"

Cookies ussage consent

Our site saves small pieces of text information (cookies) on your device in order to deliver better content and for statistical purposes. You can disable the usage of cookies by changing the settings of your browser. By browsing our site without changing the browser settings you grant us permission to store that information on your device.

I agree, do not show this message again.

Benchmarking flow for testing sentiment coherence for information transfer models assessment

A. V. CHIROȘCA1,* , G. CHIROȘCA2,* , S. G. MUȘAT2

Affiliation

  1. Faculty of Physics, University of Bucharest, Atomiștilor 405, 077125, Magurele, Ilfov, Romania
  2. National Institute of Research and Development for Optoelectronics INOE 2000, 077125, Măgurele, Ilfov,, Romania

Abstract

During this work we developed a benchmarking flow for subtle elements propagation through standard machine learning application. We started from the Stable Diffusion image generation models, one of the most active and innovative tools for artistic creation. While we focus on sentiment coherence by measuring the way image generation and image captioning models relate to this information. The benchmarking system incorporates a custom model that analyzes textual input (using Recurrent Neural Networks) providing text classification for the six primary emotional states: joy, sadness, love, anger, fear, and surprise. The classified emotion is used to construct a semantically rich textual prompt, which conditions the generative model to produce imagery aligned with the affective context of the input. The resulting visual outputs aim to faithfully encapsulate the emotional nuances conveyed in the source text. While working on-premises, we also compared the results with LLM (Large Language Model) to better grasp the capabilities of diffusion models in image generation. Then the produced image is fed through an image captioner model (Efficient Net B0) that once again produces a text caption that is analyzed through the sentiment analysis network. This approach shows that sentiments are hard to grasp by modern machine learning models providing a coherence score of 21%, the rest of the results being, most likely provided by model hallucinations. This approach demonstrates potential for a flexible application that can be used in many fields, providing feedback towards the model coherence (as a benchmark tool) for reducing hallucinations thus allowing model applications in fields such as digital art creation, art-based therapeutic interventions, affective computing, and the design of emotionally responsive human-computer interfaces..

Keywords

AI art, Generative AI, Stable diffusion, Diffusion models, LLM, Sentiment analysis, Machine learning, Benchmarking, AI and creativity.

Submitted at: Oct. 6, 2025
Accepted at: Dec. 4, 2025

Citation

A. V. CHIROȘCA, G. CHIROȘCA, S. G. MUȘAT, Benchmarking flow for testing sentiment coherence for information transfer models assessment, Journal of Optoelectronics and Advanced Materials Vol. 27, Iss. 11-12, pp. 604-612 (2025)