ViFA-Council: Multi-Agent LLM Deliberation for Vietnamese Folk Art Generation

ViFA-Council is a three-stage multi-agent framework aimed at two tricky generative tasks: image outpainting and educational story generation grounded in traditional Vietnamese folk paintings. Instead of relying on a single model to improvise its way through culturally specific content, the pipeline coordinates GPT-4o, Gemini 3.1 Pro, and Claude Sonnet 4.6 so the models can critique and refine each other’s output before synthesis.

The key idea is structured deliberation. Task-specific JSON schemas act as the handoff layer between natural-language reasoning and diffusion-based image generation through Banana Pro, which helps keep the system aligned with cultural constraints. That matters because low-resource artistic domains are exactly where generic image and text models tend to drift into visual clichés, invented details, or flattened symbolism.

The authors say experiments and a user study show the approach improves cultural fidelity and narrative coherence. The source code and data are being released, which makes this more than a one-off demo: it gives teams a concrete reference for building multi-agent creative pipelines that need guardrails, not just prompt engineering.

For game developers, the practical angle is clear. If you’re building concept-art tools, lore generators, localization-adjacent content systems, or any workflow that has to respect a specific visual tradition, this is a reminder that model orchestration can matter as much as model quality. The same pattern could be useful anywhere hallucination is less a novelty and more a production risk.