A Novel Method for Reducing Hallucinations in Artificial Intelligence-Generated Images Using Reinforcement Learning with Artificial Intelligence Feedback

Abstract

Image synthesis has been significantly enhanced by generative AI models particularly diffusion-based frameworks but they frequently produce irrational artefacts known as hallucinations. And in order to increase the realism and fidelity of Artificial Intelligence-generated images that’s why this paper proposes a reinforcement learning from AI feedback (RLAIF)-inspired pipelin in which Artificial Intelligence-generated feedback (visual-language captioning plus rule-based hallucination detection) guides iterative prompt refinement for image-to-image diffusion, without an explicit trainable reward model in the current prototype. By combining textual refinement from sophisticated language models such as (DeepSeek and Phi-1) those automated hallucination detection using visual-language models and image enhancement using the Stable Diffusion model this work methodically has addressed common artefacts such as distorted text and anatomical inaccuracies. A quantitative analysis using Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS) reveals that Phi-based prompt modification attains the minimum LPIPS values in both the testing environments (0.38 for the landscape scenario and 0.58 for the factory scenario) and the minimum FID value in the factory scenario (282.98, as opposed to 283.71 for the base-line model Stable Diffusion and 389.70 for DeepSeek-based prompt refinement). This translates into an FID reduction of up to 27.4% and an LPIPS reduction of up to 4.9% with respect to the less effective baseline model, showing that AI-based feedback helps minimize hallucinations.