SAVIER: Source-Anchored Visual Inspection and Evidence-Guided Repair for Vietnamese Cultural Text-to-Image Generation
| Source: arXiv AI
Tags: SAVIER, text-to-image, SDXL, FLUX.1-dev, Culture-TRIP, ViCuP
SAVIER is a training-free inspect-and-repair loop for Vietnamese cultural text-to-image prompts that raises visual contract fulfillment from 55.6 to 69.6 on SDXL and from 59.1 to 67.3 on FLUX.1-dev.
Details
SAVIER repairs how a named Vietnamese cultural entity is depicted in a generated image while preserving user-requested scene details, in one pass without training. Starting from a Culture-TRIP expanded prompt, a Curator separates cultural cues from preservation requirements, a prompt-blind Observer describes the first image, and a Refiner turns discrepancies into at most three instructions before one regeneration with the same seed and a fixed bank of reference photos. On the 50 bilingual ViCuP prompts it improves contract fulfillment on SDXL and FLUX.1-dev. An ablation shows the reference photographs account for most of the gain, while agent roles add country-level relevance on FLUX.1-dev.