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

Set up

This example uses Anaconda to manage virtual Python environments.

Create a new virtual Python environment for VQGAN-CLIP:

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

Install Pytorch in the new enviroment:

Note: This installs the CUDA version of Pytorch, if you want to use an AMD graphics card, read the AMD section below.

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

Install other required Python packages:

pip install ftfy regex tqdm omegaconf pytorch-lightning IPython kornia imageio imageio-ffmpeg einops torch_optimizer

Clone required repositories:

git clone 'https://github.com/nerdyrodent/VQGAN-CLIP'
cd VQGAN-CLIP
git clone 'https://github.com/openai/CLIP'
git clone 'https://github.com/CompVis/taming-transformers'

Note: In my development environment both CLIP and taming-transformers are present in the local directory, and so aren't present in the requirements.txt or vqgan.yml files.

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

mkdir checkpoints

curl -L -o checkpoints/vqgan_imagenet_f16_16384.yaml -C - 'https://heibox.uni-heidelberg.de/d/a7530b09fed84f80a887/files/?p=%2Fconfigs%2Fmodel.yaml&dl=1' #ImageNet 16384
curl -L -o checkpoints/vqgan_imagenet_f16_16384.ckpt -C - 'https://heibox.uni-heidelberg.de/d/a7530b09fed84f80a887/files/?p=%2Fckpts%2Flast.ckpt&dl=1' #ImageNet 16384

The download_models.sh script is an optional way to download a number of models. By default, it will download just 1 model.

See https://github.com/CompVis/taming-transformers#overview-of-pretrained-models for more information about VQGAN pre-trained models, including download links.

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.

Video guides are also available:

• Linux - https://www.youtube.com/watch?v=1Esb-ZjO7tw
• Windows - https://www.youtube.com/watch?v=XH7ZP0__FXs

Using an AMD graphics card

Note: This hasn't been tested yet.

ROCm can be used for AMD graphics cards instead of CUDA. You can check if your card is supported here: https://github.com/RadeonOpenCompute/ROCm#supported-gpus

Install ROCm accordng to the instructions and don't forget to add the user to the video group: https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html

The usage and set up instructions above are the same, except for the line where you install Pytorch. Instead of pip install torch==1.9.0+cu111 ..., use the one or two lines which are displayed here (select Pip -> Python-> ROCm): https://pytorch.org/get-started/locally/

Using the CPU

If no graphics card can be found, the CPU is automatically used and a warning displayed.

Regardless of an available graphics card, the CPU can also be used by adding this command line argument: -cd cpu

This works with the CUDA version of Pytorch, even without CUDA drivers installed, but doesn't seem to work with ROCm as of now.

Uninstalling

Remove the Python enviroment:

conda remove --name vqgan --all

and delete the VQGAN-CLIP directory.

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. You can also use a colon followed by a number to set a weight for that prompt. For example:

python generate.py -p "A painting of an apple in a fruit bowl | psychedelic | surreal:0.5 | weird:0.25"

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"

Story mode

Sets of text prompts can be created using the caret symbol, in order to generate a sort of story mode. For example:

python generate.py -p "A painting of a sunflower|photo:-1 ^ a painting of a rose ^ a painting of a tulip ^ a painting of a daisy flower ^ a photograph of daffodil" -cpe 1500 -zvid -i 6000 -zse 10 -vl 20 -zsc 1.005 -opt Adagrad -lr 0.15 -se 6000

"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

A video style transfer effect can be achived by specifying a directory of video frames in video_style_dir. Output will be saved in the steps directory, using the original video frame filenames. You can also use this as a sort of "batch mode" if you have a directory of images you want to apply a style to. This can also be combined with Story Mode if you don't wish to apply the same style to every images, but instead roll through a list of styles.

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

There is also a simple zoom video creation option available. For example:

python generate.py -p "The inside of a sphere" -zvid -i 4500 -zse 20 -vl 10 -zsc 0.97 -opt Adagrad -lr 0.15 -se 4500

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] [-ip IMAGE_PROMPTS] [-i MAX_ITERATIONS] [-se DISPLAY_FREQ]
[-s SIZE SIZE] [-ii INIT_IMAGE] [-in INIT_NOISE] [-iw INIT_WEIGHT] [-m CLIP_MODEL]
[-conf VQGAN_CONFIG] [-ckpt VQGAN_CHECKPOINT] [-nps [NOISE_PROMPT_SEEDS ...]]
[-npw [NOISE_PROMPT_WEIGHTS ...]] [-lr STEP_SIZE] [-cuts CUTN] [-cutp CUT_POW] [-sd SEED]
[-opt {Adam,AdamW,Adagrad,Adamax,DiffGrad,AdamP,RAdam,RMSprop}] [-o OUTPUT] [-vid] [-zvid]
[-zs ZOOM_START] [-zse ZOOM_FREQUENCY] [-zsc ZOOM_SCALE] [-cpe PROMPT_FREQUENCY]
[-vl VIDEO_LENGTH] [-ofps OUTPUT_VIDEO_FPS] [-ifps INPUT_VIDEO_FPS] [-d]
[-aug {Ji,Sh,Gn,Pe,Ro,Af,Et,Ts,Cr,Er,Re} [{Ji,Sh,Gn,Pe,Ro,Af,Et,Ts,Cr,Er,Re} ...]]
[-cd CUDA_DEVICE]

optional arguments:
  -h, --help            show this help message and exit
  -p PROMPTS, --prompts PROMPTS
                        Text prompts
  -ip IMAGE_PROMPTS, --image_prompts IMAGE_PROMPTS
                        Image prompts / target image
  -i MAX_ITERATIONS, --iterations MAX_ITERATIONS
                        Number of iterations
  -se DISPLAY_FREQ, --save_every DISPLAY_FREQ
                        Save image iterations
  -s SIZE SIZE, --size SIZE SIZE
                        Image size (width height) (default: [512, 512])
  -ii INIT_IMAGE, --init_image INIT_IMAGE
                        Initial image
  -in INIT_NOISE, --init_noise INIT_NOISE
                        Initial noise image (pixels or gradient)
  -iw INIT_WEIGHT, --init_weight INIT_WEIGHT
                        Initial weight
  -m CLIP_MODEL, --clip_model CLIP_MODEL
                        CLIP model (e.g. ViT-B/32, ViT-B/16)
  -conf VQGAN_CONFIG, --vqgan_config VQGAN_CONFIG
                        VQGAN config
  -ckpt VQGAN_CHECKPOINT, --vqgan_checkpoint VQGAN_CHECKPOINT
                        VQGAN checkpoint
  -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
  -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
  -sd SEED, --seed SEED
                        Seed
  -opt, --optimiser {Adam,AdamW,Adagrad,Adamax,DiffGrad,AdamP,RAdam,RMSprop}
                        Optimiser
  -o OUTPUT, --output OUTPUT
                        Output file
  -vid, --video         Create video frames?
  -zvid, --zoom_video   Create zoom video?
  -zs ZOOM_START, --zoom_start ZOOM_START
                        Zoom start iteration
  -zse ZOOM_FREQUENCY, --zoom_save_every ZOOM_FREQUENCY
                        Save zoom image iterations
  -zsc ZOOM_SCALE, --zoom_scale ZOOM_SCALE
                        Zoom scale
  -cpe PROMPT_FREQUENCY, --change_prompt_every PROMPT_FREQUENCY
                        Prompt change frequency
  -vl VIDEO_LENGTH, --video_length VIDEO_LENGTH
                        Video length in seconds
  -ofps OUTPUT_VIDEO_FPS, --output_video_fps OUTPUT_VIDEO_FPS
                        Create an interpolated video (Nvidia GPU only) with this fps (min 10. best set to 30 or 60)
  -ifps INPUT_VIDEO_FPS, --input_video_fps INPUT_VIDEO_FPS
                        When creating an interpolated video, use this as the input fps to interpolate from (>0 & <ofps)
  -d, --deterministic   Enable cudnn.deterministic?
  -aug, --augments {Ji,Sh,Gn,Pe,Ro,Af,Et,Ts,Cr,Er,Re} [{Ji,Sh,Gn,Pe,Ro,Af,Et,Ts,Cr,Er,Re} ...]
                        Enabled augments
  -cd CUDA_DEVICE, --cuda_device CUDA_DEVICE
                        Cuda device to use

Troubleshooting

CUSOLVER_STATUS_INTERNAL_ERROR

For example:

RuntimeError: cusolver error: CUSOLVER_STATUS_INTERNAL_ERROR, when calling cusolverDnCreate(handle)

Make sure you have specified the correct size for the image. For more information please refer to #6

RuntimeError: CUDA out of memory

For example:

RuntimeError: CUDA out of memory. Tried to allocate 150.00 MiB (GPU 0; 23.70 GiB total capacity; 21.31 GiB already allocated; 78.56 MiB free; 21.70 GiB reserved in total by PyTorch)

Your request doesn't fit into your GPU's VRAM. Reduce the image size and/or number of cuts.

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