So you want to get better at AI image generation. Good timing—the tools have never been more accessible, and the learning curve has never been gentler.
Here's what nobody tells you at first: prompting isn't really about knowing secret keywords. It's about learning to describe what you see in your head in a way a machine can understand. That's a skill, and like any skill, it improves with practice.
Start with What You Want to See
Before you type anything, close your eyes for a second. What's the image? Not the prompt—the image. What do you actually want to end up with?
A lot of beginners jump straight to typing without a clear vision. They type "cool robot" and get... something. Then they're frustrated it's not what they imagined. But they never really imagined anything specific in the first place.
The best prompters I know spend time looking at art, photography, illustration. They're building a visual vocabulary. They know the difference between chiaroscuro lighting and flat lighting. They know what "analog film grain" looks like versus "clean digital." They know because they've seen it, not because they read a list of magic words.
So first: know what you want. Then describe it.
Describe Like You're Talking to a Smart Alien
AI models have seen billions of images. They understand visual concepts incredibly well. But they're not mind readers—you have to tell them.
Think of it like describing an image to someone over the phone. You wouldn't say "make it look cool." You'd say "a woman standing in rain at night, neon signs reflecting on wet pavement, blue and pink lighting, cinematic, shallow depth of field."
Be specific about:
- Subject: What or who is in the image?
- Setting: Where is this happening?
- Lighting: How is it lit? (natural, studio, golden hour, dramatic, neon)
- Style: What medium or aesthetic? (photograph, oil painting, digital art, anime)
- Mood: What feeling? (peaceful, tense, whimsical, noir)
The more specific you are, the less the model has to guess. And when the model guesses, you get random results.
Style Keywords Actually Help
Okay, I said they're not magic words. But certain terms do push outputs in particular directions:
Art styles: impressionist, digital art, concept art, oil painting, watercolor, ukiyo-e, art nouveau, brutalist, baroque
Mediums: photograph, illustration, 3D render, pencil sketch, digital painting, film still
Lighting: golden hour, dramatic lighting, studio lighting, natural light, neon, cinematic, rim lighting, backlit
Camera terms (for photorealistic): depth of field, bokeh, wide angle, macro, telephoto, 85mm, shallow focus
Quality indicators: highly detailed, 4k, masterpiece—these work better on some models than others
The trick is using these as nudges, not crutches. A great prompt with no style keywords will still beat a mediocre prompt with every keyword thrown in.
Negative Prompts: What You Don't Want
Many platforms (including Artfelt) let you specify what to avoid. This is surprisingly powerful.
Common negative prompts:
blurry, low quality, distorted, deformed, uglytext, watermark, signature(keeps the model from adding these)cartoon, anime, illustration(if you want photorealistic)
Negative prompts are especially useful for cleaning up persistent artifacts. If your model keeps adding an extra finger, add extra fingers, deformed hands to negatives.
Seed: Your Secret Weapon for Consistency
Every generated image has a "seed"—a number that determines the starting randomness. Same seed + same prompt + same model = same image.
Why does this matter?
Say you generate an image that's almost perfect. The composition is right, but the lighting is slightly off. If you use the same seed and adjust your prompt, you'll get a variation of the same base image rather than something completely different.
This is how you iterate. You don't start from scratch each time—you refine.
Your First 100 Generations Are Practice
This is the most important thing: expect to generate a lot of images you won't use.
Photographers take hundreds of shots to get one great image. AI artists do the same. Your hit rate will improve with practice, but even experienced prompters generate 20-50 images to find the one they want.
Don't let this discourage you. It's not failure—it's the process.
Different Models, Different Strengths
Not all AI image generators are the same:
- SDXL (what Artfelt uses): Excellent for artistic and photorealistic work, struggles with text
- Midjourney: Distinctive aesthetic, even simple prompts look "artistic," Discord-only interface
- DALL-E 3: Follows instructions precisely including text, but has a more clinical look
- Fine-tuned models: Specialized for anime, architecture, realism, etc.
The same prompt on different platforms will give different results. Learn what works where.
Curation Is Part of the Art
You generate. You review. You select the best. You share the best.
That's not cheating. That's curation. It's as much a creative act as the prompting itself. Knowing which image from a batch captures what you were going for—developing that eye—is a skill.
Start Creating
There's only so much you can learn from reading. The real education happens when you start generating.
Go to artfelt.ai/create. Type something. See what happens. Adjust. Try again.
You'll get better faster than you think. The gap between "a cat in space" and "a Russian blue cat floating through a cosmic nebula, stardust swirling around it, deep purple and gold color palette, digital painting, dreamlike atmosphere" is smaller than you'd expect.
You just have to start.