---
license: apache-2.0
datasets:
- multi-train/coco_captions_1107
- visual_genome
language:
- en
pipeline_tag: text-to-image
tags:
- scene_graph
- transformers
- laplacian
- autoregressive
- vqvae
---

# trf-sg2im

Model card for the paper __"[Transformer-Based Image Generation from Scene Graphs](https://arxiv.org/abs/2303.04634)"__.
Original GitHub implementation [here](https://github.com/perceivelab/trf-sg2im).

![teaser](docs/teaser.gif)

## Model

This model is a two-stage scene-graph-to-image approach. It takes a scene graph as input and generates a layout using a transformer-based architecture with Laplacian Positional Encoding.
Then, it uses this estimated layout to condition an autoregressive GPT-like transformer to compose the image in the latent, discrete space, converted into the final image by a VQVAE.

![architecture](docs/architecture.png)

## Usage
For usage instructions, please refer to the original [GitHub repo](https://github.com/perceivelab/trf-sg2im).

## Results

Comparison with other state-of-the-art approaches

![results](docs/results.png)