Home

Awesome

TITAN: Large-Scale Visual ObjecT DIscovery Through Text Attention using StAble DiffusioN

TITAN is an All-In-One destination if you are willing to create a synthetic segmentation/object annotated dataset if you have access to only prompts! The entire pipeline is very intuitive and you can have your dataset ready with less than 30 lines of code!

It relies on Stable Diffusion/Diffusion Models and Diffusion Attentive Attribution Map.

child annotationchairlift annotation
child annotationchairlift annotation

Getting Started

First, install PyTorch for your platform. You may also check the Colab Tutorial.

Installation

The following steps are required for setting up the titan package. The instructions are made keeping in mind the Colab environment for ease of understanding. Feel free to adapt it to work on your local machine/cloud server.

pip install daam==0.0.12
pip install git+https://github.com/RishiDarkDevil/TITAN.git

Using TITAN as a Library

Import and use TITAN as follows:

# For Stable Diffusion
from diffusers import StableDiffusionPipeline

# For Heatmap Generation
import daam

# For TITAN workflow
from titan import *

The several parts of a Data Generation Pipeline supported by TITAN:

Citation

Add Link to this Repo for citation purpose.