Non_Representational_Art_AI

Dataset: Generative AI for Non-Representational Art — LoRA-Based Style Learning on FLUX This dataset accompanies the paper "Generative Artificial Intelligence for Non-Representational Art: LoRA-Based Style Learning on FLUX" (Conference'26). It contains all materials necessary to reproduce the quantitative evaluation and inspect the generative outputs presented in the study. Contents dataset_with_artist_grounded_annotations.zip — The curated training corpus of 127 digitized non-representational artworks by the artist, including LLM-assisted textual annotations generated via a reference-image-calibrated GPT-4o pipeline. generic_captions.zip — The alternative caption set used in the annotation ablation study, generated by GPT-4o without access to the artist's reference annotation. output_original_lora.zip — 127 images generated by the LoRA model trained on artist-grounded captions, one per prompt. output_generic_lora.zip — 127 images generated by the LoRA model trained on generic captions, one per prompt. output_baseline.zip — 127 images generated by the non-finetuned FLUX-dev baseline model using identical prompts, serving as the comparative reference. output_nano_banana.zip — 127 images generated by Nano Banana using identical prompts, included as a contemporary non-adapted baseline. artist_specific_LoRA.safetensors — The trained LoRA weights for the artist-grounded caption model, compatible with standard FLUX inference pipelines. generic_caption_LoRA.safetensors — The trained LoRA weights for the generic caption model, compatible with standard FLUX inference pipelines. Usage The LoRA weights can be loaded into any FLUX-dev compatible inference setup. Prompts should follow the annotation style of the training corpus for best stylistic alignment. A guidance scale of 1 and 50 inference steps are recommended, as used in the exhibition deployment.