# PixArt-$\alpha$
FAST TRAINING OF DIFFUSION TRANSFORMER FOR PHOTOREALISTIC TEXT-TO-IMAGE
SYNTHESIS

* **Link:** [[pdf]](https://arxiv.org/pdf/2310.00426)
* **Authors:** Huawei
* **Comments:** ICLR 2024

## Introduction
> High resolution image synthesis up to 1024x2024 with low training cost

* Training stratgy decomposition
* Optimize pixel dependency
* Text-image alignment
* Image aesthetic quality
* Efficient T2I Transformer
* High-informative data

## Method
* **TRAINING STRATEGY DECOMPOSITION**
* Stage1: Pixel dependency learning
* Initialize DiT from ImageNet trained model: inexpensive, learn dependency between features.
* Stage2: Text-image alignment learning
* Construct a dataset consisting of precise text-image pairs
* Stage3: High-resolution and aesthetic image generation
* Fine-tune the model using high-quality aesthetic data for high-resolution image generation
* **DATASET CONSTRUCTION**
* Use LLaVA for captioning: better captions

* MODEL ARCHITECTURE

## Experiments

### Datasets

### Results

## Misc

# PixArt-$\delta$
FAST AND CONTROLLABLE IMAGE GENERATION WITH LATENT CONSISTENCY MODELS

* **Link:** [[pdf]](https://arxiv.org/pdf/2401.05252)
* **Authors:** Huawei
* **Comments:**

## Introduction

## Method

## Experiments

### Datasets

### Results

## Misc

# PixArt-$\Sigma$
Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation

* **Link:** [[pdf]](https://arxiv.org/pdf/2403.04692)
* **Authors:** Huawei
* **Comments:**

## Introduction

## Method

## Experiments

### Datasets

### Results

## Misc