zai-org/CogView

zai-org/CogView

Description: Text-to-Image generation. The repo for NeurIPS 2021 paper "CogView: Mastering Text-to-Image Generation via Transformers".

Language: Python

License: Apache-2.0

Stars: 1797

Forks: 176

Open issues: 19

Created: 2021-05-25T14:48:31Z

Pushed: 2023-09-25T04:07:19Z

Default branch: main

Fork: no

Archived: no

README:

Generate vivid Images for Any (Chinese) text


News! The paper of ImageReward is accepted by NeurIPS 2023!

News! The codes of ImageReward (paper link) have been released at https://github.com/THUDM/ImageReward! ImageReward is the first general-purpose text-to-image human preference RM.

News! The codes of CogView2 (paper link) have been released at https://github.com/THUDM/CogView2!

News! The demo for a better and faster CogView2 (formal version, March 2022) is available! The lastest model also supports English input, but to translate them into Chinese often could be better.

News! The demo for a better and faster CogView2 (new version) is available!

News! The paper of CogView is accepted by NeurIPS 2021!

CogView is a pretrained (4B-param) transformer for text-to-image generation in general domain.

• Read our paper CogView: Mastering Text-to-Image Generation via Transformers on ArXiv for a formal introduction. The *PB-relax* and *Sandwich-LN* can also help you train large and deep transformers stably (e.g. eliminating NaN losses).
• Visit our demo at Github Page or Wudao! (Without post-selection or super-resolution, currently only supports simplified Chinese input, but one can translate text from other languages into Chinese for input. Note: *Wudao* provides faster access for users from China mainland.)
• Download our pretrained models from Tsinghua Cloud.
• Cite our paper if you find our work is helpful~

@article{ding2021cogview,
title={CogView: Mastering Text-to-Image Generation via Transformers},
author={Ding, Ming and Yang, Zhuoyi and Hong, Wenyi and Zheng, Wendi and Zhou, Chang and Yin, Da and Lin, Junyang and Zou, Xu and Shao, Zhou and Yang, Hongxia and Tang, Jie},
journal={arXiv preprint arXiv:2105.13290},
year={2021}

• Google Colab Two contributors successfully setup up CogView on Colab !

Getting Started

Setup

• Hardware: Linux servers with Nvidia V100s or A100s are recommended, but it is also okay to run the pretrained models with smaller --max-inference-batch-size or training smaller models on less powerful GPUs.
• Environment (Option 1): Please first install PyTorch (>=1.7.0) and apex, and then install other dependencies via pip install -r requirements.txt.

• Environment (Option 2): We prepare a docker image in case that you fail to handle the environments. Pull the image, create a (background) container and get into it via:

docker pull cogview/cuda111_torch181_deepspeed040
./env/start_docker.sh && docker exec -it bg-cogview bash

cd /root/cogview # in the container

Download

0. Download the image tokenizer vqvae_hard_biggerset_011.pt from BAAI website or Tsinghua Cloud. Place the file under pretrained/vqvae.

wget 'https://cloud.tsinghua.edu.cn/f/71607a5dca69417baa8c/?dl=1' -O pretrained/vqvae/vqvae_hard_biggerset_011.pt

1. Download models from Project Wudao-Wenhui. | FileName | Discription | | ---- | ---- | | cogview-base.tar | The pretrained text-to-image model. | | cogview-caption.tar | Finetuned image-to-text model, also used for reranking. | | cogview-sr.tar | Finetuned super-resolution model. (warning: it runs slow.) |

Uncompress them into pretrained/cogview/. The following command should be modified based on the model name.

tar -xvf cogview-{base, sr, caption}.tar -C pretrained/cogview/

2. (Only for training tutorial, skip it for inference.) Download a small "bird-and-animal" example dataset from our link at Tsinghua Cloud.

wget https://cloud.tsinghua.edu.cn/f/1e4963ec8ac84941ba68/?dl=1 -O data/bird_animal.bin

Run CogView! (Model Inference)

We encapsulate the generation functions into scripts. See generate_samples.py and arguments.py for details.

Text-to-Image Generation

Write text queries (one per line) into input.txt and run:

./scripts/text2image.sh --debug

The results will in a new folder samples_text2image/.

Arguments useful in inference are mainly:

• --input-source [path or "interactive"]. The path of the input file, can also be "interactive", which will launch a CLI.
• --output-path [path]. The folder containing the results.
• --batch-size [int]. The number of samples will be generated per query.
• --max-inference-batch-size [int]. Maximum batch size per forward. Reduce it if OOM.
• --debug. Only save concatenated images for all generated samples, and name them by input text and date.
• --with-id. When it toggled, you must specify an "id" before each input, e.g. 001\t一个漂亮的女孩, \t denoting TAB (NOT space). It will generate batch-size split images in a folder named "id" for each input. Confict with --debug.
• --device [int]. Running on which GPU.

Super-resolution

Run the following script and input text\t{image_path}, where {image_path} means the path of a previously generated image.

./scripts/super_resolution.sh

Note: *It is only effective for generated images from our Image Tokenizer (due to the token distribution).*

Image-to-Text

The input is "one image path per line", and will print the results to stdout.

./scripts/image2text.sh

Note: *Not optimized for this task, so it might not very competitive (but okay). We will consider to release a version funetuning for a longer period on this task in the future.* (*TODO*)

Post-selection

This application only takes file inputs, where each line is…