Inspiration Seeds

Learning Non-Literal Visual Combinations for Generative Exploration

Kfir Goldberg1, Elad Richardson2, Yael Vinker3   1Bria   2Runway   3MIT

Given two input images, Inspiration Seeds produces diverse, visually coherent compositions that reveal latent relationships between the inputs — no text prompt required. The model is a LoRA on top of FLUX.1 Kontext [dev], trained on synthetic triplets of decomposed visual aspects derived via CLIP Sparse Autoencoders.

Upload two images (or pick an example below), then click Generate to sample several diverse fused compositions.

Note: black-forest-labs/FLUX.1-Kontext-dev is a gated model. This Space works because the hosting account has accepted its license.

1 10
1 8
4 40
Example pairs

Citation

@article{goldberg2026inspirationseeds,
  title   = {Inspiration Seeds: Learning Non-Literal Visual Combinations for Generative Exploration},
  author  = {Goldberg, Kfir and Richardson, Elad and Vinker, Yael},
  journal = {arXiv preprint arXiv:2602.08615},
  year    = {2026}
}