EdiBERT is a generative model based on a bi-directional transformer, suited for image manipulation

Last update: Dec 07, 2022

Related tags

Overview

EdiBERT, a generative model for image editing

EdiBERT is a generative model based on a bi-directional transformer, suited for image manipulation. The same EdiBERT model, derived from a single training, can be used on a wide variety of tasks.

We follow the implementation of Taming-Transformers (https://github.com/CompVis/taming-transformers). Main modifications can be found in: taming/models/bert_transformer.py ; scripts/sample_mask_likelihood_maximization.py.

Requirements

A suitable conda environment named edibert can be created and activated with:

conda env create -f environment.yaml
conda activate edibert

FFHQ

Download FFHQ dataset (https://github.com/NVlabs/ffhq-dataset) and put it into data/ffhq/.

Training BERT

In the logs/ folder, download and extract the FFHQ VQGAN:

gdown --id '1P_wHLRfdzf1DjsAH_tG10GXk9NKEZqTg'
tar -xvzf 2021-04-23T18-19-01_ffhq_vqgan.tar.gz

Training on 1 GPUs:

python main.py --base configs/ffhq_transformer_bert_2D.yaml -t True --gpus 0,

Training on 2 GPUs:

python main.py --base configs/ffhq_transformer_bert_2D.yaml -t True --gpus 0,1

Running pre-trained BERT on composite/scribble-edited images

In the logs/ folder, download and extract the FFHQ VQGAN:

gdown --id '1P_wHLRfdzf1DjsAH_tG10GXk9NKEZqTg'
tar -xvzf 2021-04-23T18-19-01_ffhq_vqgan.tar.gz

In the logs/ folder, download and extract the FFHQ BERT:

gdown --id '1YGDd8XyycKgBp_whs9v1rkYdYe4Oxfb3'
tar -xvzf 2021-10-14T16-32-28_ffhq_transformer_bert_2D.tar.gz

folders and place them into logs.

Then, launch the following script for composite images:

python scripts/sample_mask_likelihood_maximization.py -r logs/2021-10-14T16-32-28_ffhq_transformer_bert_2D/checkpoints/epoch=000019.ckpt \
--image_folder data/ffhq_collages/ --mask_folder data/ffhq_collages_masks/ --image_list data/ffhq_collages.txt --keep_img \
--dilation_sampling 1 -k 100 -t 1.0 --batch_size 5 --bert --epochs 2  \
--device 0 --random_order \
--mask_collage --collage_frequency 3 --gaussian_smoothing_collage

Then, launch the following script for edits images:

python scripts/sample_mask_likelihood_maximization.py -r logs/2021-10-14T16-32-28_ffhq_transformer_bert_2D/checkpoints/epoch=000019.ckpt \
--image_folder data/ffhq_edits/ --mask_folder data/ffhq_edits_masks/ --image_list data/ffhq_edits.txt --keep_img \
--dilation_sampling 1 -k 100 -t 1.0 --batch_size 5 --bert --epochs 2  \
--device 0 --random_order \
--mask_collage --collage_frequency 3 --gaussian_smoothing_collage

The samples can then be found in logs/my_model/samples/. Here, the --batch_size argument corresponds to the number of EdiBERT generations per image.

Notebooks for playing with completion/denoising with BERT

Notebooks for image denoising and image inpainting can also be found in the main folder.

EdiBERT is a generative model based on a bi-directional transformer, suited for image manipulation

Related tags

Overview

EdiBERT, a generative model for image editing

Requirements

FFHQ

Training BERT

Running pre-trained BERT on composite/scribble-edited images

Notebooks for playing with completion/denoising with BERT

Owner

Code release for our paper, "SimNet: Enabling Robust Unknown Object Manipulation from Pure Synthetic Data via Stereo"

PConv-Keras - Unofficial implementation of "Image Inpainting for Irregular Holes Using Partial Convolutions". Try at: www.fixmyphoto.ai

An implementation of the BADGE batch active learning algorithm.

PyTorch implementation of Pay Attention to MLPs

🍅🍅🍅YOLOv5-Lite: lighter, faster and easier to deploy. Evolved from yolov5 and the size of model is only 1.7M (int8) and 3.3M (fp16). It can reach 10+ FPS on the Raspberry Pi 4B when the input size is 320×320~

aka "Bayesian Methods for Hackers": An introduction to Bayesian methods + probabilistic programming with a computation/understanding-first, mathematics-second point of view. All in pure Python ;)

HeartRate detector with ArduinoandPython - Use Arduino and Python create a heartrate detector.

Anderson Acceleration for Deep Learning

Official code of our work, Unified Pre-training for Program Understanding and Generation [NAACL 2021].

[SIGGRAPH Asia 2021] Pose with Style: Detail-Preserving Pose-Guided Image Synthesis with Conditional StyleGAN

Sample code from the Neural Networks from Scratch book.

The Power of Scale for Parameter-Efficient Prompt Tuning

Fast image augmentation library and easy to use wrapper around other libraries. Documentation: https://albumentations.ai/docs/ Paper about library: https://www.mdpi.com/2078-2489/11/2/125

This library provides an abstraction to perform Model Versioning using Weight & Biases.

This repo provides function call to track multi-objects in videos

A set of examples around hub for creating and processing datasets

The official implementation of You Only Compress Once: Towards Effective and Elastic BERT Compression via Exploit-Explore Stochastic Nature Gradient.

High level network definitions with pre-trained weights in TensorFlow

The Face Mask recognition system uses AI technology to detect the person with or without a mask.

PyTorch inference for "Progressive Growing of GANs" with CelebA snapshot