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

Abstract

This paper presents ViFA-Council, a three-stage multi-agent framework that employs multiple large language models (LLMs) to tackle two culturally complex generative tasks: image outpainting and educational story generation based on traditional Vietnamese folk paintings.

Current single-model generative pipelines frequently struggle with stylistic hallucinations and cultural misrepresentations because they lack mechanisms for cross-model critique. ViFA-Council addresses this challenge by orchestrating collaboration among GPT-4o, Gemini 3.1 Pro, and Claude Sonnet 4.6. It enforces rigorous cultural constraints through structured agent deliberation. This deliberation is mediated by task-specific JSON schemas that effectively bridge natural language discussions with diffusion-based image synthesis using Banana Pro.

Experiments and a user study demonstrate that structured multi-agent deliberation is a promising direction for improving cultural fidelity and narrative coherence in culturally sensitive, low-resource artistic domains.

Method

A three-stage Council that drafts, cross-refines and adjudicates

Given a folk painting I and a user directive q, the Council produces one Chairman-selected JSON configuration J*. Every stage reads and writes the same task-specific schema, so cultural and stylistic constraints are tightened before any image is generated.

Pipeline diagram: an input image and task prompt go to GPT-4o, Gemini and Claude; their three JSON drafts are cross-refined into a three by three grid; a Gemini 3.1 Pro Chairman selects one JSON, which Nano Banana Pro renders as an outpainting or a story page — The LLM Council architecture. Parallel drafting, then *M × N* cross-model refinement, then Chairman-based adjudication.

Parallel drafting

GPT-4o, Gemini 3.1 Pro and Claude Sonnet 4.6 each fill the schema from the same painting and directive, without seeing one another’s output. Calls run concurrently with exponential-backoff retries.

3 independent drafts
Cross-model refinement

Every agent reviews and improves every draft, whoever wrote it: checking structural validity, improving cultural fidelity, and returning the same schema. A failed call falls back to the original draft, so no candidate is lost.

up to 3 × 3 = 9 refined candidates
Chairman adjudication

A separate Gemini 3.1 Pro instance, isolated from the earlier stages, receives all drafts and refinements under alphabetical labels. It scores structural validity, logical coherence, cultural fidelity and whether refinement helped, then names one winner.

1 selected configuration J*

JSON schema fieldsThe schema is the bridge between LLM deliberation and image synthesis.
Task	Field	Description	Purpose
Outpainting	`input_image_analysis`	Original style and dominant colors	Source-aware visual analysis
Outpainting	`expansion_settings`	Direction, pixel amount, and mask blur	Synthesis spatial control
Outpainting	`seamless_blending`	Style, texture, color, pattern constraints	Boundary and appearance continuity
Outpainting	`outpainting_scenarios[]`	Expansion proposals with prompts	Multi-candidate generation
Story	`source_material`	Title, era, cultural context, and symbols	Cultural grounding
Story	`art_style_preservation`	Character, line, and perspective constraints	Prevent visual-style drift
Story	`characters[]`	Appearance, style, and cultural significance	Cross-panel character consistency
Story	`story_panels[]`	Layout, dialogue, motifs, and transitions	Sequential narrative specification
Story	`educational_content`	Objectives, explanations, and glossary	Educational adaptation

Task 1

Folk-painting outpainting

We compare ViFA-Council against GPT-4o, Gemini 3.1 Pro and Claude Sonnet 4.6, each prompted with the identical Stage 1 instruction but without cross-refinement or Chairman adjudication, and rendered through the same image pipeline. Experiments use a curated collection of 257 paintings: 112 Đông Hồ, 74 Hàng Trống and 71 Kim Hoàng works.

Four source paintings, each shown beside the ViFA-Council outpainting and the outpaintings from GPT-4o, Gemini 3.1 Pro and Claude Sonnet 4.6 — Outpainting results. From left: original, ViFA-Council, GPT-4o, Gemini 3.1 Pro, Claude Sonnet 4.6. Select the figure to enlarge it.

GPT-4o

Typographical content injection

Modern Vietnamese text is written straight onto the canvas, such as “VƯỜN NHÀ BÌNH YÊN” and “BÉ VUI ĐÙA”.
Gemini 3.1 Pro

Cultural hallucination

Imperial swirling clouds, ornamental rock gardens and curved-roof pagodas replace naive rural aesthetics with generic East Asian tropes.
Claude Sonnet 4.6

Textural and chromatic degradation

Boundaries are structurally plausible, but the earthy palette is flattened and the fibrous texture of giấy điệp paper is lost.
Gemini 3.1 Pro, stories

Structural format drift

Comic panels morph into modern educational infographics with rigid typographic headers.

Outpainting, automated metricsCLIP measures style consistency; LPIPS and DreamSim measure content consistency; JSON Semantic Score rates the specification from 0 to 100.
Model	CLIP ↑	LPIPS ↓	DreamSim ↓	JSON Semantic (%) ↑
GPT-4o	0.8348	0.7832	0.5896	70.10
Claude Sonnet 4.6	0.8485	0.7986	0.6169	84.30
Gemini 3.1 Pro	0.8520	0.7705	0.6134	76.70
ViFA-Council (ours)	0.8524	0.7317	0.5745	93.90

More outpainting results

Pick a painting to see the source beside the ViFA-Council extension.

Source · Đông Hồ

ViFA-Council

Task 2

Educational story generation

The same Council turns a single painting into a multi-panel comic for children: characters stay consistent with the source print, dialogue is written in Vietnamese, and the story carries the painting’s cultural lesson. The baseline is Gemini 3.1 Pro with the Stage 1 instruction alone.

Three source paintings, each shown beside the ViFA-Council comic page and the Gemini 3.1 Pro page — Story generation results. From left: original, ViFA-Council, Gemini 3.1 Pro. ViFA-Council keeps a comic-strip format with speech bubbles; the baseline drifts toward infographic layouts with typographic headers.

Story generation, automated metricsDINOv2 measures protagonist consistency across panels; PickScore measures text-image alignment.
Model	CLIP ↑	DINO ↑	PickScore ↑	JSON Semantic (%) ↑
Gemini 3.1 Pro	0.667	0.504	18.573	87.0
ViFA-Council (ours)	0.703	0.601	19.61	93.0

Story pages from ViFA-Council

Hai Bà Trưng đánh giặcSource: Đông Hồ print

Thầy đồ CócSource: Đông Hồ print

Gà thư hùngSource: Đông Hồ print

Nông nhànSource: Đông Hồ print

Đám cưới chuộtSource: Đông Hồ print

Bát tiên náo hảiSource: Hàng Trống painting

Select a page to enlarge it.1 / 6

User study

Blind ratings from 15 participants

Ten STEM students, three educators and two AI engineers rated anonymized results in randomized order on a five-point Likert scale. ViFA-Council receives the highest overall rating on both tasks, with the lowest variance. For stories, the largest gains are in dialogue quality (3.96 vs. 3.54) and narrative coherence (3.94 vs. 3.66).

OutpaintingSeamless blending, style consistency, cultural relevance, overall.
Method	Blend	Style	Culture	Overall
GPT-4o	3.18	3.42	3.52	3.37 ± 1.16
Gemini 3.1 Pro	3.36	3.48	3.55	3.46 ± 1.11
Claude Sonnet 4.6	3.95	3.64	3.96	3.85 ± 1.00
ViFA-Council	4.08	4.08	4.00	4.05 ± 0.89

Story generationCultural relevance, narrative coherence, dialogue quality, overall.
Method	Culture	Narrative	Dialogue	Overall
Gemini 3.1 Pro	3.56	3.66	3.54	3.59 ± 1.05
ViFA-Council	3.89	3.94	3.96	3.93 ± 0.88

Limitations

ViFA-Council is evaluated zero-shot, without fine-tuning on Vietnamese folk-art data, which may limit adaptation to subtle school-specific conventions. The user study measures perceived cultural relevance and does not include professional folk-art historians or artisans, so historical and iconographic authenticity still needs expert assessment. The framework depends on the underlying LLMs and image model, and its multiple parallel API calls cost more and take longer than single-model generation.

Citation

BibTeX

@inproceedings{Dang2026MAPR,
  title     = {ViFA-Council: Multi-Agent LLM Deliberation for Vietnamese Folk Art Generation},
  author    = {Nguyen, Hai-Dang and Pham, Minh-Phuong and Dao, Thao Thi Phuong and Do, Trong-Le and Nguyen, Vinh-Tiep and Le, Trung-Nghia},
  booktitle = {International Conference on Multimedia Analysis and Pattern Recognition (MAPR)},
  year      = {2026}
}

Acknowledgment

This research is funded by Vietnam National University - Ho Chi Minh City (VNU-HCM) under Grant Number B2026-18-17.

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

A three-stage Council that drafts, cross-refines and adjudicates

Parallel drafting

Cross-model refinement

Chairman adjudication

Folk-painting outpainting

Typographical content injection

Cultural hallucination

Textural and chromatic degradation