TrendMorph: A Multimodal Generative AI Pipeline for Real-Time Trend-Based Content Generation
Content creators face a persistent gap between spotting a trend and producing something usable from it: trendtracking tools such as Google Trends report what is popular but do not generate content, while generic AI writing and image tools generate content without any awareness of what is currently trending. We present TrendMorph, a system that connects the two by combining real-time trend collection from YouTube, X (Twitter), Reddit, and Pinterest with a multimodal generative pipeline for captions, hashtags, and images. Captions are produced by pairing BLIP image understanding with a fine-tuned T5-small text generator; images are produced wit
Resource information
- Authors
- Pandey, Rajesh, Pandit, Ravi, Niraula, Ritika, Chaudhary, Sony Kumari
- Research area
- Artificial intelligence
- Research topics
- Generative AI, Stable Diffusion XL, LoRA, BLIP, T5, Multimodal AI, Trend Detection, Content Generation
- Publication date
- 2026
- DOI
- 10.5281/zenodo.22247715
- Upstream last modified
- Sep 2, 2026
- License
- Creative Commons Attribution 4.0 International
- Last record update
- Nov 6, 2026
About this resource
Content creators face a persistent gap between spotting a trend and producing something usable from it: trendtracking tools such as Google Trends report what is popular but do not generate content, while generic AI writing and image tools generate content without any awareness of what is currently trending. We present TrendMorph, a system that connects the two by combining real-time trend collection from YouTube, X (Twitter), Reddit, and Pinterest with a multimodal generative pipeline for captions, hashtags, and images. Captions are produced by pairing BLIP image understanding with a fine-tuned T5-small text generator; images are produced with a LoRAadapted Stable Diffusion XL (SDXL) pipeline. On a held-out caption evaluation set, the fine-tuned captioning model reached a BERTScore F1 of 0.873 despite modest n-gram overlap (BLEU 0.123), which we attribute to the open-ended nature of social media captions, where many different phrasings can be equally valid. The image generation pipeline achieved an average CLIP score of 33.69 across four qualitative test cases spanning surrealism, composition, stylized environments, and mechanical design. We describe the system architecture, training configuration, and evaluation results, and discuss the practical limitations of the evaluation, including dataset size and the absence of human judgment, as directions for follow-up work.Keywords: Generative AI, Stable Diffusion XL, LoRA, BLIP, T5, Multimodal AI, Trend Detection, Content Generation.