Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/PixArt-alpha/PixArt-alpha
PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis
https://github.com/PixArt-alpha/PixArt-alpha
Last synced: 9 days ago
JSON representation
PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis
- Host: GitHub
- URL: https://github.com/PixArt-alpha/PixArt-alpha
- Owner: PixArt-alpha
- License: agpl-3.0
- Created: 2023-10-12T14:16:33.000Z (about 1 year ago)
- Default Branch: master
- Last Pushed: 2024-08-01T06:27:17.000Z (3 months ago)
- Last Synced: 2024-10-29T15:34:27.347Z (10 days ago)
- Language: Python
- Homepage: https://pixart-alpha.github.io/
- Size: 59.5 MB
- Stars: 2,765
- Watchers: 47
- Forks: 177
- Open Issues: 4
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- StarryDivineSky - PixArt-alpha/PixArt-alpha - α：用于逼真文本到图像合成的 Diffusion Transformer 的快速训练。提出了三个核心设计：（1） 训练策略分解：我们设计了三个不同的训练步骤，分别优化像素依赖性、文本图像对齐和图像审美质量;（2） 高效的 T2I Transformer：我们将交叉注意力模块整合到 Diffusion Transformer （DiT） 中，以注入文本条件并简化计算密集型类条件分支;（3） 信息量大的数据：我们强调文本-图像对中概念密度的重要性，并利用大型视觉-语言模型自动标记密集的伪标题，以辅助文本-图像对齐学习。因此，PIXART-α 的训练速度明显超过现有的大规模 T2I 模型，例如，PIXART-α 仅占用 Stable Diffusion v1.5 训练时间的 10.8%（~675 对 ~6,250 个 A100 GPU 日），节省了近 300,000 美元（26,000 美元对 320,000 美元），并减少了 90% 的二氧化碳排放。此外，与更大的 SOTA 模型 RAPHAEL 相比，我们的训练成本仅为 1%。大量实验表明，PIXART-α 在图像质量、艺术性和语义控制方面表现出色。 我们希望 PIXART-α 能为 AIGC 社区和初创公司提供新的见解，以加速从头开始构建自己的高质量、低成本的生成模型。PixArt-alpha/PixArt-LCM-XL-2-1024-MS 检查点， (其他_机器视觉 / 网络服务_其他)
README
###
👉 PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis
### ICLR 2024 Spotlight
---
This repo contains PyTorch model definitions, pre-trained weights and inference/sampling code for our paper exploring
Fast training diffusion models with transformers. You can find more visualizations on our [project page](https://pixart-alpha.github.io/).
**PixArt-α Community**: Join our PixArt-α discord channels
for discussions. Coders are welcome to contribute.
> [**PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis**](https://pixart-alpha.github.io/)
> [Junsong Chen*](https://lawrence-cj.github.io/), [Jincheng Yu*](https://lovesykun.cn/about.html),
> [Chongjian Ge*](https://chongjiange.github.io/), [Lewei Yao*](https://scholar.google.com/citations?user=hqDyTg8AAAAJ&hl=zh-CN&oi=ao),
> [Enze Xie](https://xieenze.github.io/)†,
> [Yue Wu](https://yuewuhkust.github.io/), [Zhongdao Wang](https://zhongdao.github.io/),
> [James Kwok](https://www.cse.ust.hk/~jamesk/), [Ping Luo](http://luoping.me/),
> [Huchuan Lu](https://scholar.google.com/citations?hl=en&user=D3nE0agAAAAJ),
> [Zhenguo Li](https://scholar.google.com/citations?user=XboZC1AAAAAJ)
>
Huawei Noah’s Ark Lab, Dalian University of Technology, HKU, HKUST
> [**PIXART-δ: Fast and Controllable Image Generation with Latent Consistency Models**](https://pixart-alpha.github.io/)
> [Junsong Chen](https://lawrence-cj.github.io/), [Yue Wu](https://yuewuhkust.github.io/), [Simian Luo](https://luosiallen.github.io/), [Enze Xie](https://xieenze.github.io/)†,
> [Sayak Paul](https://sayak.dev/), [Ping Luo](http://luoping.me/), [Hang Zhao](), [Zhenguo Li](https://scholar.google.com/citations?user=XboZC1AAAAAJ)
>
Huawei Noah’s Ark Lab, DLUT, Tsinghua University, HKU, Hugging Face
---
## Breaking News 🔥🔥!!
- (🔥 New) Apr. 12, 2024. 💥 A better version of [PixArt-Σ](https://github.com/PixArt-alpha/PixArt-sigma) training & inference code, checkpoints are all released!!!
Welcome to collaborate and contribute. Star 🌟us if you think it is helpful!!
- (🔥 New) Jan. 19, 2024. 💥 [PixArt-δ](https://arxiv.org/abs/2401.05252) ControlNet [app_controlnet.py](app/app_controlnet.py) and [Checkpoint](https://huggingface.co/PixArt-alpha/PixArt-ControlNet/tree/main) are released!!!
- (🔥 New) Jan. 16, 2024. 💥 Glad to announce that [PixArt-α](https://arxiv.org/abs/2310.00426) is accepted by ICLR 2024 (Spotlight).
- (🔥 New) Dec. 17, 2023. 💥 PixArt supports [ComfyUI](https://github.com/comfyanonymous/ComfyUI#manual-install-windows-linux). Thanks to [@city96](https://github.com/city96/ComfyUI_ExtraModels) with his great work.
- (🔥 New) Nov. 30, 2023. 💥 PixArt collaborates with [LCMs](https://github.com/luosiallen/latent-consistency-model) team to make the **fastest** [Training & Inference Text-to-Image Generation System](https://github.com/PixArt-alpha/PixArt-alpha).
Here, [Training code](train_scripts/train_pixart_lcm.py) & [Inference code](scripts/inference_lcm.py) & [Weights](https://huggingface.co/PixArt-alpha/PixArt-LCM-XL-2-1024-MS) & [HF Demo](https://huggingface.co/spaces/PixArt-alpha/PixArt-LCM) [OpenXLab Demo](https://openxlab.org.cn/apps/detail/houshaowei/PixArt-LCM) are all released, we hope users will enjoy them.
Detailed **inference speed** and **code guidance** can be found in [docs](asset/docs/pixart_lcm.md). At the same time, we update the codebase for better user experience and fix some bugs in the newest version.
---
## 🚩 **New Features/Updates**
- ✅ Jan. 11, 2024. 💥 [PixArt-δ](https://arxiv.org/abs/2401.05252): We are excited to announce the release of the [PixArt-δ](https://arxiv.org/abs/2401.05252) technical report!!!
This report offers valuable insights into the training of LCM and ControlNet-like modules in Transformer Models. Along with the report, we have also released all the training and inference code for LCM & ControlNet [in this repository](https://github.com/PixArt-alpha/PixArt-alpha).
We encourage you to try them out and warmly welcome any Pull Requests from our users. Your contributions and feedback are highly appreciated!
- ✅ Feb. 07, 2024. [train_diffusers.py](train_scripts/train_diffusers.py) can directly train with diffusers model and visualize during training.
- ✅ Jan. 26, 2024. 💥 All checkpoints of [PixArt-α](https://github.com/PixArt-alpha/PixArt-alpha), including 256px checkpoints are all available here [Download Models](#-download-models).
- ✅ Jan. 19, 2024. 💥 [PixArt-δ](https://arxiv.org/abs/2401.05252) ControlNet [app_controlnet.py](app/app_controlnet.py) and [Checkpoint](https://huggingface.co/PixArt-alpha/PixArt-ControlNet/tree/main) is released!!!
- ✅ Jan. 12, 2024. 💥 We release the [SAM-LLaVA-Captions](https://huggingface.co/datasets/PixArt-alpha/SAM-LLaVA-Captions10M) used in PixArt-α training.
- ✅ Dec. 27, 2023. [PixArt-α](https://github.com/PixArt-alpha/PixArt-alpha) incorporates into [ControlLLM](https://github.com/OpenGVLab/ControlLLM)!
- ✅ Dec. 17, 2023. [PixArt-LCM-Lora](train_scripts/train_pixart_lcm_lora.py) & [PixArt-Lora](train_scripts/train_pixart_lora_hf.py) training scripts in Hugging Face style is released.
- ✅ Dec. 13, 2023. Add multi-scale vae feature extraction in [tools/extract_features.py](https://github.com/PixArt-alpha/PixArt-alpha/blob/3b4f0afdbe39def80b41ab05c664c963edeebbcd/tools/extract_features.py#L276).
- ✅ Dec. 01, 2023. Add a [Notebook folder](./notebooks) to help users get started with PixArt quickly! Thanks to [@kopyl](https://github.com/kopyl) for his contribution!
- ✅ Nov. 27, 2023. 💥 **PixArt-α Community**: Join our PixArt-α discord channels
for discussions. Coders are welcome to contribute.
- ✅ Nov. 21, 2023. 💥 [SA-Sovler](https://arxiv.org/abs/2309.05019) official code first release [here](asset/docs/sasolver.md).
- ✅ Nov. 19, 2023. Release `PixArt + Dreambooth` training scripts.
- ✅ Nov. 16, 2023. Diffusers support `random resolution` and `batch images` generation now. Besides,
running `Pixart` in under 8GB GPU VRAM is available in 🧨 [diffusers](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pixart).
- ✅ Nov. 10, 2023. Support DALL-E 3 Consistency Decoder in 🧨 diffusers.
- ✅ Nov. 06, 2023. Release pretrained weights with 🧨 diffusers integration, Hugging Face demo, and Google Colab example.
- ✅ Nov. 03, 2023. Release the LLaVA-captioning inference code.
- ✅ Oct. 27, 2023. Release the training & feature extraction code.
- ✅ Oct. 20, 2023. Collaborate with Hugging Face & Diffusers team to co-release the code and weights. (plz stay tuned.)
- ✅ Oct. 15, 2023. Release the inference code.
---
## Contents
* [Training](#-how-to-train)
* [Inference](#-how-to-test)
* [Download Models](#-download-models)
* [Use diffusers](#1---using-in--diffusers)
* [Data Processing](#-how-to-extract-t5-and-vae-features)
* [PixArt-**α** Demo](#3---gradio-with-diffusers--faster-)
* [PixArt-**α** 8GB VRAM](asset/docs/pixart.md)
* [PixArt-**δ** (LCM)](asset/docs/pixart_lcm.md)
* [PixArt-**δ** (ControlNet)](asset/docs/pixart_controlnet.md)
* [PixArt-**δ** (Dreambooth)](asset/docs/pixart-dreambooth.md)
* [Acknowledgement](#acknowledgements)
* [Citation](#bibtex)
* [PixArt-**Σ** Releasing](https://github.com/PixArt-alpha/PixArt-sigma)
---
## 🐱 Abstract
TL; DR: PixArt-α is a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators (e.g., Imagen, SDXL, and even Midjourney), and the training speed markedly surpasses existing large-scale T2I models, e.g., PixArt-α only takes 10.8% of Stable Diffusion v1.5's training time (675 vs. 6,250 A100 GPU days).
CLICK for the full abstract
The most advanced text-to-image (T2I) models require significant training costs (e.g., millions of GPU hours),
seriously hindering the fundamental innovation for the AIGC community while increasing CO2 emissions.
This paper introduces PixArt-α, a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators (e.g., Imagen, SDXL, and even Midjourney),
reaching near-commercial application standards. Additionally, it supports high-resolution image synthesis up to 1024px resolution with low training cost.
To achieve this goal, three core designs are proposed:
(1) Training strategy decomposition: We devise three distinct training steps that separately optimize pixel dependency, text-image alignment, and image aesthetic quality;
(2) Efficient T2I Transformer: We incorporate cross-attention modules into Diffusion Transformer (DiT) to inject text conditions and streamline the computation-intensive class-condition branch;
(3) High-informative data: We emphasize the significance of concept density in text-image pairs and leverage a large Vision-Language model to auto-label dense pseudo-captions to assist text-image alignment learning.
As a result, PixArt-α's training speed markedly surpasses existing large-scale T2I models,
e.g., PixArt-α only takes 10.8% of Stable Diffusion v1.5's training time (675 vs. 6,250 A100 GPU days),
saving nearly $300,000 ($26,000 vs. $320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL,
our training cost is merely 1%. Extensive experiments demonstrate that PixArt-α excels in image quality, artistry, and semantic control.
We hope PixArt-α will provide new insights to the AIGC community and startups to accelerate building their own high-quality yet low-cost generative models from scratch.
---
---
# 🔥🔥🔥 Why PixArt-α?
## Training Efficiency
PixArt-α only takes 12% of Stable Diffusion v1.5's training time (753 vs. 6,250 A100 GPU days), saving nearly $300,000 ($28,000 vs. $320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL, our training cost is merely 1%.
| Method | Type | #Params | #Images| FID-30K ↓ | A100 GPU days |
|-----------|------|---------|--------|------------------|---------------|
| DALL·E | Diff | 12.0B | 250M | 27.50 | |
| GLIDE | Diff | 5.0B | 250M | 12.24 | |
| LDM | Diff | 1.4B | 400M | 12.64 | |
| DALL·E 2 | Diff | 6.5B | 650M | 10.39 | 41,66 |
| SDv1.5 | Diff | 0.9B | 2000M | 9.62 | 6,250 |
| GigaGAN | GAN | 0.9B | 2700M | 9.09 | 4,783 |
| Imagen | Diff | 3.0B | 860M | 7.27 | 7,132 |
| RAPHAEL | Diff | 3.0B | 5000M+ | 6.61 | 60,000 |
| PixArt-α | Diff | 0.6B | 25M | 7.32 (zero-shot) | 753 |
| PixArt-α | Diff | 0.6B | 25M | 5.51 (COCO FT) | 753 |
## Inference Efficiency
PIXART-δ successfully generates **1024x1024 high resolution** images within **0.5 seconds** on an A100. With the implementation
of 8-bit inference technology, PIXART-δ requires **less than 8GB of GPU VRAM**.
Let us stress again how liberating it is to explore image generation so easily with PixArt-LCM.
| Hardware | PIXART-δ (4 steps) | SDXL LoRA LCM (4 steps) | PixArt-α (14 steps) | SDXL standard (25 steps) |
|-----------------------------|--------------------|-------------------------|---------------------|---------------------------|
| T4 (Google Colab Free Tier) | 3.3s | 8.4s | 16.0s | 26.5s |
| V100 (32 GB) | 0.8s | 1.2s | 5.5s | 7.7s |
| A100 (80 GB) | 0.51s | 1.2s | 2.2s | 3.8s |
These tests were run with a batch size of 1 in all cases.
For cards with a lot of capacity, such as A100, performance increases significantly when generating multiple images at once, which is usually the case for production workloads.
## High-quality Generation from PixArt-α
- More samples
- PixArt + [Dreambooth](https://dreambooth.github.io/)
- PixArt + [ControlNet](https://github.com/lllyasviel/ControlNet)
# 🔧 Dependencies and Installation
- Python >= 3.9 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))
- [PyTorch >= 1.13.0+cu11.7](https://pytorch.org/)
```bash
conda create -n pixart python=3.9
conda activate pixart
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
git clone https://github.com/PixArt-alpha/PixArt-alpha.git
cd PixArt-alpha
pip install -r requirements.txt
```
# ⏬ Download Models
All models will be automatically downloaded. You can also choose to download manually from this [url](https://huggingface.co/PixArt-alpha/PixArt-alpha).
| Model | #Params | url | Download in OpenXLab |
|:----------------------------|:--------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:----------------------------------------------------------------------------------------------------------------|
| T5 | 4.3B | [T5](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) | [T5](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/t5-v1_1-xxl.zip) |
| VAE | 80M | [VAE](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/sd-vae-ft-ema) | [VAE](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/sd-vae-ft-ema.zip) |
| PixArt-α-SAM-256 | 0.6B | [PixArt-XL-2-SAM-256x256.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-SAM-256x256.pth) or [diffusers version](https://huggingface.co/PixArt-alpha/PixArt-XL-2-SAM-256x256) | [256-SAM](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/PixArt-XL-2-SAM-256x256.pth) |
| PixArt-α-256 | 0.6B | [PixArt-XL-2-256x256.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-256x256.pth) or [diffusers version](https://huggingface.co/PixArt-alpha/PixArt-XL-2-256x256) | [256](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/PixArt-XL-2-256x256.pth) |
| PixArt-α-256-MSCOCO-FID7.32 | 0.6B | [PixArt-XL-2-256x256.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-256x256-MSCOCO-FID732.pth) | [256]() |
| PixArt-α-512 | 0.6B | [PixArt-XL-2-512x512.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-512x512.pth) or [diffusers version](https://huggingface.co/PixArt-alpha/PixArt-XL-2-512x512) | [512](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/PixArt-XL-2-512x512.pth) |
| PixArt-α-1024 | 0.6B | [PixArt-XL-2-1024-MS.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/resolve/main/PixArt-XL-2-1024-MS.pth) or [diffusers version](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS) | [1024](https://download.openxlab.org.cn/models/PixArt-alpha/PixArt-alpha/weight/PixArt-XL-2-1024-MS.pth) |
| PixArt-δ-1024-LCM | 0.6B | [diffusers version](https://huggingface.co/PixArt-alpha/PixArt-LCM-XL-2-1024-MS) | |
| ControlNet-HED-Encoder | 30M | [ControlNetHED.pth](https://huggingface.co/PixArt-alpha/PixArt-alpha/blob/main/ControlNetHED.pth) | |
| PixArt-δ-512-ControlNet | 0.9B | [PixArt-XL-2-512-ControlNet.pth](https://huggingface.co/PixArt-alpha/PixArt-ControlNet/tree/main) | [512](https://openxlab.org.cn/models/detail/PixArt-alpha/PixArt-ControlNet) |
| PixArt-δ-1024-ControlNet | 0.9B | [PixArt-XL-2-1024-ControlNet.pth](https://huggingface.co/PixArt-alpha/PixArt-ControlNet/tree/main) | [1024](https://openxlab.org.cn/models/detail/PixArt-alpha/PixArt-ControlNet) |
ALSO find all models in [OpenXLab_PixArt-alpha](https://openxlab.org.cn/models/detail/PixArt-alpha/PixArt-alpha)
# 🔥 How to Train
## 1. PixArt Training
**First of all.**
Thanks to [@kopyl](https://github.com/kopyl), you can reproduce the full fine-tune training flow on [Pokemon dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions) from HugginFace with notebooks:
1. Train with [notebooks/train.ipynb](https://github.com/PixArt-alpha/PixArt-alpha/blob/53dac066f60fe5fdbdde4f0360145ca96d4cc38c/notebooks/train.ipynb).
2. Convert to Diffusers with [notebooks/convert-checkpoint-to-diffusers.ipynb](https://github.com/PixArt-alpha/PixArt-alpha/blob/master/notebooks/convert-checkpoint-to-diffusers.ipynb).
3. Run the inference with converted checkpoint in step 2 with [notebooks/infer.ipynb](https://github.com/PixArt-alpha/PixArt-alpha/blob/master/notebooks/infer.ipynb).
**Then, for more details.**
Here we take SAM dataset training config as an example, but of course, you can also prepare your own dataset following this method.
You **ONLY** need to change the **config** file in [config](./configs/pixart_config) and **dataloader** in [dataset](./diffusion/data/datasets).
```bash
python -m torch.distributed.launch --nproc_per_node=2 --master_port=12345 train_scripts/train.py configs/pixart_config/PixArt_xl2_img256_SAM.py --work-dir output/train_SAM_256
```
The directory structure for SAM dataset is:
```
cd ./data
SA1B
├──images/ (images are saved here)
│ ├──sa_xxxxx.jpg
│ ├──sa_xxxxx.jpg
│ ├──......
├──captions/ (corresponding captions are saved here, same name as images)
│ ├──sa_xxxxx.txt
│ ├──sa_xxxxx.txt
├──partition/ (all image names are stored txt file where each line is a image name)
│ ├──part0.txt
│ ├──part1.txt
│ ├──......
├──caption_feature_wmask/ (run tools/extract_caption_feature.py to generate caption T5 features, same name as images except .npz extension)
│ ├──sa_xxxxx.npz
│ ├──sa_xxxxx.npz
│ ├──......
├──img_vae_feature/ (run tools/extract_img_vae_feature.py to generate image VAE features, same name as images except .npy extension)
│ ├──train_vae_256/
│ │ ├──noflip/
│ │ │ ├──sa_xxxxx.npy
│ │ │ ├──sa_xxxxx.npy
│ │ │ ├──......
```
**Here we prepare data_toy for better understanding**
```bash
cd ./data
git lfs install
git clone https://huggingface.co/datasets/PixArt-alpha/data_toy
```
Then,
[Here](https://huggingface.co/datasets/PixArt-alpha/data_toy/blob/main/part0.txt) is an example of partition/part0.txt file.
---
Besides, for json file guided [training](https://github.com/PixArt-alpha/PixArt-alpha/blob/fe0cb78065d64c18ecd8955a04e4f29138d47946/configs/pixart_config/PixArt_xl2_img1024_internalms.py#L3C2-L3C2),
[here](https://huggingface.co/datasets/PixArt-alpha/data_toy/blob/main/data_info.json) is a toy json file for better understand.
---
## 2. PixArt + DreamBooth Training
Following the `Pixart + DreamBooth` [training guidance](asset/docs/pixart-dreambooth.md)
## 3. PixArt + LCM / LCM-LoRA Training
Following the `PixArt + LCM` [training guidance](asset/docs/pixart_lcm.md)
## 4. PixArt + ControlNet Training
Following the `PixArt + ControlNet` [training guidance](asset/docs/pixart_controlnet.md)
## 4. PixArt + LoRA Training
```bash
pip install peft==0.6.2
accelerate launch --num_processes=1 --main_process_port=36667 train_scripts/train_pixart_lora_hf.py --mixed_precision="fp16" \
--pretrained_model_name_or_path=PixArt-alpha/PixArt-XL-2-1024-MS \
--dataset_name=lambdalabs/pokemon-blip-captions --caption_column="text" \
--resolution=1024 --random_flip \
--train_batch_size=16 \
--num_train_epochs=200 --checkpointing_steps=100 \
--learning_rate=1e-06 --lr_scheduler="constant" --lr_warmup_steps=0 \
--seed=42 \
--output_dir="pixart-pokemon-model" \
--validation_prompt="cute dragon creature" --report_to="tensorboard" \
--gradient_checkpointing --checkpoints_total_limit=10 --validation_epochs=5 \
--rank=16
```
# 💻 How to Test
Inference requires at least `23GB` of GPU memory using this repo, while `11GB and 8GB` using in 🧨 [diffusers](#using-in--diffusers).
Currently support:
- [x] [IDDPM](https://arxiv.org/abs/2102.09672)
- [x] [DPM-Solver](https://arxiv.org/abs/2206.00927)
- [x] [SA-Solver](https://arxiv.org/abs/2309.05019)
- [ ] [DPM-Solver-v3](https://arxiv.org/abs/2310.13268v2)
## 1. Quick start with [Gradio](https://www.gradio.app/guides/quickstart)
To get started, first install the required dependencies. Make sure you've downloaded the [models](https://huggingface.co/PixArt-alpha/PixArt-alpha) to the output/pretrained_models folder, and then run on your local machine:
```bash
DEMO_PORT=12345 python app/app.py
```
As an alternative, a sample [Dockerfile](Dockerfile) is provided to make a runtime container that starts the Gradio app.
```bash
docker build . -t pixart
docker run --gpus all -it -p 12345:12345 -v :/root/.cache/huggingface pixart
```
Or use docker-compose. Note, if you want to change context from the 1024 to 512 or LCM version of the app just change the APP_CONTEXT env variable in the docker-compose.yml file. The default is 1024
```bash
docker compose build
docker compose up
```
Let's have a look at a simple example using the `http://your-server-ip:12345`.
## 2. Integration in diffusers
### 1). Using in 🧨 diffusers
Make sure you have the updated versions of the following libraries:
```bash
pip install -U transformers accelerate diffusers SentencePiece ftfy beautifulsoup4
```
And then:
```python
import torch
from diffusers import PixArtAlphaPipeline, ConsistencyDecoderVAE, AutoencoderKL
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too.
pipe = PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16, use_safetensors=True)
# If use DALL-E 3 Consistency Decoder
# pipe.vae = ConsistencyDecoderVAE.from_pretrained("openai/consistency-decoder", torch_dtype=torch.float16)
# If use SA-Solver sampler
# from diffusion.sa_solver_diffusers import SASolverScheduler
# pipe.scheduler = SASolverScheduler.from_config(pipe.scheduler.config, algorithm_type='data_prediction')
# If loading a LoRA model
# transformer = Transformer2DModel.from_pretrained("PixArt-alpha/PixArt-LCM-XL-2-1024-MS", subfolder="transformer", torch_dtype=torch.float16)
# transformer = PeftModel.from_pretrained(transformer, "Your-LoRA-Model-Path")
# pipe = PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-LCM-XL-2-1024-MS", transformer=transformer, torch_dtype=torch.float16, use_safetensors=True)
# del transformer
# Enable memory optimizations.
# pipe.enable_model_cpu_offload()
pipe.to(device)
prompt = "A small cactus with a happy face in the Sahara desert."
image = pipe(prompt).images[0]
image.save("./catcus.png")
```
Check out the [documentation](./asset/docs/sasolver.md) for more information about SA-Solver Sampler.
This integration allows running the pipeline with a batch size of 4 under 11 GBs of GPU VRAM.
Check out the [documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pixart) to learn more.
### 2). Running the `PixArtAlphaPipeline` in under 8GB GPU VRAM
GPU VRAM consumption under 8 GB is supported now, please refer to [documentation](asset/docs/pixart.md) for more information.
### 3). Gradio with diffusers (Faster)
To get started, first install the required dependencies, then run on your local machine:
```bash
# diffusers version
DEMO_PORT=12345 python app/app.py
```
Let's have a look at a simple example using the `http://your-server-ip:12345`.
You can also click [here](https://colab.research.google.com/drive/1jZ5UZXk7tcpTfVwnX33dDuefNMcnW9ME?usp=sharing) to have a free trial on Google Colab.
### 4). Convert .pth checkpoint into diffusers version
```bash
python tools/convert_pixart_alpha_to_diffusers.py --image_size your_img_size --multi_scale_train (True if you use PixArtMS else False) --orig_ckpt_path path/to/pth --dump_path path/to/diffusers --only_transformer=True
```
## 3. Online Demo [](https://huggingface.co/spaces/PixArt-alpha/PixArt-alpha)
# ✏️ How to LLaVA captioning
Thanks to the code base of [LLaVA-Lightning-MPT](https://huggingface.co/liuhaotian/LLaVA-Lightning-MPT-7B-preview),
we can caption the LAION and SAM dataset with the following launching code:
```bash
python tools/VLM_caption_lightning.py --output output/dir/ --data-root data/root/path --index path/to/data.json
```
We present auto-labeling with custom prompts for LAION (left) and SAM (right). The words highlighted in green represent the original caption in LAION, while those marked in red indicate the detailed captions labeled by LLaVA.
# ✏️ How to extract T5 and VAE features
Prepare T5 text feature and VAE image feature in advance will speed up the training process and save GPU memory.
```bash
python tools/extract_features.py --img_size=1024 \
--json_path "data/data_info.json" \
--t5_save_root "data/SA1B/caption_feature_wmask" \
--vae_save_root "data/SA1B/img_vae_features" \
--pretrained_models_dir "output/pretrained_models" \
--dataset_root "data/SA1B/Images/"
```
## 💪To-Do List (Congratulations🎉)
- [x] Inference code
- [x] Training code
- [x] T5 & VAE feature extraction code
- [x] LLaVA captioning code
- [x] Model zoo
- [x] Diffusers version & Hugging Face demo
- [x] Google Colab example
- [x] DALLE3 VAE integration
- [x] Inference under 8GB GPU VRAM with diffusers
- [x] Dreambooth Training code
- [x] SA-Solver code
- [x] PixArt-α-LCM will release soon
- [x] Multi-scale vae feature extraction code
- [x] PixArt-α-LCM-LoRA scripts will release soon
- [x] PixArt-α-LoRA training scripts will release soon
- [x] ControlNet code will be released
- [x] SAM-LLaVA caption dataset
- [x] ControlNet checkpoint
- [x] 256px pre-trained models
- [x] PixArt-Σ: Next version model with much better ability is training!
# Other Source
We make a video comparing PixArt with current most powerful Text-to-Image models.
[](https://www.youtube.com/watch?v=7_6KsIITgWY)
# 📖BibTeX
@misc{chen2023pixartalpha,
title={PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis},
author={Junsong Chen and Jincheng Yu and Chongjian Ge and Lewei Yao and Enze Xie and Yue Wu and Zhongdao Wang and James Kwok and Ping Luo and Huchuan Lu and Zhenguo Li},
year={2023},
eprint={2310.00426},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@misc{chen2024pixartdelta,
title={PIXART-{\delta}: Fast and Controllable Image Generation with Latent Consistency Models},
author={Junsong Chen and Yue Wu and Simian Luo and Enze Xie and Sayak Paul and Ping Luo and Hang Zhao and Zhenguo Li},
year={2024},
eprint={2401.05252},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
# 🤗Acknowledgements
- Thanks to [Diffusers](https://github.com/huggingface/diffusers) for their wonderful technical support and awesome collaboration!
- Thanks to [Hugging Face](https://github.com/huggingface) for sponsoring the nicely demo!
- Thanks to [DiT](https://github.com/facebookresearch/DiT) for their wonderful work and codebase!
## Star History
[](https://star-history.com/#PixArt-alpha/PixArt-alpha&Date)