Text-to-Image Summary – Part 1

What Are Text-to-Image Systems

Text-to-Image systems/models/scripts/networks (what is the official correct term for these?) are machine learning based models that take a descriptive phrase as input and attempt to generate images that match the input phrase.

Requirements

You do need a decent NVIDIA GPU. 3090 recommended for 768×768 resolution, 2080 for smaller 256×256 images, 10xx possibly for tiny images or if you want to try reduced settings and wait ages for results. If you have a commercial grade GPU with more memory you will be able to push these resolutions higher. VRAM matters more than GPU model, ie you can get 3090s with only 16GB of VRAM and others with 24GB. You may see a laptop with an advertised 3080 GPU, but the total VRAM will likely be much smaller than a desktop 3080.

To run these scripts from Visions of Chaos you need to have installed these prerequisites. Once you get all the prerequisites setup it really is as simple as typing your prompt text and clicking a button. I do include a lot of other settings so you can tweak the script parameters as you do more experimentation.

Text-to-Image GUI

Visions of Chaos Text-to-Image Tutorial

You can watch the following tutorial video to get an idea of how the Text-to-Image mode works in Visions of Chaos.

Text-to-Image Scripts Included With Visions of Chaos

The rest of this blog post (and other parts) lists the 49 (so far) Text-to-Image scripts that I have been able to get working with Visions of Chaos.

If you are the author of one of these scripts then many thanks to you for sharing the code publicly. If you are a creator of a script I do not include here, please leave a comment with a link or send me an email so I can try it out. If you are a better coder than I am and improve any of these also let me know and I will share your fixes with the world.

I have included sample image outputs from each script. Most of the text prompts for these samples come from a prompt builder I include with Visions of Chaos that randomly combines subjects, adjectives, styles and artists.

Note also that these samples all use the default settings for GAN and CLIP models. Most of the included scripts allow tweaking of settings and different models to alter the outputs. There is a much wider range of output images possible. Download Visions of Chaos to experiment with all the combinations of scripts, models, prompts and settings.


Name: Deep Daze
Author: Phil Wang
Original script: https://github.com/lucidrains/deep-daze
Time for 512×512 on a 3090: 1 minutes 53 seconds.
Maximum resolution on a 24 GB 3090: 1024×1024
Description: This was the first Text-to-Image script I ever found and tested. The output images from the original script are very washed out and pastel shaded, but after adding some torchvision transforms for brightness, contrast and sharpness tweaks they are a little better. Very abstract output compared to the other scripts.

'a bronze sculpture of a colorful parrot in the style of Kandinsky' Deep Daze Text-to-Image
a bronze sculpture of a colorful parrot in the style of Kandinsky

'a crying person' Deep Daze Text-to-Image
a crying person

'a desert oasis' Deep Daze Text-to-Image
a desert oasis

'a surrealist painting of the Terminator made of silver' Deep Daze Text-to-Image
a surrealist painting of the Terminator made of silver

'a zombie in the style of Turner' Deep Daze Text-to-Image
a zombie in the style of Turner


Name: Big Sleep
Author: Phil Wang
Original script: https://github.com/lucidrains/big-sleep
Time for 512×512 on a 3090: 4 minutes 0 seconds
Maximum resolution on a 24 GB 3090: 512×512
Description: Can give a good variety of images for any prompt text and does not suffer from the coloring or tiled image issues some of the other methods do. See here for my older post with a lot of Big Sleep examples. If you give it a chance and run repeated batches of the same prompt you can get some very nice results.

'H R Giger' Big Sleep Text-to-Image
H R Giger

'surrealism' Big Sleep Text-to-Image
surrealism

'colorful surrealism' Big Sleep Text-to-Image
colorful surrealism

'a charcoal drawing of a landscape' Big Sleep Text-to-Image
a charcoal drawing of a landscape


Name: VQGAN+CLIP z-quantize
Author: Katherine Crowson
Original script: https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN
Time for 512×512 on a 3090: 2 minutes 28 seconds
Maximum resolution on a 24 GB 3090: 768×768 or 1120×480
Description: The outputs tend to be divided up into rectangular regions, but the resulting imagery can be interesting.

'a drawing of a bouquet of flowers made of cardboard' VQGAN+CLIP z-quantize Text-to-Image
a drawing of a bouquet of flowers made of cardboard

'a rose made of silver' VQGAN+CLIP z-quantize Text-to-Image
a rose made of silver

'a tilt shift photo of traffic' VQGAN+CLIP z-quantize Text-to-Image
a tilt shift photo of traffic

'an abstract painting of a house made of crystals' VQGAN+CLIP z-quantize Text-to-Image
an abstract painting of a house made of crystals

'an abstract painting of a skull' VQGAN+CLIP z-quantize Text-to-Image
an abstract painting of a skull

VQGAN+CLIP z-quantize allows specifying an image as the input starting point. If you take the output, stretch it very slightly, and then feed it back into the system each frame you get a movie zooming in. For this movie I used SRCNN Super Resolution to double the resolution of the frames and then Super Slo-Mo for optical flow frame interpolation (both SRCNN and Super Slo-Mo are included with Visions of Chaos). The VQGAN model was “vqgan_imagenet_f16_16384” and the CLIP model was “ViT-B/32”. The prompts were the seven deadly sins, ie “a watercolor painting depicting pride”, “a watercolor painting depicting greed” etc.

The more astute viewers among you will notice there are only 6 of the sins in the previous video. What happened to “lust”? A while back one of my uploads was flagged as porn by the YouTube robots. Their (what I assume is) machine learning based system detected my upload as porn when there was no porn in it. An appeal was met with instant denial and so I now have a permanent “warning” on my channel with no way to talk to a person who could spend 1 minute looking at the video to tell it isn’t porn. Another warning would lead to a strike, so I am being overly cautious and omitting the lust part from the YouTube video. Those who want to see the full 7 part movie can click the following link to watch it on my LBRY channel.

https://open.lbry.com/@Softology:5/Seven-Deadly-Sins:6

Thanks LBRY!


Name: VQGAN+CLIP codebook
Author: Katherine Crowson
Original script: https://colab.research.google.com/drive/15UwYDsnNeldJFHJ9NdgYBYeo6xPmSelP
Time for 512×512 on a 3090: 3 minutes 19 seconds
Maximum resolution on a 24 GB 3090: 768×768 or 1120×480
Description: VQGAN-CLIP codebook seem to give very similar images for the same prompt phrase, so repeatedly running the script (with different seed values) does not give a wide variety of resulting images. Still gives interesting results.

'a happy alien' VQGAN+CLIP codebook Text-to-Image
a happy alien

'a library' VQGAN+CLIP codebook Text-to-Image
a library

'a teddy bear' VQGAN+CLIP codebook Text-to-Image
a teddy bear

'digital art of a colorful parrot' VQGAN+CLIP codebook Text-to-Image
digital art of a colorful parrot

'digital art of an amusement park' VQGAN+CLIP codebook Text-to-Image
digital art of an amusement park


Name: Aleph2Image Gamma
Author: Ryan Murdock
Original script: https://colab.research.google.com/drive/1VAO22MNQekkrVq8ey2pCRznz4A0_jY29
Time for 512×512 on a 3090: 2 minutes 1 second
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: This one seems to evolve white blotches that grow and take over the entire image. Before the white out stage the images tend to have too much contrast.

'H R Giger' Aleph2Image Gamma Text-to-Image
H R Giger

'surrealism' Aleph2Image Gamma Text-to-Image
surrealism

'seascape painting' Aleph2Image Gamma Text-to-Image
seascape painting


Name: Aleph2Image Delta
Author: Ryan Murdock
Original script: https://colab.research.google.com/drive/1oA1fZP7N1uPBxwbGIvOEXbTsq2ORa9vb
Time for 512×512 on a 3090: 2 minutes 1 second
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: A newer revision of Aleph2Image that doesn’t have the white out issues. The resulting images have much more vibrant colors and that may be a good or bad point depending on your preferences.

'a sketch of an angry person' Aleph2Image Delta Text-to-Image
a sketch of an angry person

'a spooky forest' Aleph2Image Delta Text-to-Image
a spooky forest

'a sunset in the style of Rembrandt' Aleph2Image Delta Text-to-Image
a sunset in the style of Rembrandt

'a surrealist painting of a forest path' Aleph2Image Delta Text-to-Image
a surrealist painting of a forest path

'a tropical beach' Aleph2Image Delta Text-to-Image
a tropical beach


Name: Aleph2Image Delta v2
Author: Ryan Murdock
Original script: https://colab.research.google.com/drive/1NGM9L8qP0gwl5z5GAuB_bd0wTNsxqclG
Time for 512×512 on a 3090: 3 minutes 42 seconds
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: A newer revision of Aleph2Image Delta that gives much sharper results. The resulting images tend to be similar to each other for each prompt text so not a lot of variety.

'a cartoon of love in the style of Claude Monet' Aleph2Image Delta v2 Text-to-Image
a cartoon of love in the style of Claude Monet

'a detailed painting of a rose' Aleph2Image Delta v2 Text-to-Image
a detailed painting of a rose

'a drawing of a volcano' Aleph2Image v2 Delta Text-to-Image
a drawing of a volcano

'a house' Aleph2Image v2 Delta Text-to-Image
a house

'a submarine' Aleph2Image v2 Delta Text-to-Image
a submarine


Name: Deep Daze Fourier
Author: Vadim Epstein
Original script: https://colab.research.google.com/gist/afiaka87/e018dfa86d8a716662d30c543ce1b78e/text2image-siren.ipynb
Time for 512×512 on a 3090: 4 minutes 54 seconds
Maximum resolution on a 24 GB 3090: 512×512 or 640×360
Description: Creates more collaged images with sharp, crisp bright colors.

'a pencil sketch of a vampire made of bones' Deep Daze Fourier Text-to-Image
a pencil sketch of a vampire made of bones

'H R Giger' Deep Daze Fourier Text-to-Image
H R Giger

'medusa made of wood' Deep Daze Fourier Text-to-Image
medusa made of wood

'Shrek eating pizza' Deep Daze Fourier Text-to-Image
Shrek eating pizza

'surrealist Homer Simpson' Deep Daze Fourier Text-to-Image
surrealist Homer Simpson


Name: Text2Image v2
Author: Denis Malimonov
Original script: https://colab.research.google.com/github/tg-bomze/collection-of-notebooks/blob/master/Text2Image_v2.ipynb
Time for 512×512 on a 3090: 1 minute 48 seconds
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: Can give more abstract results of the input phrase. Colors and details can be sharp, but not always. Good variety of output for each input phrase. Definitely worth a try.

'a fireplace made of voxels' Text2Image v2 Text-to-Image
a fireplace made of voxels

'a green tree frog in the style of M C Escher' Text2Image v2 Text-to-Image
a green tree frog in the style of M C Escher

'a pencil sketch of an evil alien' Text2Image v2 Text-to-Image
a pencil sketch of an evil alien

'a sea monster' Text2Image v2 Text-to-Image
a sea monster

'The Incredible Hulk made of silver' Text2Image v2 Text-to-Image
The Incredible Hulk made of silver


Name: The Big Sleep Customized
Author: NMKD
Original script: https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO
Time for 512×512 on a 3090: 1 minute 45 seconds
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: Another good one. Worth exploring further.

'a forest path' The Big Sleep Customized Text-to-Image
a forest path

'a watercolor painting of a colorful parrot in the style of Kandinsky' The Big Sleep Customized Text-to-Image
a watercolor painting of a colorful parrot in the style of Kandinsky

'a western town' The Big Sleep Customized Text-to-Image
a western town

'Christmas' The Big Sleep Customized Text-to-Image
Christmas

'medusa made of vines' The Big Sleep Customized Text-to-Image
medusa made of vines


Name: Big Sleep Minmax
Author: @!goose
Original script: https://colab.research.google.com/drive/12CnlS6lRGtieWujXs3GQ_OlghmFyl8ch
Time for 512×512 on a 3090: 1 minute 45 seconds
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: Another interesting Big Sleep variation. Allows a second phrase to be specified that is minimized in the output. For example if your prompt for a landscape painting has too many clouds you could specify clouds as the minimize prompt so the system outputs less clouds in the resulting image.

'a charcoal drawing of an eyeball' Big Sleep Minmax Text-to-Image
a charcoal drawing of an eyeball

'an ultrafine detailed painting of a crying person made of voxels' Big Sleep Minmax Text-to-Image
an ultrafine detailed painting of a crying person made of voxels

'dense woodland' Big Sleep Minmax Text-to-Image
dense woodland

'King Kong made of wrought iron in the style of Frida Kahlo' Big Sleep Minmax Text-to-Image
King Kong made of wrought iron in the style of Frida Kahlo

'Michael Myers' Big Sleep Minmax Text-to-Image
Michael Myers


Name: CLIP Pseudo Slime Mold
Author: hotgrits
Original script: https://discord.com/channels/729741769192767510/730484623028519072/850857930881892372
Time for 512×512 on a 3090: 2 minutes 57 seconds
Maximum resolution on a 24 GB 3090: Locked to 512×512
Description: This one gives unique output compared to the others. Really nicely defined sharp details. The colors come from any color palette you select (currently all the 3,479 palettes within Visions of Chaos can be used) so you can “tint” the resulting images with color shades you prefer.

'H R Giger' CLIP Pseudo Slime Mold Text-to-Image
H R Giger

'H R Giger' CLIP Pseudo Slime Mold Text-to-Image
H R Giger with a different color palette

'Shrek eating pizza' CLIP Pseudo Slime Mold Text-to-Image
Shrek eating pizza

'seascape painting' CLIP Pseudo Slime Mold Text-to-Image
seascape painting


Name: Aleph2Image Dall-E Remake
Author: Daniel Russell
Original script: https://colab.research.google.com/drive/17ZSyxCyHUnwI1BgZG22-UFOtCWFvqQjy
Time for 512×512 on a 3090: 3 minutes 42 seconds
Maximum resolution on a 24 GB 3090: 768×768
Description: Another Aleph2Image variant.

'a color pencil sketch of Jason Vorhees made of plastic' Aleph2Image Dall-E Remake Text-to-Image
a color pencil sketch of Jason Vorhees made of plastic

'a cubist painting of a science laboratory' Aleph2Image Dall-E Remake Text-to-Image
a cubist painting of a science laboratory

'a green tree frog in the style of Kandinsky' Aleph2Image Dall-E Remake Text-to-Image
a green tree frog in the style of Kandinsky

'a watercolor painting of Godzilla' Aleph2Image Dall-E Remake Text-to-Image
a watercolor painting of Godzilla

'an octopus' Aleph2Image Dall-E Remake Text-to-Image
an octopus


Name: VQGAN+CLIP v3
Author: Eleiber
Original script: https://colab.research.google.com/drive/1go6YwMFe5MX6XM9tv-cnQiSTU50N9EeT
Time for 512×512 on a 3090: 2 minutes 52 seconds
Maximum resolution on a 24 GB 3090: 768×768 or 1120×480
Description: “v3” because it is the third VQGAN system I have tried and it didn’t have a unique specific name. Gives clear sharp images. Can give very painterly results with visible brush strokes if you use “a painting of” before the prompt subject.

'a pencil sketch of a campfire in the style of Da Vinci' VQGAN+CLIP v3 Text-to-Image
a pencil sketch of a campfire in the style of Da Vinci

'a pop art painting of a lush rainforest' VQGAN+CLIP v3 Text-to-Image
a pop art painting of a lush rainforest

'a storybook illustration of a cityscape' VQGAN+CLIP v3 Text-to-Image
a storybook illustration of a cityscape

'an airbrush painting of frogs' VQGAN+CLIP v3 Text-to-Image
an airbrush painting of frogs

'the Amazon Rainforest' VQGAN+CLIP v3 Text-to-Image
the Amazon Rainforest

VQGAN+CLIP v3 allows specifying an image as the input starting point. If you take the output and repeatedly use it as the input with some minor image stretching each frame you can get a movie zooming into the Text-to-Image output. For this movie I used SRCNN Super Resolution to double the resolution of the frames and then Super Slo-Mo for optical flow frame interpolation (both SRCNN and Super Slo-Mo are included with Visions of Chaos). The VQGAN model was “vqgan_imagenet_f16_16384” and the CLIP model was “ViT-B/32”.

This next example movie is showing a “Self-Driven” zoom movie. As in a regular zoom movie the output frames are slightly stretched and fed back into the system each frame. The self-driven difference with this movie is that the Text-to-Image prompt text is automatically changed every 2 seconds by CLIP detecting what it “sees” in the current frame. This way the movie subjects are automatically changed and steered in new directions in a totally automated way. There is no human control except me setting the initial “A landscape” prompt. After that it was fully automated.

By default the CLIP Image Captioning script is very good at detecting what is in an image. Using the default accuracy resulted in a zoom movie that got stuck with a single topic or subject. One got stuck on a slight variation of a prompt dealing with kites, so as the zoom movie went deeper it only showed kites. Luckily after tweaking and decreasing the accuracy of the CLIP captioning the predicitons allow the resulting subjects to drift to new topics during the movie.


Name: VQGAN+CLIP v4
Author: crimeacs
Original script: https://colab.research.google.com/drive/1ZAus_gn2RhTZWzOWUpPERNC0Q8OhZRTZ
Time for 512×512 on a 3090: 2 minutes 37 seconds
Maximum resolution on a 24 GB 3090: 768×768 or 1120×480
Description: Another improved VQGAN system utilizing pooling. “v4” because it is the forth VQGAN system I have tried and it didn’t have a unique specific name.

'a fine art painting of a cozy den' VQGAN+CLIP v4 Text-to-Image
a fine art painting of a cozy den

'a king in the style of Kandinsky' VQGAN+CLIP v4 Text-to-Image
a king in the style of Kandinsky

'a nurse in the style of Edward Hopper' VQGAN+CLIP v4 Text-to-Image
a nurse in the style of Edward Hopper

'a pastel of a demon' VQGAN+CLIP v4 Text-to-Image
a pastel of a demon

'a watercolor painting of a mountain path' VQGAN+CLIP v4 Text-to-Image
a watercolor painting of a mountain path

VQGAN+CLIP v4 allows specifying an image as the input starting point. If you take the output and repeatedly use it as the input with some minor image stretching each frame you can get a movie zooming into the Text-to-Image output. For this movie I used SRCNN Super Resolution to double the resolution of the frames and then Super Slo-Mo for optical flow frame interpolation (both SRCNN and Super Slo-Mo are included with Visions of Chaos). The VQGAN model was “vqgan_imagenet_f16_16384” and the CLIP model was “ViT-B/32”.

The text prompts for each part came from an idea in a YouTube comment to try more non-specific terms to see what happens, so here are the results of “an image of fear”, “an image of humanity”, “an image of knowledge”, “an image of love”, “an image of morality” and “an image of serenity”.

Here is another example. This time using the prompt of various directors, ie “Stanley Kubrick imagery”, “David Lynch imagery” etc. No super resolution this time. Super Slo-Mo was used for optical flow. I wasn’t sure if YouTube would accept the potentially unsettling horror visuals and I do not want to risk the hassle of a strike, so being on the safe side I am hosting this one on my LBRY channel only. Click the following image to open the movie in a new window. Note that LBRY can be a lot slower to buffer, so you may need to pause it for a while to let the movie load in.

Directors Text-to-Image

If you find that too slow to buffer/load I also have a copy on my BitChute channel here.


Continued in Part 2, Part 3 and Part 4.

Jason.

Deep Daze Fourier Text-to-Image

NOTE: Make sure you also see this post that has a summary of all the Text-to-Image scripts supported by Visions of Chaos with example images.

More Fascinating Text-to-Image

This time “Deep Daze Fourier” from Vadim Epstein. Code available in this notebook.

Compared to the last Deep Daze that generated washed out and pastel shaded results this Deep Daze creates images with sharp, crisp bright colors.

Sample results

“Shrek eating pizza”

Deep Daze Fourier - Shrek Eating Pizza

Deep Daze Fourier - Shrek Eating Pizza

Deep Daze Fourier - Shrek Eating Pizza

Deep Daze Fourier - Shrek Eating Pizza

“H R Giger”

Deep Daze Fourier - H R Giger

Deep Daze Fourier - H R Giger

Deep Daze Fourier - H R Giger

Deep Daze Fourier - H R Giger

“Freddy Krueger”

Deep Daze Fourier - Freddy Krueger

Deep Daze Fourier - Freddy Krueger

Deep Daze Fourier - Freddy Krueger

Deep Daze Fourier - Freddy Krueger

“Surrealist Homer Simpson”

Deep Daze Fourier - Surrealist Homer Simpson

Deep Daze Fourier - Surrealist Homer Simpson

Deep Daze Fourier - Surrealist Homer Simpson

Deep Daze Fourier - Surrealist Homer Simpson

“rose bush”

Deep Daze Fourier - Rose Bush

Deep Daze Fourier - Rose Bush

Deep Daze Fourier - Rose Bush

Deep Daze Fourier - Rose Bush

Availability

This and the previous Text-to-Image systems I have experimented with (here, here and here) are now supported by a GUI front end in Visions of Chaos. As long as you install these prerequisites and have a decent GPU you will be able to run these systems yourself.

Text-to-Image GUI

For those who love to tinker I have now added a bunch more of the script parameters so you no longer have to edit the Python source code outside Visions of Chaos.

Other Text-to-Image

If you know of any other Text-to-Image systems (with sharable open-source code) then please let me know. All of the Text-to-Image systems I have tested so far all have their own unique behaviors and outputs so I will always be on the lookout for more new variations.

Jason.

Aleph2Image Text-to-Image

NOTE: Make sure you also see this post that has a summary of all the Text-to-Image scripts supported by Visions of Chaos with example images.

Previously I experimented with Big Sleep and other Text-to-Image systems.

This post covers variations of Aleph2Image Text-to_Image. Originally coded by Ryan Murdock.


Aleph2Image “Gamma”

Code from this colab. This one seems to evolve white blotches that grow and take over the entire image. Before the white out stage the images tend to have too much contrast. Previous results from Deep Daze were too washed out, this one is too “contrasty”. If they could both be pushed towards that “sweet spot” they would both look much better.

“surrealism”

Aleph2Image Gamma - Surrealism

Aleph2Image Gamma - Surrealism

Aleph2Image Gamma - Surrealism

Aleph2Image Gamma - Surrealism

“H R Giger”

Aleph2Image Gamma - H R Giger

Aleph2Image Gamma - H R Giger

Aleph2Image Gamma - H R Giger

Aleph2Image Gamma - H R Giger

“seascape oil painting”

Aleph2Image Gamma - Seascape Oil Painting

Aleph2Image Gamma - Seascape Oil Painting

Aleph2Image Gamma - Seascape Oil Painting

Aleph2Image Gamma - Seascape Oil Painting

“frogs in the rain”

Aleph2Image Gamma - Frogs In The Rain

Aleph2Image Gamma - Frogs In The Rain

Aleph2Image Gamma - Frogs In The Rain

Aleph2Image Gamma - Frogs In The Rain


Aleph2Image “Delta”

Code from this colab. A newer revision of Aleph2Image that doesn’t have the white out issues. The resulting images have much more vibrant colors.

“surrealism”

Aleph2Image Delta - Surrealism

Aleph2Image Delta - Surrealism

Aleph2Image Delta - Surrealism

Aleph2Image Delta - Surrealism

“H R Giger”

Aleph2Image Delta - H R Giger

Aleph2Image Delta - H R Giger

Aleph2Image Delta - H R Giger

Aleph2Image Delta - H R Giger

“seascape oil painting”

Aleph2Image Delta - Seascape Oil Painting

Aleph2Image Delta - Seascape Oil Painting

Aleph2Image Delta - Seascape Oil Painting

Aleph2Image Delta - Seascape Oil Painting

“frogs in the rain”

Aleph2Image Delta - Frogs In The Rain

Aleph2Image Delta - Frogs In The Rain

Aleph2Image Delta - Frogs In The Rain

Aleph2Image Delta - Frogs In The Rain


Improved Aleph2Image “Delta” v2

Code from this colab. A newer revision of Aleph2Image Delta that gives much better results, although the results tend to be similar to each other for each prompt text. This and Big Sleep would be the best 2 Text-to-Image systems I have experimented with so far.

“surrealism”

Aleph2Image Delta v2 - Surrealism

Aleph2Image Delta v2 - Surrealism

Aleph2Image Delta v2 - Surrealism

Aleph2Image Delta v2 - Surrealism

“H R Giger”

Aleph2Image Delta v2 - H R Giger

Aleph2Image Delta v2 - H R Giger

Aleph2Image Delta v2 - H R Giger

Aleph2Image Delta v2 - H R Giger

“seascape oil painting”

Aleph2Image Delta v2 - Seascape Oil Painting

Aleph2Image Delta v2 - Seascape Oil Painting

Aleph2Image Delta v2 - Seascape Oil Painting

Aleph2Image Delta v2 - Seascape Oil Painting

“frogs in the rain”

Aleph2Image Delta v2 - Frogs In The Rain

Aleph2Image Delta v2 - Frogs In The Rain

Aleph2Image Delta v2 - Frogs In The Rain

Aleph2Image Delta v2 - Frogs In The Rain


Easy GUI Front End

I include a simple GUI dialog front end for these Text-to-Image systems in Visions of Chaos. As long as you have the prerequisites installed you will be able to convert text prompts into single or multiple images.

Text-to-Image GUI

You do need a GPU with lots of VRAM for these to work (especially the 512×512 image models).

Jason.

Further Explorations Into Text-to-Image Machine Learning

NOTE: Make sure you also see this post that has a summary of all the Text-to-Image scripts supported by Visions of Chaos with example images.

After my initial experiments with Big Sleep Text-to-Image generation I looked around for some more examples to play with. I was really impressed with Big Sleep and you can see some examples of Big Sleep output in that original post. I still think Big Sleep is the best Text-to-Image code I have used so far and better than what is in this post.


Deep Daze

Deep Daze is by Phil Wang and the source code is available here.

Deep Daze tends to generate collage-like images. As the first example image shows the resulting images have a washed out or faded look. I put the rest of the example Deep Daze images through a quick Auto White Balance pass in GIMP.

“H R Giger”

DeepDaze - H R Giger

DeepDaze - H R Giger

“Rainforest”

DeepDaze - Rainforest

“night club”

DeepDaze - Night Club

“seascape painting”

DeepDaze - Seascape Painting

“flowing water”

DeepDaze - Flowing Water


VQGAN-CLIP z+quantize

VQGAN-CLIP using a z+quantize method is from Katherine Crowson. Source code is available here.

This method also has the option to use an image to seed the initial model rather than just random noise, but the following examples were all seeded with noise. The resulting images tend to be divided up into rectangular regions, but the resulting imagery is interesting.

“H R Giger”

VQGAN-CLIP z+quantize - H R Giger

VQGAN-CLIP z+quantize - H R Giger

VQGAN-CLIP z+quantize - H R Giger

VQGAN-CLIP z+quantize - H R Giger

“rainforest”

VQGAN-CLIP z+quantize - Rainforest

VQGAN-CLIP z+quantize - Rainforest

VQGAN-CLIP z+quantize - Rainforest

VQGAN-CLIP z+quantize - Rainforest

“night club”

VQGAN-CLIP z+quantize - Night Club

VQGAN-CLIP z+quantize - Night Club

VQGAN-CLIP z+quantize - Night Club

VQGAN-CLIP z+quantize - Night Club

“seascape painting”

VQGAN-CLIP z+quantize - Seascape Painting

VQGAN-CLIP z+quantize - Seascape Painting

VQGAN-CLIP z+quantize - Seascape Painting

VQGAN-CLIP z+quantize - Seascape Painting

“flowing water”

VQGAN-CLIP z+quantize - Flowing Water

VQGAN-CLIP z+quantize - Flowing Water

VQGAN-CLIP z+quantize - Flowing Water

VQGAN-CLIP z+quantize - Flowing Water


VQGAN-CLIP codebook

VQGAN-CLIP using a codebook method is also from Katherine Crowson. Source code is available here.

VQGAN-CLIP codebook seem to give very similar images for different seeds, so I have only shown two examples for each phrase.

“H R Giger”

VQGAN-CLIP codebook - H R Giger

VQGAN-CLIP codebook - H R Giger

“rainforest”

VQGAN-CLIP codebook - Rainforest

VQGAN-CLIP codebook - Rainforest

“night club”

VQGAN-CLIP codebook - Night Club

VQGAN-CLIP codebook - Night Club

“seascape painting”

VQGAN-CLIP codebook - Seascape Painting

VQGAN-CLIP codebook - Seascape Painting

“flowing water”

VQGAN-CLIP codebook - Flowing Water

VQGAN-CLIP codebook - Flowing Water


Other Text-to-Image Models?

If you know of any other available Text-to-Image systems (that are freely available and shareable) let me know.


Availability

You can follow the above links and download the Python code yourself if you are so inclined.

I do include a basic GUI front-end for these Text-to-Image generators in Visions of Chaos. As long as you have the prerequisites installed (which you would need to install to run these outside Visions of Chaos) then you can experiment with these models yourself without needing to use the command line.

Text-to-Image GUI

Jason.

Super Resolution

The Dream

For years now you would have seen scenes in TV shows like CSI or movies like Blade Runner the “enhance” functionality of software that allows details to be enhanced in images that are only a blur or a few pixels in size. In Blade Runner, Deckard’s system even allowed him to look around corners.

The Reality

I have recently been testing machine learning neural network enhancers (aka super resolution) models. They resize an image while trying to maintain or enhance details without losing detail (or with losing a lot less detail than if the image was zoomed with an image editing tool using linear or bicubic zoom).

Some of my results with these models follows. I am using the following test image from here.

Unprocessed Test Image

To best see the differences between the algorithms I recommend you open the x4 zoomed images in new tabs and switch between them.

SRCNN – Super-Resolution Convolutional Neural Network

To see the original paper on SRCNN, click here.
I am using the PyTorch script by Mirwaisse Djanbaz here.

SRCNN x4

SRCNN x4

SRRESNET

To see the original paper on SRRESNET, click here.
I am using the PyTorch script by Sagar Vinodababu here.

SRRESNET x4

SRRESNET x4

SRGAN – Super Resolution Generative Adversarial Network

To see the original paper on SRGAN, click here.
I am using the PyTorch script by Sagar Vinodababu here.

SRGAN x4

SRGAN x4

ESRGAN – Enhanced Super Resolution Generative Adversarial Network

I am using the PyTorch script by Xintao Wang et al here.

ESRGAN x4

ESRGAN x4

PSNR

I am using the PyTorch script by Xintao Wang et al here.

PSNR x4

PSNR x4

Real-ESRGAN

This is the best super sampler here. I am using the executable by Xintao Wang et al here.

Real-ESRGAN x4

Real-ESRGAN x4

Real-ESRNET

I am using the executable by Xintao Wang et al here.

Real-ESRNET x4

Real-ESRNET x4

SwinIR

Very nice results. May be equal to or better than Real-ESRGAN depending on the input image. I am using the code from this colab.

SwinIR x4

SwinIR x4

Differences

Each of the algorithms gives different results. For an unknown source image it would probably be best to run it through them all and then see which gives you the best result. These are not the Hollywood or TV enhance magic fix just yet.

If you know of any other PyTorch implementations of super resolution I missed, let me know.

Availability

You can follow the links to the original GitHub repositories to get the software, but I have also added a simple GUI front end for these scripts in Visions of Chaos. That allows you to try the above algorithms on single images or batch process a directory of images.

Jason.

Text-to-Image Machine Learning

NOTE: Make sure you also see this post that has a summary of all the Text-to-Image scripts supported by Visions of Chaos with example images.

Text-to-Image

Input a short phrase or sentence into a neural network and see what image it creates.

I am using Big Sleep from Phil Wang (@lucidrains).

Phil used the code/models from Ryan Murdock (@advadnoun). Ryan has a blog post explaining the basics of how all the parts connect up here. Ryan has some newer Text-to-Image experiments but they are behind a Patreon paywall, so I have not played with them. Hopefully he (or anyone) releases the colabs publicly sometime in the future. I don’t want to experiment with a Text-to-Image system that I cannot share with everyone, otherwise it is just a tease.

The most simple explanation is that BigGAN generates images that try to satisfy CLIP which rates how closely the image matches the input phrase. BigGAN creates an image and CLIP looks at it and says “sorry, that does not look like a cat to me, try again”. As each repeated iteration is performed BigGAN gets better at generating an image that matches the desired phrase text.

Big Sleep Examples

Big Sleep uses a seed number which means you can have thousands/millions of different outputs for the same input phrase. Note there is an issue with the seed not always being able to create the same images though. From my testing, even with the torch_deterministic flag set to True and setting the CUDA envirnmental variable does not help. Every time Big Sleep is called it will generate a different image with the same seed. That means you will never be able to reproduce the same output in the future.

These images are 512×512 pixels square (the largest resolution Big Sleep supports) and took 4 minutes each to generate on an RTX 3090 GPU. The same code takes 6 minutes 45 seconds per image on an older 2080 Super GPU.

Also note that these are the “cherry picked” best results. Big Sleep is not going to create awesome art every time. For these examples or when experimenting with new phrases I usually run a batch of multiple images and then manually select the best 4 or 8 to show off (4 or 8 because that satisfies one or two tweets).

To start, these next four images were created from the prompt phrase “Gandalf and the Balrog”

Big Sleep - Gandalf and the Balrog

Big Sleep - Gandalf and the Balrog

Big Sleep - Gandalf and the Balrog

Big Sleep - Gandalf and the Balrog

Here are results from “disturbing flesh”. These are like early David Cronenberg nightmare visuals.

Big Sleep - Disturbing Flesh

Big Sleep - Disturbing Flesh

Big Sleep - Disturbing Flesh

Big Sleep - Disturbing Flesh

A suggestion from @MatthewKafker on Twitter “spatially ambiguous water lillies painting”

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

Big Sleep - Spatially Ambiguous Water Lillies Painting

“stormy seascape”

Big Sleep - Stormy Seascape

Big Sleep - Stormy Seascape

Big Sleep - Stormy Seascape

Big Sleep - Stormy Seascape

After experimenting with acrylic pour painting in the past I wanted to see what BigSleep could generate from “acrylic pour painting”

Big Sleep - Acrylic Pour Painting

Big Sleep - Acrylic Pour Painting

Big Sleep - Acrylic Pour Painting

Big Sleep - Acrylic Pour Painting

I have always enjoyed David Lynch movies so let’s see what “david lynch visuals” results in. This one got a lot of surprises and worked great. These images really capture the feeling of a Lynchian cinematic look. A lot of these came out fairly dark so I have tweaked exposure in GIMP.

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

More from “david lynch visuals” but these are more portraits. The famous hair comes through.

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

Big Sleep - David Lynch Visuals

“H.R.Giger”

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

Big Sleep - H.R.Giger

“metropolis”

Big Sleep - Metropolis

Big Sleep - Metropolis

Big Sleep - Metropolis

Big Sleep - Metropolis

“surrealism”

Big Sleep - Surrealsim

Big Sleep - Surrealsim

Big Sleep - Surrealsim

Big Sleep - Surrealsim

“colorful surrealism”

Big Sleep - Colorful Surrealsim

Big Sleep - Colorful Surrealsim

Big Sleep - Colorful Surrealsim

Big Sleep - Colorful Surrealsim

Availability

I have now added a simple GUI front end for Big Sleep into Visions of Chaos, so once you have installed all the pre-requisites you can run these models on any prompt phrase you feed into them. The following images shows Big Sleep in the process of generating an image for the prompt text “cyberpunk aesthetic”.

Text-to-Image GUI

After spending a lot of time experimenting with Big Sleep over the last few days, I highly encourage anyone with a decent GPU to try these. The results are truly fascinating. This page says at least a 2070 8GB or greater is required, but Martin in the comments managed to generate a 128×128 image on a 1060 6GB GPU after 26 (!!) minutes.

Jason.

Adding PyTorch support to Visions of Chaos

TensorFlow 2

Recently after getting a new 3090 GPU I needed to update TensorFlow to version 2. Going from TensorFlow version 1 to TensorFlow version 2 had way too many code breaking changes for me. Looking at other github examples for TensorFlow 2 code (eg an updated Style Transfer script) gave me all sorts of errors. Not just one git repo either, lots of supposed TensorFlow 2 code would not work for me. If it is a pain for me it is going to be a bigger annoyance for my users. I already get enough emails saying “I followed your TensorFlow instructions exactly, but it doesn’t work”. I am in no way an expert in Python, TensorFlow or PyTorch, so I need something that for most of the time “just works”.

I did manage to get the current TensorFlow 1 scripts in Visions of Chaos running under TensorFlow 2, so at least the existing TensorFlow functionality will still work.

PyTorch

After having a look around and watching some YouTube videos I wanted to give PyTorch a go.

The install is one pip command they build for you on their home page after you select OS, CUDA, etc. So for my current TensorFlow tutorial (maybe I now need to change that to “Machine Learning Tutorial”) all I needed to do was add 1 more line to the pip install section.


pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

PyTorch Style Transfer

First Google hit is the PyTorch tutorial here. After spending most of a day banging my head against the wall with TensorFlow 2 errors, that single self contained Python script using PyTorch “just worked”! The settings do seem harder to tweak to get a good looking output compared to the previous TensorFlow Style Transfer script I use. After making the following examples I may need to look for another PyTorch Style Transfer script.

Here are some example results using Biscuit as the source image.

Biscuit

Biscuit Style Transfer

Biscuit Style Transfer

Biscuit Style Transfer

Biscuit Style Transfer

PyTorch DeepDream

Next up was ProGamerGov’s PyTorch DeepDream implementation. Again, worked fine. I have used ProGamerGov‘s TensorFlow DeepDream code in the past and it worked just as well this time. It gives a bunch of other models to use too, so more different DeepDream outputs for Visions of Chaos are now available.

Biscuit DeepDream

Biscuit DeepDream

Biscuit DeepDream

Biscuit DeepDream

PyTorch StyleGAN2 ADA

Using NVIDIA’s official PyTorch implentation from here. Also easy to get working. You can quickly generate images from existing models.

StyleGAN2 ADA

Metropolitan Museum of Art Faces – NVIDIA – metfaces.pkl

StyleGAN2 ADA

Trypophobia – Sid Black – trypophobia.pkl

StyleGAN2 ADA

Alfred E Neuman – Unknown – AlfredENeuman24_ADA.pkl

StyleGAN2 ADA

Trypophobia – Sid Black – trypophobia.pkl

I include the option to train your own models from a bunch of images. Pro tip: if you do not want to have nightmares do not experiment with training a model based on a bunch of naked women photos.

Going Forward

After these early experiments with PyTorch, I am going to use PyTorch from now on wherever possible.

Jason.

TensorFlow 2 and RTX 3090 Performance

A Pleasant Surprise

Just in time for when 3090 GPUs started to become available again in Sydney I was very generously gifted the funds to finally purchase a new GeForce RTX 3090 for my main development PC. After including a 1000 Watt power supply the total cost came to around $4000 AUD ($3000 USD). Such a rip off.

GeForce RTX™ 3090 GAMING X TRIO 24G

The card itself is really heavy and solid. They include a bracket to add support to help it not sag over time which is a nice touch. Like all recent hardware parts it lights up in various RGB colors and shades. These RGB rigs are going to look so out of date once this fad goes away. After upgrading my PC parts over the last few years I now have PCs that flash and blink more than my Christmas tree does when fully setup and lit.

Who needs a Christmas tree?

Not So Fast

I naively assumed that a quick GPU swap would give me the boost in performance that previous GPU upgrades did (like when I upgraded to the 1080 and then to the 2080 Super). Not this time. I ran a few quick machine learning TensorFlow (version 1) tests from Visions of Chaos and the result was either the Python scripts ran extremely slow (around 10x to 30x SLOWER) or they just crashed. So much for a simple upgrade for more power.

Turns out the Ampere architecture the 3090 GPUs use is only supported by CUDA 11.0 or higher. After updating CUDA, cuDNN, all the various Python libraries and the Python scripts I was back to where I was before the upgrades. If you have been through the tedious process of installing TensorFlow before for Visions of Chaos, you will need to follow my new instructions to get TensorFlow version 2 support. Updating TensorFlow v1 code to TensorFlow v2 code is a pain. From now on I am going to use PyTorch scripts for all my machine learning related needs.

High Temperatures

These newer GPUs can run hot. Under 100% load (when I was doing a Style Transfer movie with repeated frames being calculated one after the other) the 3090 peaks around 80 degrees C (176 F). I do have reasonable cooling in the case, but the air being blown out is noticeably hot. The 2080 running the same test peaks around 75 degrees.

The 2080 and my older 1080 seem to push all the hot exhaust air out the rear vents of the card, but the 3090 has no rear exhaust so all the hot air goes directly into the case. I can only assume this is due to them not being able to handle all that heat going “through” the card and out the back, so it needs to just vent that heat anywhere it can. This means if the card is running hot a lot of hot air goes straight into the case. When I touched the side of the case next to the GPU it was very warm.

Apparently 80 and under is perfectly fine and safe for a GPU, but they would say that wouldn’t they. They would be bragging about low temps if they could manufacture cooler running cards.

After some experimenting with Afterburner I lowered the temp limit from the GPU default of 83 degrees down to 75 degrees. This resulted in more throttling but only a slight performance hit (style transfer took 1 minute 21 seconds rather than 1 minute 14 seconds). The case was noticeably cooler and the average temp was now down to a much more chilly 65 degrees. Afterburner allows tweaking (overclocking/underclocking) of your GPU, but the most useful feature is its graphing capabilities to see what is really going on. You can monitor temperatures and throttling as you run complex GPU operations.

Extra Cooling

I wanted to see if more case fans would help, so I removed the current 3 stock case fans and installed 6 of these fans (2 sucking in at the front, 3 blowing out at the top, and 1 blowing out at the rear of the case). My silent PC is no longer silent. I set the GPU back to its default profile with a temp limit of 83 degrees and started another Style Transfer movie to keep the GPU pegged as close to 100% usage as possible for an extended period of time. Watching the temp graph in Afterburner shows peaks still up to 76 degrees, but much less throttling with the core clock graph around 95% to 100% maximum possible MHz when running so that results in a better overall performance.

After a week the extra noise annoyed me too much though so I replaced the Gamdias fans with Corsair fans. 6 of these fans and one of these controllers. Setting the fans to the default “quiet” profile gets the noise back down to near silent sound levels. When I start a machine learning batch run the temp sensors detect the increased heat in the case and ramp up the fans to compensate. Watching Afterburner graphs shows they may even be slightly better at cooling than the Gamdias fans. The problem with the auto-adjust speed control is that there is this noticeable ramping up and down of the fan speeds as they compensate for temp changes all the time (not just when the GPU is 100%). That was more annoying than always on full speed fans. After some adjustments and tests with the excellent Corsair software I settled on a custom temp curve that only cranked up as necessary when I start full 100% GPU load processing. Once the GPU usage drops back to normal the fans ramp down and are silent again.

Power Usage

Using one of those cheap inline watt meters shows the PC pulls 480 watts when the GPU is at 100% usage. Afterburner reports the card using around 290 watts under full load.

I have basically been using the 3090 24 hours a day training and testing machine learning setups since I bought it. 3 weeks with the 3090 usage made my latest quarterly electricity bill go up from $284 to $313. That works out to roughly $1.40 a day to power the GPU full time. If you can afford the GPU you should be able to afford the cost of powering it.

Final Advice

When experimenting with one of these newer GPUs use all the monitoring you need to make sure the card is not overheating and throttling performance. Afterburner is great to setup a graph showing GPU temp, usage, MHz and power usage. Monitor levels when the GPU is under 100% load and over an extended period of time.

Temperature controlled fans like the Corsair Commander Pro setup can work as a set and forget cooling solution once you tweak a custom temp curve that suits your usage and hardware.

Final Opinion

Was it worth spending roughly three times the cost of the 2080 on the 3090? No, definitely not. These current over inflated priced GPUs are not worth the money. But if you need or want one you have to pay the price. If the prices were not so artificially inflated and they sold at the initial recommended retail prices then it would be a more reasonable price (still expensive, but not ridiculously so).

After testing the various GPU related modes in Visions of Chaos, the 3090 is only between 10% to 70% faster than the 2080 Super depending on what GPU calculations are being made, and more often on the slower end of that scale. OpenGLSL shader performance is a fairly consistent speed boost between 10% and 15%.

The main reason I wanted the 3090 was for the big jump in VRAM from 8GB to 24GB so I am now able to train and run larger machine learning models without the dreaded out of memory errors. StyleGAN2 ADA models are the first things I have now successfully been able to train.

StyleGAN2 ADA - Alfred E Neuman

Upgrading the 1080 in my older PC to the 2080 Super is a big jump in performance and allows me to run less VRAM intensive sessions. Can you tell I am trying to convince myself this was a good purchase? I just expected more. Cue the “Ha ha, you idiot! Serves you right for not researching first.” comments.

Jason.

GPT-2 Text Generation

What is it?

GPT-2 is a Generative Pre-Training machine learning model created by OpenAI. The basic purpose of it is to predict what word comes next after a prompt of some seed text. The model was trained on over 40 GB of Internet text. That is an enormous amount of data. Being text only without any images means a lot more text to be used. Estimations on the Internet give approximately 680,000 pages of text per GB. So the 40 GB of text GPT-2 was trained on equates to roughly 27.2 million pages of text!

Originally OpenAI was worried about releasing the AI models publicly because they feared it could be used to auto-generate copious amounts of fake news and spam etc. Since then they have generously released all their models (even the largest with 1.5 billion neural network parameters) for anyone to experiment with.

If you want to use GPT-2 outside Visions of Chaos you can download the code at their GitHub here.

Visions of Chaos front end GUI for GPT-2

I have wrapped all the GPT-2 text generation behind a simple GUI dialog now in Visions of Chaos. As long as you have all the pre-requisite programs and libraries installed. See my TensorFlow Tutorial for steps needed to get this and other machine learning systems working in Visions of Chaos.

You give the model a sentence and after a minute it spits out what it thinks the continued text should be after that prompt. Each time you run the model you get a new unique result.

There is an option for which model to use as on my 2080 Super with 8GB VRAM it cannot handle the largest 1.5 billion parameter model without getting out of memory errors. The 774 million parameter model works fine.

Some example results

What does AI need to do to get rid of us

GPT-2 Text Generation

A nightmare

GPT-2 Text Generation

The future for the human race

GPT-2 Text Generation

How to be happy

GPT-2 Text Generation

These early test results are really interesting. At first I thought the model was just assembling sentences of text it found online, but if you take random chunks of the generated text and do a Google search (in quotes so it searches for the complete sentence) you get no results. The model is really assembling these mostly grammatically correct sentences and paragraphs by itself.

It can be accurate in answering “what is” questions, but then again it can spit out grammatically correct nonsense, so don’t take anything it says as truth.

More to come

A future use I want to use GPT-2 for is a basic chat bot you can talk with. OpenAI’s MuseNet is very promising for generating music and gives much better results than my previous best LSTM results.

OpenAI have also since released GPT-3 with limited access. I hope they also release the model to the general public like they did GPT-2. There are some very impressive results I have seen using GPT-3. GPT-3’s largest model is 175 billion parameters, compared to 1.5 billion for GPT-2. Although if my 8GB GPU cannot handle the 1.5 billion GPT-2 model it will have no hope of using the 175 billion parameter model.

Jason.

DeepDream – Part 3

DeepDream

This is the third part in a series of posts. See Part 1 and Part 2.

DeepDream

ProGamerGov Code

The script from Part 2 supports rendering 59 layers of the Inception model. Each of the 59 DeepDream layers have multiple channels that allow many more unique patterns and outputs.

I found this out thanks to ProGamerGov‘s script here.

DeepDream

There are 7,548 channels total. A huge number of patterns to explore and create movies from. If I followed the same principal as in part 2 and created a movie changing the channel every 10 seconds that would result in a movie almost 21 hours long. If each frame took around 25 seconds to render it would take 1310 DAYS to render all the frames. Not even I am that patient.

DeepDream

Channel Previews

The following links are previews of each layer and available channels within them. The layer, channel and other settings are included so you can reproduce them in Visions of Chaos if required.

DeepDream

As the layers get deeper the images get more complex. If you notice a layer name shown twice it is because it had too many channels within that layer to render into a valid image file so it had to be split into two separate images.

DeepDream

conv2d0_pre_relu
conv2d1_pre_relu
conv2d2_pre_relu

DeepDream

head0_bottleneck_pre_relu
head1_bottleneck_pre_relu

DeepDream

mixed3a_1x1_pre_relu
mixed3a_3x3_bottleneck_pre_relu
mixed3a_3x3_pre_relu
mixed3a_5x5_bottleneck_pre_relu
mixed3a_5x5_pre_relu
mixed3a_pool_reduce_pre_relu

DeepDream

mixed3b_1x1_pre_relu
mixed3b_3x3_bottleneck_pre_relu
mixed3b_3x3_pre_relu
mixed3b_5x5_bottleneck_pre_relu
mixed3b_5x5_pre_relu
mixed3b_pool_reduce_pre_relu

DeepDream

mixed4a_1x1_pre_relu
mixed4a_3x3_bottleneck_pre_relu
mixed4a_3x3_pre_relu
mixed4a_5x5_bottleneck_pre_relu
mixed4a_5x5_pre_relu
mixed4a_pool_reduce_pre_relu

DeepDream

mixed4b_1x1_pre_relu
mixed4b_3x3_bottleneck_pre_relu
mixed4b_3x3_pre_relu
mixed4b_5x5_bottleneck_pre_relu
mixed4b_5x5_pre_relu
mixed4b_pool_reduce_pre_relu

DeepDream

mixed4c_1x1_pre_relu
mixed4c_3x3_bottleneck_pre_relu
mixed4c_3x3_pre_relu
mixed4c_5x5_bottleneck_pre_relu
mixed4c_5x5_pre_relu
mixed4c_pool_reduce_pre_relu

DeepDream

mixed4d_1x1_pre_relu
mixed4d_3x3_bottleneck_pre_relu
mixed4d_3x3_pre_relu
mixed4d_3x3_pre_relu
mixed4d_5x5_bottleneck_pre_relu
mixed4d_5x5_pre_relu
mixed4d_pool_reduce_pre_relu

DeepDream

mixed4e_1x1_pre_relu
mixed4e_3x3_bottleneck_pre_relu
mixed4e_3x3_pre_relu
mixed4e_3x3_pre_relu
mixed4e_5x5_bottleneck_pre_relu
mixed4e_5x5_pre_relu
mixed4e_pool_reduce_pre_relu

DeepDream

mixed5a_1x1_pre_relu
mixed5a_3x3_bottleneck_pre_relu
mixed5a_3x3_pre_relu
mixed5a_3x3_pre_relu
mixed5a_5x5_bottleneck_pre_relu
mixed5a_5x5_pre_relu
mixed5a_pool_reduce_pre_relu

DeepDream

mixed5b_1x1_pre_relu
mixed5b_1x1_pre_relu
mixed5b_3x3_bottleneck_pre_relu
mixed5b_3x3_pre_relu
mixed5b_3x3_pre_relu
mixed5b_5x5_bottleneck_pre_relu
mixed5b_5x5_pre_relu
mixed5b_pool_reduce_pre_relu

Individual Sample Images

DeepDream

I was going to render each layer/channel combination as a 4K single image to really show the possible results, but after seeing it would take 15 minutes to generate each image I was looking at nearly 79 days to render all the example images. HDV 1920×1080 resolution will have to do for now (at least until the next generation of hopefully much faster consumer GPUs are released by Nvidia).

DeepDream

Even using two PCs (one with a 1080 GPU and one with a 2080 Super GPU) these images still took nearly 3 weeks to generate. Each image took 6 minutes on a 1080 GPU and 5 minutes on a 2080 Super GPU. Since working with neural networks and GPU computations (especially these week long all day sessions) I can see they do have a noticeable impact on my power bill. These GPUs are not electricity friendly.

DeepDream

See this gallery for all of the individual 7,548 channel images. Starts at page 4 to skip the more plain images from the initial layers.

DeepDream

Movie Samples

The following movies use a series of channels that follow a basic theme.

Eye imagery.

Architecture imagery.

Furry imagery.

Trypophobia imagery.

Availability

DeepDream Dialog

As long as you setup the TensorFlow pre-requisites you can run DeepDream processing from within Visions of Chaos.

Tutorial

The following tutorial goes into much more detail on using the DeepDream functionality within Visions of Chaos.

Jason.