Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{209372,
        author = {Allu Sowrya Deepika and Dhavala Lalitha bhaskari},
        title = {A Multimodal AI Framework for Telugu Poetry Understanding and Visual Representation},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {5},
        pages = {1572-1581},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=209372},
        abstract = {Classical Telugu poems (padyams) express the essence in poetic metrical verse, with the use of archaic words and cultural allusion, making them difficult to read by the modern reader and even by general language models. This paper presents the Padyam2Gadyam+ pipeline, which takes a Telugu poem and translates the same into Telugu prose (gadyam) and into an image. Using only 549 poem–meaning–English triples from a set of 599 poem–meaning–English triples, Gemma-2-9B-it is trained with 4-bit Quantized Low-Rank Adaptation (QLoRA), which updates 0.58% of the model's parameters. In Module 2, the English meaning is given as a structured text instruction to be passed to Gemini API as a text-to-image prompt. For Module 3, a prompt is created with Stable Diffusion 1.5 and LoRA adapter. The model is fine-tuned for a test set of 20 poems which were not used in the training set, and a relative BLEU gain of 86.7% is observed compared to the same base model with three shots. The baseline scores for all three measures are all below the 95% confidence interval limits for the fine-tuned model. Qualitative analysis is used for representative examples to analyze prompt and image outputs. The modules are wrapped in a Flask Web App which will provide the English translation and image of a given poem. The results indicate that it is possible to develop an adaptation of a large-scale quantized model for the task of poetry interpretation using a small and carefully selected subset of data.},
        keywords = {Telugu poetry, low-resource machine translation, QLoRA, Gemma, Stable Diffusion, multimodal AI.},
        month = {October},
        }