Kuang, Tanqing and Liu, Xuanming and Mei, Kangfu and Wen, Yandong and Liu, Weiyang},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026},
}