
In January 2025 I wrote to a friend of mine:
When it comes to AI art, I’ve been deeply impressed. Just like human-made poor writing is everywhere, so is human-made poor art. I realize “art is subjective” – and that applies to writing too – but as someone who deeply appreciates artistic impressionism, poetic writing, and even interpretative dance, I feel there is a lot of room for intentional, carefully crafted “messiness,” and expressive impressions which aren’t fully defined. There is a lot of room for statement-making uncanniness, and odd composition. In my somewhat controversial opinion, there is also just bad art – where not much thought or care was really put into it, or where the creator is just not very skilled. But AI-generated art has proven to be capable of surpassing beauty, creating masterpieces that would have taken hundreds of hours for me to personally create in mere seconds. I can’t help but be impressed and excited.
My friend also brought up the issue of recompense for artists. Again, this is already an issue. There is a reason we have a phrase, the starving artist.
I am concerned about recompense for original artists. We completely need real artists. But I don’t think the crux of this issue lies with the AI technology itself. I believe this is an economic and political issue of tremendous scale.
We need a way to ensure everyone’s basic needs are met, regardless of whether machines can “do it better.” People should be able to make art, get educated, and have every enriching experience they desire regardless of whether it is “profitable.” This speaks to the need for some sort of universal basic income or a similar notion that provides people with their needs – which goes beyond clean water, food, and shelter. People also need beauty, rest, and mental stimulation.
People don’t need “jobs” to be healthy; they need the autonomy and empowerment to do things that feel good to do, such as helping a neighbor or painting an artwork. So I don’t think the issue is whether people are getting paid enough (or at all) to be artists, but whether we are moving toward a world where we simply meet people’s needs, regardless of things like intergenerational wealth and technological power.
Again and again, when people tell me about the problems they see with AI – from copyright issues, to energy usage – I feel that AI is merely magnifying the existing problems with humanity: how we have exploited the planet and one another, and how our systems do not cater to the everyday person (or ecosystem).
Most recently I learned that cooling the servers used to give us AI-powered programs requires an insane amount of water. People in the United States (I hear) are running out of water because of this . . . And there are no regulations against this use of water. But once again, companies have been stealing the fresh drinking water of citizens – often illegally stealing from so-called protected sources – since the dawn of industrialization. This has happen in the US before, and it has been ongoing in third-world countries for a very long time.
One of personal drives to shop exclusively for organic food (and clothing!) was not just for my personal health, but because I discovered that almost all conventional food (and clothing) products are linked to atrocities – from abused work forces, slavery, ecological destruction (including cutting down old-growth rain forests), and literally taking the entire water supply from entire villages in Africa to make soda for Americans – causing hundreds of people to either die of literal dehydration, or leave their homes and lives entirely behind and walk dozens (if not a hundred) miles to the next village to beg for water.
Because AI uses tremendous amounts of energy, it is absolutely a political concern. I admonish you to only utilize AI in ways that you feel genuinely enrich your life and nourish your soul. Please do not use AI as a frivolous tool to clutter up the internet with more absurdism. (Of course, there is almost zero chance that someone who is likely to do that would even land on this website to begin with!)
This may seem like a heavy way to begin a mostly fun-centric article about making AI-art, but I feel we must learn to enjoy life whilst simultaneously holding space for the truth that our systems and institutions are failing to serve their purposes.
I will explore some further philosophical concerns toward the end of this article, but for now, I want to dive into telling you about my personal favorite AI-art creation platform. I tried a dozen of them at the beginning of 2025 seeking the perfect fit for creating the creatures for my latest board-game idea. It was a perfect fit for AI art because I needed a hundred different artworks of a fantasy creature I had invented in my mind, but these artworks would be about the size of a quarter or beer cap. It would have taken me a year to do the artwork the long way.
As it was, I still spent a couple months doing the vector artwork (such as the logo and symbols for the game) the long-form way. Making a board game is not a small project even with AI assistance in the decoration department! And AI-generation does not replace the need for an actual artist to judge, arrange, select, combine, and edit as needed.
My board game creature artwork is one of the most perfect applications for AI-created art. It would be repetitive for a real artist and become droll quite quickly. And beyond that, it would be cost prohibitive for a small game company to hire an artist to do quality work.
Getting Started with NightCafe: A Guide
NightCafe the most engaging and user-friendly AI (artificial intelligence) art platform I have tried. Whether you’re here to make a single image for a project or to immerse yourself in the community, this guide will help you get the most out of it. Please do check out my profile on NightCafe.
How the Credit System Works
NightCafe’s free model is surprisingly generous:
Baseline Credits – You start with a set amount of free credits (plus some “pro credits,” which I recommend saving for important projects unless you plan to subscribe).
Earn More Credits – You gain additional credits by:
Publishing creations
Participating in chats and following ten other users
Liking and commenting on other people’s work
Participating in the daily challenge
This gamified engagement encouraged me to experiment beyond my original purpose (which was creating images for a board game) and helped me learn the platform faster. And because you earn credits for these actions, it is worth it to dip your toes in the various features of the platform. In the beginning, I tried everything and enjoyed it.
Baseline Credits – You start with a set amount of free credits (plus some “pro credits,” which I recommend saving for important projects unless you plan to subscribe).
Earn More Credits – You gain additional credits by:
Publishing creations
Participating in chats and following ten other users
Liking and commenting on other people’s work
Participating in the daily challenge
My Quick-Start Recommendations
Complete the “onboarding” actions to earn extra credits right away (publish creations, like others’ work, join the daily challenge, etc.).
Search for “open prompt gang” and experiment with prompts that catch your eye.
Try the same prompt in different models to see how output changes.
Complete the “onboarding” actions to earn extra credits right away (publish creations, like others’ work, join the daily challenge, etc.).
Search for “open prompt gang” and experiment with prompts that catch your eye.
Try the same prompt in different models to see how output changes.
Why I Prefer Prompt Revision Over Image Revision
Early on, I learned it’s often more effective to revise the prompt instead of trying to edit an existing image.
Yes, it can be a little sad to lose elements you liked, but improving your prompt-writing skills yields better, faster results in the long run. In the world of AI-image generation, prompt mastery is your artistry.
Learning From Others’ Prompts
One of NightCafe’s best features is the ability to see other creators’ settings and prompts – if they choose to share them.
-
Look for the “open prompt gang” tag – This tag signals that the creator is openly sharing their prompt for educational purposes.
-
How to View a Prompt – Open any image, go to “Settings” (location depends on your interface: beside or under comments), and check if the prompt and settings are visible.
You can then adapt prompts from images with a style or realism level you like.
Example:
My most popular creation – View Here – includes the full prompt and settings I used. I also applied a Creative Upscale, which adds detail without changing the composition.
Note: Creative Upscale is best used on images that already have good hands and faces. It’s about 50/50 for fixing faces, and it won’t fix hands.
Choosing the Right Model
NightCafe offers multiple AI models – each with its strengths. Here are some of my favorites:
-
Flux – Excellent for prompt accuracy; one of the best at capturing specific details. Excellent for people and buildings, but it does not seem to understand basic plants. (I asked for basil once and it gave me a plant that had no resemblance to basil.)
-
Google Imagine – Strong prompt adherence and clean style. This model can be so realistic that it is almost droll at times. When I asked for a beaded dragon necklace from many different models, I was surprised to find that the one that Google Imagine generated was so realistic it almost looked boring by comparison to the others.
-
Boltning – Consistently visually appealing, even if prompt adherence is looser. The composition and colors are always striking, but once you have used it a lot, it begins to seem repetitive, as if it is stuck in its own style (which it is).
-
Crystal Clear – Polished, vibrant visuals.
-
DALL·E 3 (Pro) – Versatile with a broad creative range. This is excellent for creating artistic renditions of something. But be sure to toggle prompt revision off if you want to understand how your prompt relates to your actual creation.
-
Ideogram – Handles text better than most, but AI text is still hit-or-miss – often easier to add text later in a separate program.
And many others, including video-creation models. Also, new models are coming out all the time, so eventually all the above will be outdated.
Example Collection:
I created a multi-model prompt experiment to show how different models interpret the same text prompt in wildly different ways.
For instance, in this example, only Flux managed to include the reins specified in the prompt, while Boltning and Crystal Clear produced the most visually appealing results despite skipping that detail.
Prompt Writing
The Anatomy of a Prompt
Strong prompts tend to contain:
-
Subject and details – The main focus of the image, and what it is doing, like: A gorgeous woman with long, red, curly hair is sitting with her legs crossed wearing a haughty expression bordering on a smirk.
-
Background and setting – The models know exactly what you mean when you say to put something in the background. So you can specify something as simply as galactic background, or the background is coral reef, and you will get what you asked for. You also influence the overall vibe and setting when you specify the background you want.
-
Composition cues – Camera angles, shot types, and perspective details. This includes phrases like three-quarter view, top-down perspective, or isometric sketches.
-
Stylistic references – Art movements, artists, or mediums. This can be something like traditional Japanese painting, or ancient Egyptian wall art, or large brush strokes.
-
Quality and stylistic cues – Phrases like highly detailed, ultra realistic, cinematic lighting, or minimalist.
Vibe and mood – This includes descriptive words for the setting (such as futuristic), and for the tone (whimsical). The end of your prompt could read: The overall tone is bright, solarpunk and cheerful.
Example:
A majestic white Pegasus, standing atop a mossy cliff at dawn. Epic sun rays. Ultra realistic yet with fantasy flourish. Cinematic lighting with dramatic contrasts. Soft pastel clouds in artistic swirls. Misty mountain background. The whole image is in warm tones, with fiery red highlights and gilded gold elements.
Structuring Your Prompt
NightCafe interprets prompts linearly. In other words, the order of your words matter. Think of it as “weighting” by position. The earlier a word or phrase appears, the more important the AI considers it.
Typical practice:
-
Put the subject first. If you care more about the style, put some of the style prompt at the beginning such as watercolor painting of . . . And then your subject. Write more details about the style at the end.
-
Follow with the most critical modifiers.
-
End with secondary details and style influences.
Note: The more subjects your image has, the more likely your result will be poor. AI images of a single subject, such as one cat or one human tend to be superb. Even images of couples (whether that is a human couple or a pair of boats) are also often quite good. However, once you expand to three or more subjects, the quality tends to dramatically decrease.
Example:
Intricate layered artwork of a flying goth steampunk clockwork metallic cat with spread wings. Atmospheric landscape, futuristic silhouette, and imaginative lighting, featuring subtle gradient textures, and nuanced shading, with a mix of sharp and soft edges, set against a vibrant, purple-indigo galaxy night sky background, with gear details and atmospheric depth.
Using Style References Effectively
NightCafe allows you to guide the AI toward specific visual styles by referencing:
-
Art movements: Art Nouveau, Impressionism, Surrealism
-
Artists: Studio Ghibli, Moebius, Klimt, Escher
-
Mediums: oil painting, charcoal sketch, 3D render, alcohol ink, gauche, stained glass
Tip: You likely will not want to use more than three style references in one prompt to avoid confusing the model.
Note: One can argue that borrowing styles from your favorite artists is a form of stealing from them. I am unsure if I personally would argue this because the nature of art is borrowing techniques and styles from our favorites. (Long before AI we noticed the massive number of duplicates on Deviant Art from artists tracing each other.)
Composition and Perspective
Adding composition cues greatly influences the final image:
-
Shot type: portrait, close-up, medium shot, wide shot, aerial view, full body view*
-
Perspective: isometric, fisheye lens, low-angle, over-the-shoulder
-
Framing: symmetrical, rule of thirds, centered subject
Example:
“An isometric view of a cozy cob cottage in a snowstorm, glowing windows, warm interior light spilling out. The cottage is nestled in a gorgeous snow-shimmering forest. High detail, watercolor painting with alcohol ink.”
Lighting and Atmosphere
Lighting cues set tone and realism:
-
Natural: golden hour, overcast, moonlit, sunrise glow
-
Artificial: neon lighting, candlelight, studio lighting
-
Mood: dreamy, ominous, ethereal, gritty
Example:
“A gorgeous canvas painting of a stone temple at dusk with warm golden-hour light. The temple has a huge arched doorway which is open and emitting a soft purple glow. Large brush strokes describe mist is rising from the ground around the temple. Misty trees surround the temple. Myriad candles, some square and some round, overflow every window with wax dripping down the stone walls of the temple. Some of the candles are deep-eggplant-purple, but most are white. Most of the candles are very small, but some are very large. Cinematic shadows. Dramatic moonscape sky. Warm tones, acrylic fantasy painting, brush stroke texture.”
Controlling Complexity
Too many ideas in a prompt can overwhelm the AI – particularly subjects. I know I wrote this above a couple times, but here is another reminder: if you describe six different cats in your image, you may not get what you asked for. The more subjects, the more likely you are to get surrealism, or something that is not what you requested.
Pro tip: Start simple. If it’s close to what you want, add modifiers gradually in new runs or evolutions. Creative upscale is one of my favorite tools for making an image more complex.
Example:
“A beautiful pastel artwork of six cats. The first cat is a large orange tabby with an attitude: she is glaring at the black cat. The black cat is indifferent and looking away, with a white marking under their chin. On top of the black cat is a tiny purple cat with faerie wings. In the background is the fourth cat: a large gray panther looking elegant in a sprawling position. The fifth cat is a fluffy white cat looking at the viewer with startling blue eyes. Beside the white cat is a sleek silver Siberian cat with extra long whiskers who is licking their paw. Artwork is done in detailed, meticulous pastel.”
The fact that the number of cats here is correct, and that the faerie cat is on top of the black cat is a small miracle.
Model-Specific Prompting
NightCafe models interpret prompts differently:
-
Flux / Google Imagine – Best for prompt adherence and realism. Use precise, literal descriptions.
-
Boltning / Crystal Clear – Great for aesthetics. Focus on mood or vibe more, worry less about specifics.
-
DALL·E 3 – Strong for conceptual prompts; good balance of adherence and creativity.
-
Ideogram – Best for text in images, but still hit-or-miss.
Example Test Method:
-
Write one detailed prompt.
-
Run it in Flux Schell and Boltning. (These two are cheap.)
-
Compare – does one capture your intended detail better?
Prompt Weighting & Loras
NightCafe supports weighting through parentheses and colons (depending on the model).
This lets you tell the AI “this part is more important.” This becomes particularly relevant when using community-made models, where you can assign a weight to different models you are applying. To use community-made “loras,” you will start by selecting the appropriate model (so far, I only know this to be available with Flux). You can select up to four of them.
In this example I used three loras, including one I made myself using photos of myself. I used a weight of two (instead of the standard one) on myself to ensure the image would have my face. Then I used a very low weight on one of the loras because my experience told me that full weight did not give me the results I wanted. (I wanted the image to have a gothic flare, not a dark vibe.)
Example:
“A beautiful fantasy painting of a powerful beautiful <lora:Raederle:2.0> woman with rockstar black hair, tight leather one-piece long sleeved bodysuit with many chains, lace, zippers, high neckline, buckles, and jewelry. She is singing into a microphone. Color scheme is black, white, and purple. Goth <lora:Gothic:0.2> aesthetic. Snowy grunge setting. <lora:Goth Grunge Beauty:1.0> Snowdrop flowers on the ground.”
Negative Prompting
Negative prompts tell the AI what to avoid. You can use the “Negative Prompt” field for things like:
-
“blurry, deformed hands, text, watermark, signature, extra limbs”
-
“low contrast, pixelated, overexposed”
I rarely use this feature, but it can be useful, in particular for avoiding signatures and watermarks. Because so many images contain these by default, the AI often invents a signature or watermark to apply to your image. Of course, this makes it seem like you stole the image from someone else. While it is trivial to remove these in most cases, it makes sense to avoid them in the first place.
Iteration Workflow
If you have a complex idea in mind, worry less about trying to get perfection in one shot.
-
Draft – Write a prompt.
-
Test – Generate with at least three different models to see which model is most appropriate for your idea.
-
Refine – Choose one or two models to continue working with, and adjust your original prompt to fix major issues. I like to generate a single image at a time so I can make subtle changes along the way.
-
Evolve – Use NightCafe’s “Evolve” with low noise for subtle variations. I have not used this feature much myself, but it is one of the most popular ones on NightCafe.
-
Upscale – Apply Creative Upscale to the best image for extra polish. If you just want more detail, apply 10% or so creativity. If you want to see what other wild directions the image could go, try 60% of more creativity.
When to Use . . .
-
Evolve – When you like most of the image but want small changes in detail or pose.
-
Creative Upscale – When you want to keep composition but add detail. But be warned, this can change faces tremendously, and with the creativity high you might find some subjects in your image have changed entirely to other subjects.
-
Selective Edit – For localized changes. This works very well for removing watermarks or signatures, but terribly for fixing deformed hands.
-
Prompt Rewrite – When multiple areas need improvement or if the hands are poor. For deformed hands or badly shaped faces, I do not try to edit the image: instead, I generate from scratch.
Prompt Debugging Checklist
If your output isn’t desirable:
-
Is your main subject at the start of the prompt?
-
Have you overloaded it with too many subjects?
-
Are you mixing conflicting cues (e.g., “ultra realistic” + “abstract expressionism”)?
-
Did you remember to turn off Prompt Magic for cleaner adherence?
-
Have you tested multiple models?
Creative Upscale
Creative Upscale is one of my favorite features on NightCafe. It can be likened to adding fine brushstrokes to a painting you already like. It elaborates on your image without making fundamental changes – if you set creativity low. At low creativity (around 10%), it will mostly just polish and increase the size of your image (without any resolution loss, of course). At high creativity (around 50%), you might create large elemental shifts even though the composition remains the same. For example, a bucket on someone’s head could easily turn into a hat, and a beard might turn into the shadows on a feminine face.
Creative upscaling an image, however, is how you can get the most intricate and incredible works. If you are looking for something large enough resolution to use for a book cover or wall art, you will want to use this tool in combination with a media-editing program. I like to generate at least three different creative upscales and then use a graphic design program to combine my favorite features of each one together. This is the method used for creating the cover art on my most of my songs and poetry; a powerful example of this can be seen here with the artwork for Trillion Codes. This image composites seven different creative upscales – a process I spent hours on. This is still more efficient (by far) than creating such an image from scratch, but it does require some art skills.
Beware: creative upscale is terrible at fixing hands, and only about fifty-fifty on fixing faces (like a lazy eye or strange jawline). Use upscaling tools only when your base image is already very strong.
Q&A
Q: As an actual artist . . . What do you feel is different between AI art prompting and manual art creation (with a brush or pen)?
I think the biggest difference between slow-art and AI-co-creation (in terms of user experience) is that slow-art provides a very different learning experience.
When I’m AI-co-creating, I’m learning about how AI thinks, how prompts work, and how different AI models think. I am only learning minimal amounts about what makes good art.
In contrast, when I’m creating slow-art, I’m constantly making observations based on my reference image about the nature of shapes, and how forms evoke impressions and feelings.
I’m also learning with slow art about what looks good in a different way: instead of relying on AI to figure out a good posture, I have to figure out manually what makes an elegant tail or an attractive placement of birds. The AI seems to already be a master at composition and rarely makes compositional errors of artistic judgment (IMO; this is highly subjective). Composition is key to artistic mastery. Although, in my case, I seem to have always had a natural knack for it (I was frequently complimented on mine as a child).
All of the above said, learning prompt engineering isn’t just a matter of learning how the AI models think. You do have to learn terminology related to the medium you wish to emulate. This will be easier if you wish to create photo-realistic imagery – which AI excels at because it has been trained on so many photographs. However, if you wish to emulate an artistic style you will need to learn whether it is alcohol ink, gouache paints, watercolors over fine inkwork or something entirely different.
The same rules apply regardless of whether you’re working with visual arts, auditory arts, or text outputs. You can not ask for something if you do not know what it is called. Thus, if you are already familiar with art and art terms, it will aid you in prompting the model to generate what you’re seeking. It seems that many of my favorite creators on NightCafe are, indeed, already artists in the traditional sense.
Q: How do I create a cartoon image that is not anime? Does the model matter?
A: Yes, the model does matter. Some models are very capable of this, such as Google Imagen, but others, such as Boltning, are not. See this prompt:
“Clean black linework on a completely white background. Two stick figures with circle-shaped faces are holding hands and smiling at each other. A cute heart is above their heads. They are walking through grass although the grass is just hatch marks.”
Because I wanted the image to be very clearly a cartoon, I started my prompt with clear black linework. There is little room for interpretation there. If I said cartoon, I might get any sort of cartoon style the model is most familiar with, or that is associated with the other terms I used.
Q: Can I transfer my NightCafe credits to another user?
A: No, NightCafe doesn’t allow credit transfers. The good news is that it’s easy to earn credits through daily logins, maintaining a log-in streak (which affords an extra 25 credits per five days), liking creations, and other site activities. For many users (myself included), credits tend to build up over time rather than run out.
Q: Is it addictive?
A: Yes! But only at first. There is an element similar to gambling, as well as gamification in other ways. You more-or-less gamble each day on how many credits you will get, and then, every time you generate from a prompt, it is a gamble whether your payout (your creation) will feel exciting or disappointing. But like all novel things, eventually it stops being novel and you will do other things. Admittedly, I did have late-night NightCafe sessions during my first few weeks – I was stunned by how much beautiful artwork I found there!
Q: Is there a way to change just one detail in an image instead of generating a completely new picture?
A: Yes, this is called in-painting (in NightCafe, the “Selective Edit” tool). You can mask the area you want to change and provide a prompt for the edit.
That said, results vary. Textures, colors, and styles may not line up perfectly because the edit uses a separate AI model.
In my experience:
-
Creative Upscale with a moderate creativity setting (e.g., 45%) can sometimes change backgrounds while keeping the subject intact.
-
Evolving an image lets you control “noise” from 0% (similar) to 100% (entirely new).
-
For complex fixes – especially hands – it’s often more efficient to tweak your prompt and regenerate from scratch. I have blundered through twenty credits trying to edit an image, when I could have generated five to twenty new attempts with the same credits.
Q: What is “Prompt Magic” and should I use it?
A: Prompt Magic uses an extra processing step to hypothetically improve image quality, but it can also make your prompt less “pure” by adding interpretive flourishes. I recommend turning Prompt Magic off to see exactly how the model is interpreting your words – otherwise, you have introduced so much noise into the process that it will be harder to learn effectively.
Q: Can I delete creations I don’t like?
A: Yes – creations can be deleted or archived. Archiving keeps them off your public profile without permanently removing them.
A: LoRA (Low-Rank Adaptation) lets you train AI on a specific style or subject using custom images. For example, I trained a LoRA on photos of myself and prompted (more or less), “Raederle in a doll shop, wearing a purple fridge dress.” The result was highly amusing. Everything in the scene – dolls included – had my face. 😂 It’s a reminder that AI will sometimes over-learn its subject in unexpected ways.
Q: What about the debate over AI art being “soulless” or “meaningless”?
A: Many people worry that AI-generated works lack “soul” or emotional intent. My perspective has evolved away from that notion. Some creators revisit the same theme over months – iterating on a “city in a teacup” until it became breathtaking – and to me, that feels like the work of an artist. As someone who spent decades refining their skills with traditional arts (pencil, watercolors, acrylic paints, markers, linework, and others), I still respect the effort that goes into prompt engineering. It is still an artistic skill. It is not, however, the same skill as drawing. Not by a long shot.
AI accelerates execution, but the vision, selection, and persistence are still human. We are collectively choosing to use these tools, in my opinion, because AI art easily surpasses the beauty any artist could achieve in the same time frame. The AI is doing most of the work for us. Still, we still remain the judge of what it is that we wish to display.
Q: Is AI writing as good as AI art?
A: In my experience, not yet. AI struggles with long-form storytelling, often producing repetitive or clumsy text. But that’s not unique to AI – low-quality human writing is everywhere too. AI poetry and short formats (like haiku) tend to fare better, especially when the structure is constrained.
Q: What’s the real issue with AI art – ethics or quality?
A: For me, it’s more about economics and society than the tech itself. We need systems (like universal basic income) that allow people to create and live well regardless of whether machines can “do it better.” Artists should be free to make art because it’s fulfilling, not because it has to be profitable to survive.
Q: So did you use NightCafe to create the art on Raederle.com?
A: As of 2025, yes, I did do an overhaul on my website which utilizes NightCafe art generations. The somewhat consistent style is something I developed through much experimentation with prompt engineering. However, older posts still contain sketches and vector artwork I created long-form through more traditional methods. While I do not really claim to have created artwork when AI is involved, I did prompt it, and I did so with intention. Furthermore, I do own the rights to the art I have prompted, so please refrain from borrowing it. These days it is easy enough to prompt your own – and now you know how!