CrAIyon (formerly DALL-E after Salvador “Dali”) is a web-facing art-making generative AI tool online (https://www.craiyon.com/) that enables the uses of text (and image) prompts for the creation of watermarked, lightweight visuals. Counterintuitively, the rough visuals are much more usable for recombinations and remixes and recreations into usable digital visuals for various digital learning objects. The textual prompts are not particularly intuitive because of how the generative AI program was trained on mass-scale visuals). There is an art and occasional indirection to working prompts after each try, with the resulting nine-image proof sheets that CrAIyon outputs. The tool can be used iteratively for different outputs.
The tool sometimes turns out serendipitous surprises, including an occasional work so refined that it can be used / shared almost unedited. One challenge in using CrAIyon comes from their request for credit (for all non-subscribers to their service). Another comes from the visual watermarking (orange crayon at the bottom right of the image). However, this tool is quite useful for practical applications if one is willing to engage deep digital image editing (Adobe Photoshop, Adobe Illustrator).
Transcript: New from BookNet Canada for 2024: Loan Stars - Tech Forum 2024
Fashioning Text (and Image) Prompts for the CrAIyon Art-Making Generative AI
1. Fashioning Text (and Image) Prompts
for the
CrAIyon Art-Making Generative AI
(and Revising for Visual Effects)
Shalin Hai-Jew
Kansas State University
2. Presentation Overview
• CrAIyon (formerly DALL-E after Salvador “Dali”) is a web-facing art-
making generative AI tool online (https://www.craiyon.com/) that
enables the uses of text (and image) prompts for the creation of
watermarked, lightweight visuals. Counterintuitively, the rough
visuals are much more usable for recombinations and remixes and
recreations into usable digital visuals for various digital learning
objects. The textual prompts are not particularly intuitive because of
how the generative AI program was trained on mass-scale visuals).
There is an art and occasional indirection to working prompts after
each try, with the resulting nine-image proof sheets that CrAIyon
outputs. The tool can be used iteratively for different outputs.
3. Presentation Overview (cont.)
• The tool sometimes turns out serendipitous surprises, including an
occasional work so refined that it can be used / shared almost
unedited. One challenge in using CrAIyon comes from their request
for credit (for all non-subscribers to their service). Another comes
from the visual watermarking (orange crayon at the bottom right of
the image). However, this tool is quite useful for practical
applications if one is willing to engage deep digital image editing
(Adobe Photoshop, Adobe Illustrator).
7. CrAIyon Generative AI
• CrAIyon is a web-facing art-making generative AI available at
https://www.craiyon.com/.
• Formerly, this tool was known as DALL-e Mini (after Salvador “Dali”).
• In the current iteration, there are four classes of visuals: art, drawing,
photo, and none. “None” seems to enable the non-selection of a
class per se and leaving it to the generative AI to select.
• The free version of the CrAIyon site is replete with ads.
• The output images are also marked with a visual watermark (an
orange “crayon” at the bottom right of the generated image).
8. Terms of Use
• CrAIyon has released its images for use in “personal, academic or
commercial use” based on Craiyon LLC’s Terms of Use and FAQs.
• This endeavor is supported by subscriptions and advertising revenue.
• There is a CrAIyon app on the Google Play store, but apparently there
are emulators, so the makers advise using a link to their direct app.
9. Rights for Craiyon LLC to Use Submitted
Prompts
• The trade-off: “You hereby grant to Craiyon, its successors, and
assigns, a perpetual, worldwide, non-exclusive, sublicensable, no-
charge, royalty-free, irrevocable copyright license to reproduce,
prepare derivative works of, publicly display, publicly perform,
sublicense, or distribute any prompts (in any form) you enter into the
Site and any Images produced by the Services at your direction. This
license survives termination of this Agreement by any party for any
reason.”
10. Ensuring Legal and Ethical Approaches
• “You agree not to make use of the Site, our models, or derivatives of
our models:
• To infringe upon the intellectual property rights of any third party;
• In any way that violates any applicable law or regulation;
• To create content that exploits or abuses children, including, but not limited
to, images or depictions of child abuse, sexual abuse, or presenting children in
a sexual manner.
• To generate or disseminate verifiably false information with the purpose of
harming others; …
11. Ensuring Legal and Ethical Approaches(cont.)
• …To impersonate or attempt to impersonate others;
• To generate or disseminate personally identifying or identifiable information;
• To defame, libel, disparage, bully, threaten, stalk, or otherwise harass others;
• To create content that promotes self-harm;
• To create content that implies or promotes support or funding of, or
membership in, a terrorist organization.
• To create content that condones or promotes violence against people based
on race, ethnicity, color, national origin, religion, age, gender, sexual
orientation, disability, medical condition, veteran status, or any other
protected legal category.”
12. Limits to Commercial Usage for Non-
Subscribers
• “’Commercial Use’ means any use case intended to generate direct or
indirect financial gain. You may use the Site in connection with any
Commercial Use provided that, if you are not a Subscriber, you must
credit Craiyon in text accompanying any image(s) you use
commercially. In the sole event that it is not possible for text to
accompany such image(s), the placement of our logo in the corner
constitutes sufficient attribution.
• “Any Commercial Use of the Site inconsistent with our Code of
Conduct or the rest of these Terms is prohibited.”
13. Subscriber Packages
• Craiyon LLC offers various levels of subscription:
• $5 a month for Supporters
• $20 a month for Professionals
• $? Custom pricing for Enterprises based on the requisite level of service
• The billing is done annually.
14. “Proof Sheets”
The generative AI outputs nine candidate
images in a “proof sheet.”
Any of the images may be expanded,
examined, and downloaded.
Prompts may be re-run, verbatim, or with
textual variations.
If an image is of particular interest, it may be
used to seed a next-generation iteration,
with more visual and textural refinements.
15. CrAIyon Generative AI File Outputs
• From some months of experience using CrAIyon, it seems that…
• The generative AI seems to go more for visual description than style transfer.
• The visual works seem to be object-oriented.
• Often, the background is a basic matte one. There is less context without an
illustrated background.
16. CrAIyon Generative AI
File Outputs(cont.)
The images come out in .webp (“weppy”)
format.
The format was created by Google engineers.
It is lightweight but visually informative.
The .webp format can be transcoded to most
of the popular digital image formats.
The images are usually 8 – 10 KB in size.
The spatial resolution tends to be 72 pixels
per inch…and about 3.5 x 3.5 inches.
The spatial resolution of the images may be
“jumped,” and the image may be re-sampled
using modern digital image editing tools.
17. CrAIyon Generative AI
File Outputs(cont.)
The “Screenshot” function outputs the 9-
image “proof sheet” as a .png file.
The pixels per inch (ppi) is only 72. However,
the height-width are much bigger than the
single images.
Digital images for the Web may be at 72 ppi,
but print usage will require higher dpi (350
dpi).
19. CrAIyon LLC Advice re: Text Prompts
• The company advises users to find words that describe the visual style
of interest. (This presenter thinks style may be more important even
than the substance of the designed visual.)
• The text prompts should be “specific”.
20. Text Prompting
• It helps to have direct text prompts in some cases if one already
knows pretty much what one wants and can state it fairly clearly.
• CrAIyon does respect a number of different languages.
• One may indicate preferred style.
• It is not so easy to indicate compositing or layout.
• The generative AI tends to be fairly literalist.
• The limited file size and canvas size means that the visual can only
carry so much complexity.
21. Text Prompting(cont.)
• Sometimes, one may be in an exploratory or discovery phase and may not
have a direct idea of the type of visual desired.
• It may help to conduct an Image search on Google with various prompt
terms to acquire some ideas.
• Sometimes, an indirect text prompt may be helpful to see what CrAIyon
comes up with.
• Learning the tool and its strengths and weaknesses involves a lot of tries. It
involves going down various alleys. There is a “brute force” (in an ITS
sense) aspect that is inefficient.
• Can one predict what sort of image will come out with a particular prompt? If so,
that may suggest that one is better understanding the tool. One can perhaps start
trying to make inferences about the model and / or the training image data.
22. Text Prompts
What You Want
• A text prompt can specify what
is desirable.
What You Don’t Want (“Negative
Words”)
• Negative words may be used to
indicate what is not wanted.
23. Image Prompting
• From a proof sheet of images, one can select any of those and select
the “Upscale” button at the bottom left to have the AI re-draw the
image based on the current image as a new starting point.
• This is one way to refine a rough initial exported image into
something that may / may not be closer to what one wants. The
iterations offer a way for people to communicate with the machine
and vice versa.
• If none of the nine in the proof sheet look right, one can revise the
initial text prompt and go again and again.
24.
25. Computer-Aided Text Prompting
• Once images are generated, CrAIyon has a feature that enables other
text prompts along the same lines as the initial but with elaboration.
Their “Try this one” offers ways to elaborate, often informed by
history.
• The AI “intelligence” goes beyond visuals.
26. What CrAIyon is “Good” and “Bad” At
Current Strengths
• Shapes
• Speed of image generation
(relatively speaking)
• Numbers of options
• Additional user features being
added to the site over time
Current Weaknesses
• Appearance of 3d
• Actual 3d
• No polygonal shapes
• Difficulty with creating alpha
channels behind objects
• Counting typical number of
limbs and hands
• Drawing hands
27. What CrAIyon is “Good” and “Bad” At(cont.)
Current Strengths Current Weaknesses
• Text (not)
• Numbers (not)
• Symbols (not)
• Mapping (not)
• Grids (curvy lines, imprecision)
• Scenarios (low-res)
• Scenes (low-res)
28.
29.
30.
31. Using Others’ Visuals for Prompting
• Start with “Search” at the top.
• Review the discovered images.
• Click the orange “Generate more” button at the bottom right of the
image.
• The text used to generate that image will show in the CrAIyon page.
• Click the “Draw” button next to the pre-populated text prompt field.
(One can revise the text first before clicking “Draw.”)
32.
33.
34. A Lack of Implicit Knowledge
• The explanation for the trouble with hands and limbs, various physics,
image implications, and other issues…is that the generative AI lacks
“implicit” “in-world” knowledge. It has learned from a curated set of
visuals…and it is emulating visual understandings of terms based on
the labeling on the imagery.
• The AI of neural networks enables a bottom-up understanding of the
visuals details.
48. A Valley Amidst Mountains in the Style of
Helen Frankenthaler
49. Dilemma: Art-Making Generative AI and
Human Artists…
• Does art-making generative AI elevate human artists in the
reductionist style transfer?
• Or does it denigrate human artists?
• Or neither?
51. …A Responsive
Magical Universe
The person is a magician
The right combination of incantations will
enable the person to receive the prize
The reward comes from the universe
52. …A Robot Assistant
User is the delegator
Ask and will receive
Command
Another team member to delegate to
Lesser work to the machines
(years and years to acquire the skillset to
achieve level of output of the machines; only
a small elite have those actual skills)
53. … A Slot Machine
Put in a request and get out what one needs
sometimes and nothing useful in another
Generative AI is a gamble
One has to take one’s chances
One has to be aware of the investment
(learning, effort, time expenditure)
When it works, the generative AI may feel a
little more like a vending machine
54. … A Wishing Well
Offering some concepts (and perhaps a
seeding image) and receiving something
dreamily potent
55. … A Grocery Store
Mass production of digital visuals
Off-the-shelf visuals
Generic products
Light packaging
Meets requirements
56. … A Mechanical Turk
A machine that is fast and easy-to-use
A machine that turns out quality work
Sometimes visual suitability may be the issue
Can be a workhorse
57. … A Pixel Studio /
Workshop
Provides resources that can be reshaped to
meet particular needs
Turns out products in “pixel clay”
58. …An Informal Market
Acquiring digital goods on the “down low” (in
illicit ways)
Visiting less traveled parts of town
Going with undocumented and unexplained
work (non-crediting for machines)
Trading prompts and seeding (or base)
images for visual goods that have been
“laundered”
Engaging in transactional relationships
Products noticeable by some styles
59. … A Game
A fun challenge to see what may be elicited
A playful approach
60. Dilemma: Most Appropriate Metaphors for
Art-Making Generative AI
• Which of the prior metaphors makes the most sense in this case?
Why?
• Or if none of the prior, what would you suggest? Why?
62. A Typical Sequence
• Removal of visual watermark (erasure, color selection and removal,
clone stamping, and others)
• Background removal
• Reshaping form
• Application of texture
• Compositing
• Crediting
64. Professional Needs for Images in Academia
• Decorative / break up gray text
• Illustrative
• Exemplification
• Explanation
• Style analysis and others
65. Does CrAIyon Meet Professional Needs?
• A basic cost-effect approach (inputs for outputs) is a simple way to
assess whether it is worth going with an art-making generative AI like
CrAIyon.
• How much work goes into the prompting? The downloading? The digital
image editing?
• Are there particular dominant looks and feels in terms of the output?
Too much of a visual telltale may be a showstopper (as in stopping the
work with the generative AI dead).
• Is the generative AI adapting? Is it improving?
• Is it backed by a solid company? Does it have a solid reputation?
66. Does CrAIyon Meet Professional Needs?(cont.)
• It helps to be professionally and constructively picky. It helps to be
persistent.
• In many cases, stars will “align” on an image, and there is
convergence on something that works.
• However, it is better to return the pixels back to the ether though
than prematurely commit to a visual that is not quite relevant for the
learning context.
67. Crediting
The challenge is to correctly identify the
source (CrAIyon) without keeping a
distracting watermark at the bottom right
corner.
Things get complex if a visual from CrAIyon
has been edited…or was only used as a
reference image, and so on.
The credit should ride with the image. It
should be seeable by others. How the image
was made should be clearly communicated
especially if the visual is something that is
photorealistic and may be confused with
something actual.
68. Can a Human Artist Compete?
• Unless a human artist is widely trained on a number of historical art
styles…has a perfect sense of compositing…has a variety of color
palette expertise…and is also super prolific, there is no real way to
compete.
• Even if people could take analog artworks, they would have to scan,
digitally edit, and rework a file for the digital realm.
• Even if people could take digital stock images to seed their work, they
would have to develop the work…wrangle inspiration…and turn out a
workable file. Creating art, even with many available resources, can
be a work-laden process.
69. Conclusion and Contact
• Dr. Shalin Hai-Jew
• Instructional Design
• ITS
• Kansas State University
• 785-532-5262
• shalin@ksu.edu
• The images in the “metaphors” section were
created in the Deep Dream Generator. The
others are from the target art-making
generative AI program CrAIyon. The two
systems are not linked. Works from both
programs were included because of utilitarian
concerns.