Jacklu0831/few-shot-creative-generation-8

FSCG-8: Few-Shot Creative Generation

FSCG-8 is the few-shot creative generation benchmark introduced with ProCreate, Don't Reproduce! Propulsive Energy Diffusion for Creative Generation (ECCV 2024). It has eight categories, each with 50 caption-image pairs, for fine-tuning a text-to-image model on a few examples and measuring whether it generates novel images in the same style instead of replicating the training images.

Paper · Project page · Code

Categories

Each category is one split with 50 examples (image, text):

Split	Contents	Fine-tuned checkpoint
amedeo_modigliani	Amedeo Modigliani paintings	procreate-diffusion-amedeo-modigliani
apple	Apple product designs	procreate-diffusion-apple
burberry	Burberry designs	procreate-diffusion-burberry
frank_gehry	Frank Gehry architecture	procreate-diffusion-frank-gehry
nouns	Nouns characters	procreate-diffusion-nouns
one_piece	One Piece characters	procreate-diffusion-one-piece
pokemon	Pokemon	procreate-diffusion-pokemon
rococo	Rococo style art and furniture	procreate-diffusion-rococo

Usage

from datasets import load_dataset

ds = load_dataset("Jacklu0831/few-shot-creative-generation-8", split="pokemon")
print(ds[0]["text"])

Rights

The images belong to their respective copyright holders and are provided for research on few-shot generation.

Citation

@inproceedings{lu2024procreate,
  title     = {ProCreate, Don't Reproduce! Propulsive Energy Diffusion for Creative Generation},
  author    = {Lu, Jack and Teehan, Ryan and Ren, Mengye},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2024}
}