Jinxiu Liu1,2* · Jianru Li1* · Tanqing Kuang1* · Xuanming Liu2 · Kangfu Mei3 · Yandong Wen2 · Weiyang Liu4
1South China University of Technology 2Westlake University 3Johns Hopkins University 4The Chinese University of Hong Kong
*Equal contribution
SymbOmni is an agentic omni model designed for cumulative evolution through Symbolic Concept Learning. Instead of solving every task "from scratch", it abstracts experience into reusable symbolic knowledge and composes that knowledge to tackle novel tasks.
Visual generation has witnessed rapidly expanding applications across diverse domains, ranging from text-to-image/video synthesis to interactive creation. However, prevailing monolithic models are fundamentally limited by an inability for cumulative learning and self-evolution (i.e., the "perpetual novice" problem), as they lack mechanisms to structure experiences into reusable knowledge and must instead rely on brittle, "from-scratch" reasoning, leading to poor compositional generalization and inefficient knowledge retention. To address these limitations, we introduce SymbOmni, an agentic omni model designed for cumulative evolution through Symbolic Concept Learning. Our architecture centers on the Symbolic Concept Box (CB)---an optimizable memory space 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, where a dual-verifier system generates language-based optimization signals to drive continuous self-improvement without gradient-based fine-tuning. Extensive evaluations demonstrate that SymbOmni achieves: (I) Superior Performance, significantly outperforming agent system for interative creation and leading closed-source models 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.
- Clone the repository, create and activate conda environment:
git clone https://github.com/Sphere-AI-Lab/symbomni.git
cd symbomni
conda create -n symbomni python=3.12
conda activate symbomni- Install dependencies:
pip install -r requirements.txt- ComfyUI Installation
- ComfyUI-Manager offers custom nodes management and installation.
- Hugging Face offers Models Installation
plan: Generates multiple candidate plans, supports manual/automatic selection, and executes specified atomic workflows step-by-step. Suitable for complex multi-step tasks.interactive: Interactive dialogue mode, generating one step per round. Suitable for exploration and experimentation.instruction: Heuristic search execution mode inspired by ComfyMind, suitable for simple tasks and experimental comparison.
python main.py --mode <mode> (--resource1 'PATH') (--resource1 'PATH') (--resource_dir 'PATH')If you find our work helpful, please consider citing our paper:
@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},
}We would like to thank the authors of the following projects for their excellent works.
This code is released under the MIT License.