Daniil Dorin

Open-source projects

2024

Garage: generative augmentation framework

GitHub

Classical augmentation changes pixels; Garage changes content. Given an image and a target object, it produces new training samples where that object is swapped for a plausible alternative, keeping the scene intact.

Garage pipeline: detect and segment the object, propose a replacement and prompt, inpaint.
Garage pipeline: detect and segment the object, propose a replacement and prompt, inpaint.

Pipeline

  1. Grounding DINO + Segment Anything locate the object from a text query and produce a mask.
  2. The Augmenter model generates candidate replacement objects and prompts.
  3. PowerPaint inpaints the masked region with the new object.

A Gradio demo lets you try it in the browser; the repository includes depth-conditioned examples and scripts to download the required checkpoints.