kfirgold99/Inspiration-Seeds-Benchmark

Inspiration Seeds Benchmark

This is the evaluation benchmark for the paper "Inspiration Seeds: Learning Non-Literal Visual Combinations for Generative Exploration".

It contains 99 unique image pairs drawn from 6 visual categories (architecture, fashion, food, nature, other, sea), used to evaluate non-literal visual combination models.

[Paper] | [Project Page] | [GitHub]

Authors

• Kfir Goldberg
• Elad Richardson
• Yael Vinker

Abstract

While generative models have become powerful tools for image synthesis, they are typically optimized for executing carefully crafted textual prompts, offering limited support for the open-ended visual exploration that often precedes idea formation. In contrast, designers frequently draw inspiration from loosely connected visual references, seeking emergent connections that spark new ideas. We propose Inspiration Seeds, a generative framework that shifts image generation from final execution to exploratory ideation. Given two input images, our model produces diverse, visually coherent compositions that reveal latent relationships between inputs, without relying on user-specified text prompts. Our approach is feed-forward, trained on synthetic triplets of decomposed visual aspects derived entirely through visual means: we use CLIP Sparse Autoencoders to extract editing directions in CLIP latent space and isolate concept pairs. By removing the reliance on language and enabling fast, intuitive recombination, our method supports visual ideation at the early and ambiguous stages of creative work.

Dataset Structure

Each row contains an image pair and associated metadata:

Column	Type	Description
image_1	Image	First input image
image_2	Image	Second input image
image_1_name	string	Name of the first image (e.g. arch6)
image_2_name	string	Name of the second image (e.g. fashion1)
category_1	string	Category of the first image
category_2	string	Category of the second image

Categories: arch, fashion, food, nature, other, sea

Usage

from datasets import load_dataset

ds = load_dataset("kfirgold99/Inspiration-Seeds-Benchmark", split="test")
pair = ds[0]
pair["image_1"].show()  # PIL Image
pair["image_2"].show()

Citation

@misc{goldberg2026inspirationseedslearningnonliteral,
      title={Inspiration Seeds: Learning Non-Literal Visual Combinations for Generative Exploration},
      author={Kfir Goldberg and Elad Richardson and Yael Vinker},
      year={2026},
      eprint={2602.08615},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.08615},
}