You Don’t Need to Pay Someone to Learn AI Image Generation

You Probably Don’t Need to Pay Anyone to Learn AI Image Generation

I’ve been doing AI image generation basically since the beginning, and at this point I’m probably somewhere around 100,000 generations across different models, UIs, workflows, and whatever else has come and gone over the last few years.

The biggest thing I’ve learned is that people make this stuff seem way more complicated than it actually is.

There are a lot of people right now trying to capitalize on their “knowledge” of AI image generation. Paid Discords, paid workflows, paid prompt guides, paid tutorials, courses, all of it.

And I’m not saying nobody should ever charge for anything. If somebody builds a genuinely useful model, tool, workflow, or course, obviously there can be value in that.

But I really don’t like the idea that beginners are being made to feel like there’s some secret layer of knowledge they need to pay somebody to access.

There isn’t.

Most of what I know came from doing the exact same thing anyone else can do: generating a ridiculous amount of images, changing things, seeing what happens, breaking stuff, figuring out why it broke, and doing it again.

That’s basically it.

You learn by generating

People want a shortcut where somebody explains the perfect settings, perfect prompt structure, perfect sampler, perfect workflow, whatever.

But AI image generation changes so fast that half of that information is outdated six months later anyway.

The thing that actually carries over is intuition.

After enough generations, you start to know when the model is fighting you. You start to understand what parts of the prompt matter and which parts are basically doing nothing. You get a feel for when you need a different model instead of a different prompt. You learn what LoRAs actually help and which ones are just destroying the image.

That doesn’t come from buying somebody’s PDF.

It comes from generating.

And the nice part is you really don’t need 100,000 generations to start getting it. You improve pretty quickly if you actually pay attention to what you’re changing instead of copying settings blindly.

Krea 2 is a good example of where this is going

Krea 2 is one of the reasons I feel even stronger about this now.

Image generation is getting easier.

You tell it what you want, you generate, you adjust it, you generate again. That’s how it should work.

The model should be doing the hard part.

You shouldn’t need to feel like you’re programming a nuclear reactor just to make a picture.

Which brings me to ComfyUI.

I think ComfyUI makes this way more complicated than it needs to be

I know people love ComfyUI, and I use it too.

It’s extremely powerful.

If you’re building custom pipelines, doing weird multi-model workflows, automating generations, using ControlNet in specific ways, chaining upscalers, doing custom conditioning, or experimenting with things that normal interfaces don’t expose, ComfyUI is great.

But I also think the AI image community has convinced itself that all of that complexity is necessary.

It isn’t.

For a huge percentage of generations, ComfyUI is basically taking something that should be:

choose model → write prompt → generate

and turning it into a wall of boxes connected by noodles.

Once you understand it, it’s not that difficult. But why should somebody brand new have to understand what a VAE decode node is before they can make a good image?

They shouldn’t.

A lot of ComfyUI workflows are genuinely impressive.

A lot of others are 35 nodes doing something that should probably be a button.

Complicated doesn’t automatically mean better

This is probably one of the biggest things I’ve noticed over the years.

People see a giant ComfyUI workflow and assume the person must know some insane secret technique.

Sometimes they do.

Sometimes it’s just complicated.

Those aren’t the same thing.

There’s almost a weird incentive in AI communities to make things look technical. If your workflow looks incomprehensible, it makes whatever you’re doing seem more valuable.

Then somebody packages that knowledge up and sells it.

Meanwhile, some guy using a basic interface and actually understanding composition, prompting, reference images, lighting and what model to use is making better images.

The final image doesn’t care how many nodes were in your workflow.

Civitai already gives you an insane amount of information for free

This is what makes the whole thing kind of funny to me.

You’re already on Civitai.

There are thousands of models here.

There are example images.

There are LoRAs.

There are workflows.

There are generation settings.

There are comments where people troubleshoot things.

There are people uploading their entire process for free.

Find an image you like and actually study it.

What model did they use?

What LoRAs?

What resolution?

What prompt?

What did they do differently than you?

Try it yourself.

Then change something.

That’s how you learn.

I still do this.

I’ve generated an absurd number of images and I still see stuff where I think, how did they get it to do that?

Then I figure it out.

That’s half the fun.

You don’t need to know everything

Another thing beginners do is think they need to understand every setting before they start.

You don’t.

Learn what a checkpoint is.

Learn what a LoRA does.

Understand seeds.

Understand resolution.

Learn basic image-to-image and inpainting.

Mess around with guidance, steps and samplers enough that you understand the general effect.

That alone gets you surprisingly far.

Then when you actually run into a problem that requires ControlNet, regional prompting, masks, IP adapters, whatever else, learn it then.

There’s no reason to memorize an entire diffusion pipeline before you even know what you want to make.

Paying for convenience is different

I also don’t want this to sound like I think nobody should ever spend money on anything AI related.

I’ve paid for tools.

I’ll pay for something that saves me time.

If somebody makes an amazing model, I have no issue supporting them.

If somebody builds a workflow that would take me five hours to recreate and sells it for $10, maybe that’s worth it.

That’s paying for convenience.

What I’m talking about is paying because somebody convinced you that you’re incapable of learning this without them.

That part is bullshit.

You absolutely can.

There isn’t some hidden council of AI artists with secret prompt knowledge.

Most people who are actually good at this got good the same way.

They generated a lot.

Just make images

If you’re new, my best advice is honestly just to stop worrying about whether you’re doing it the “right” way.

Go make images.

Make bad ones.

Make weird ones.

Change one setting and see what it does.

Download a LoRA and completely ruin an image with it.

Try a different model.

Copy somebody’s settings and then change them.

Eventually you stop thinking about half of this stuff because it becomes instinct.

I’ve watched AI image generation go from feeling like some weird technical experiment to something almost anybody can pick up and use.

And tools like Krea 2 are only going to push it further in that direction.

The technology is supposed to make creating easier.

Don’t let somebody convince you that you need to pay them to make it complicated again.