Just playing with getting VQGAN+CLIP running locally, rather than having to use colab.

Last update: Jan 04, 2023

Related tags

Overview

VQGAN-CLIP Overview

A repo for running VQGAN+CLIP locally. This started out as a Katherine Crowson VQGAN+CLIP derived Google colab notebook.

Original notebook:

Some example images:

Environment:

Tested on Ubuntu 20.04
GPU: Nvidia RTX 3090
Typical VRAM requirements:
- 24 GB for a 900x900 image
- 10 GB for a 512x512 image
- 8 GB for a 380x380 image

Still a work in progress - I've not actually tested everything yet :)

Set up

Example set up using Anaconda to create a virtual Python environment with the prerequisites:

conda create --name vqgan python=3.9
conda activate vqgan

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install ftfy regex tqdm omegaconf pytorch-lightning IPython kornia imageio imageio-ffmpeg einops 

git clone https://github.com/openai/CLIP
git clone https://github.com/CompVis/taming-transformers.git

You will also need at least 1 VQGAN pretrained model. E.g.

mkdir checkpoints
curl -L -o checkpoints/vqgan_imagenet_f16_16384.yaml -C - 'http://mirror.io.community/blob/vqgan/vqgan_imagenet_f16_16384.yaml' #ImageNet 16384
curl -L -o checkpoints/vqgan_imagenet_f16_16384.ckpt -C - 'http://mirror.io.community/blob/vqgan/vqgan_imagenet_f16_16384.ckpt' #ImageNet 16384

By default, the model .yaml and .ckpt files are expected in the checkpoints directory. See https://github.com/CompVis/taming-transformers for more information on datasets and models.

Run

To generate images from text, specify your text prompt as shown in the example below:

python generate.py -p "A painting of an apple in a fruit bowl"

Multiple prompts

Text and image prompts can be split using the pipe symbol in order to allow multiple prompts. For example:

python generate.py -p "A painting of an apple in a fruit bowl | psychedelic | surreal | weird"

Image prompts can be split in the same way. For example:

python generate.py -p "A picture of a bedroom with a portrait of Van Gogh" -ip "samples/VanGogh.jpg | samples/Bedroom.png"

"Style Transfer"

An input image with style text and a low number of iterations can be used create a sort of "style transfer" effect. For example:

python generate.py -p "A painting in the style of Picasso" -ii samples/VanGogh.jpg -i 80 -se 10 -opt AdamW -lr 0.25

Output	Style
	Picasso
	Sketch
	Psychedelic

Feedback example

By feeding back the generated images and making slight changes, some interesting effects can be created.

The example zoom.sh shows this by applying a zoom and rotate to generated images, before feeding them back in again. To use zoom.sh, specifying a text prompt, output filename and number of frames. E.g.

./zoom.sh "A painting of a red telephone box spinning through a time vortex" Telephone.png 150

Random text example

Use random.sh to make a batch of images from random text. Edit the text and number of generated images to your taste!

./random.sh

Advanced options

To view the available options, use "-h".

python generate.py -h

usage: generate.py [-h] [-p PROMPTS] [-o OUTPUT] [-i MAX_ITERATIONS] [-ip IMAGE_PROMPTS]
[-nps [NOISE_PROMPT_SEEDS ...]] [-npw [NOISE_PROMPT_WEIGHTS ...]] [-s SIZE SIZE]
[-ii INIT_IMAGE] [-iw INIT_WEIGHT] [-m CLIP_MODEL] [-conf VQGAN_CONFIG]
[-ckpt VQGAN_CHECKPOINT] [-lr STEP_SIZE] [-cuts CUTN] [-cutp CUT_POW] [-se DISPLAY_FREQ]
[-sd SEED] [-opt OPTIMISER]

optional arguments:
  -h, --help            show this help message and exit
  -p PROMPTS, --prompts PROMPTS
                        Text prompts
  -o OUTPUT, --output OUTPUT
                        Number of iterations
  -i MAX_ITERATIONS, --iterations MAX_ITERATIONS
                        Number of iterations
  -ip IMAGE_PROMPTS, --image_prompts IMAGE_PROMPTS
                        Image prompts / target image
  -nps [NOISE_PROMPT_SEEDS ...], --noise_prompt_seeds [NOISE_PROMPT_SEEDS ...]
                        Noise prompt seeds
  -npw [NOISE_PROMPT_WEIGHTS ...], --noise_prompt_weights [NOISE_PROMPT_WEIGHTS ...]
                        Noise prompt weights
  -s SIZE SIZE, --size SIZE SIZE
                        Image size (width height)
  -ii INIT_IMAGE, --init_image INIT_IMAGE
                        Initial image
  -iw INIT_WEIGHT, --init_weight INIT_WEIGHT
                        Initial image weight
  -m CLIP_MODEL, --clip_model CLIP_MODEL
                        CLIP model
  -conf VQGAN_CONFIG, --vqgan_config VQGAN_CONFIG
                        VQGAN config
  -ckpt VQGAN_CHECKPOINT, --vqgan_checkpoint VQGAN_CHECKPOINT
                        VQGAN checkpoint
  -lr STEP_SIZE, --learning_rate STEP_SIZE
                        Learning rate
  -cuts CUTN, --num_cuts CUTN
                        Number of cuts
  -cutp CUT_POW, --cut_power CUT_POW
                        Cut power
  -se DISPLAY_FREQ, --save_every DISPLAY_FREQ
                        Save image iterations
  -sd SEED, --seed SEED
                        Seed
  -opt OPTIMISER, --optimiser OPTIMISER
                        Optimiser (Adam, AdamW, Adagrad, Adamax)

Citations

@misc{unpublished2021clip,
    title  = {CLIP: Connecting Text and Images},
    author = {Alec Radford, Ilya Sutskever, Jong Wook Kim, Gretchen Krueger, Sandhini Agarwal},
    year   = {2021}
}

@misc{esser2020taming,
      title={Taming Transformers for High-Resolution Image Synthesis}, 
      author={Patrick Esser and Robin Rombach and Björn Ommer},
      year={2020},
      eprint={2012.09841},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Katherine Crowson - https://github.com/crowsonkb

Public Domain images from Open Access Images at the Art Institute of Chicago - https://www.artic.edu/open-access/open-access-images

Just playing with getting VQGAN+CLIP running locally, rather than having to use colab.

Related tags

Overview

VQGAN-CLIP Overview

Set up

Run

Multiple prompts

"Style Transfer"

Feedback example

Random text example

Advanced options

Citations

Owner

Nerdy Rodent

Simple renderer for use with MuJoCo (>=2.1.2) Python Bindings.

Near-Optimal Sparse Allreduce for Distributed Deep Learning (published in PPoPP'22)

Pytorch implementation for "Implicit Semantic Response Alignment for Partial Domain Adaptation"

[CVPR'2020] DeepDeform: Learning Non-rigid RGB-D Reconstruction with Semi-supervised Data

Jittor 64*64 implementation of StyleGAN

1st Solution For ICDAR 2021 Competition on Mathematical Formula Detection

A Comprehensive Empirical Study of Vision-Language Pre-trained Model for Supervised Cross-Modal Retrieval

Supplementary code for the paper "Meta-Solver for Neural Ordinary Differential Equations" https://arxiv.org/abs/2103.08561

Guiding evolutionary strategies by (inaccurate) differentiable robot simulators @ NeurIPS, 4th Robot Learning Workshop

StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks

Deep Learning GPU Training System

An implementation of DeepMind's Relational Recurrent Neural Networks in PyTorch.

Optimizing synthesizer parameters using gradient approximation

Implementation of "DeepOrder: Deep Learning for Test Case Prioritization in Continuous Integration Testing".

A resource for learning about deep learning techniques from regression to LSTM and Reinforcement Learning using financial data and the fitness functions of algorithmic trading

Bayesian inference for Permuton-induced Chinese Restaurant Process (NeurIPS2021).

Spontaneous Facial Micro Expression Recognition using 3D Spatio-Temporal Convolutional Neural Networks

An executor that performs image segmentation on fashion items

IJON is an annotation mechanism that analysts can use to guide fuzzers such as AFL.

Detectorch - detectron for PyTorch