Spaces:
Running
on
Zero
Apply for community grant: Academic project (gpu)
Hi,
We are requesting a GPU grant for our model demo. We are the first to showcase that generative models (i.e. Stable Diffusion, MAE) can be easily adapted to segment objects. We finetuned our model on a limited set of object categories (indoor furnishings and cars), yet both models generalize to unseen object categories and styles (i.e. X-rays, animals in art, etc). Interestingly, for MAE this is outside the pretraining distribution too. This suggests generative models have learned an inherent perceptual grouping mechanism. We hope that our findings will inspire more research into the representations learned by generative pretraining, and how they can be adapted for perceptual tasks.
Please see our website for high-resolution qualitative comparisons. We would ideally prefer a GPU that can run Stable Diffusion.
Paper link: https://arxiv.org/abs/2505.15263
Website: https://reachomk.github.io/gen2seg/
