Skip to content
Calaway Solutions
Lessons Learned

I Got Tired of Paying for AI Video Credits, So I Started Making Videos on My Own Computer

My experience with ComfyUI, local AI video models, and a few spectacular failures.

6 min read

AI video generation has gotten pretty impressive.

Tools can turn a few sentences into realistic scenes, animate characters, and create footage that would've been difficult to produce without a camera crew just a few years ago.

There's just one thing that started bothering me.

Every time I wanted to experiment, I was spending credits.

I was testing video generation through ElevenLabs, and one video used approximately 13% of my monthly credits.

One video.

And here's the thing: when you're learning something new, your first result usually isn't what you wanted.

You need to experiment, change things, make mistakes, and try again.

I didn't want every bad result to feel like wasted money.

So I started looking into generating AI videos directly on my own computer.

Wait, You Can Generate AI Videos Without Paying Per Video?

Yes.

You can download AI models and run them locally using your own hardware.

There are no per-generation cloud credits when you're using locally hosted models in a local workflow.

You still pay for your computer, electricity, and potentially storage upgrades. And it can be significantly slower than commercial services.

But you get something I really value:

The freedom to experiment without watching a credit balance disappear.

The software I decided to try is called ComfyUI.

It's an interface that lets you build AI generation workflows by connecting different components together.

You can use it to generate images, create videos, animate existing images, and build customized workflows.

And you don't need to be a programmer to start.

What I'm Running

Here's the computer I've been experimenting with:

Component My setup
Graphics card NVIDIA RTX 3090, 24 GB VRAM
Processor AMD Ryzen 5 3600
Memory 32 GB DDR4
Operating system Windows 11
Software ComfyUI Desktop

It's not the newest computer.

In fact, I'm currently exploring upgrades because some of these AI models are really pushing it.

But it can generate AI videos.

My first five-second video took approximately 15 minutes.

Another, longer generation took more than 30 minutes.

Compare that to the speed of some cloud video generators, and you'll immediately see the tradeoff.

Local generation isn't automatically faster. It's about having more control and avoiding per-video fees.

How to Try It Yourself

Here's the basic process I followed.

Step 1: Download ComfyUI

Visit the official website:

https://www.comfy.org/download

Download the Desktop version for your operating system.

I used Windows with an NVIDIA graphics card.

ComfyUI handles much of the setup for you, so you don't necessarily need to manually configure Python and a bunch of dependencies.

Step 2: Check Your Available Storage

This one is important.

AI models can be huge, and video generation can require substantial additional memory and disk space.

I learned this the hard way.

My SSD filled up with downloaded models, and I eventually had to reorganize my storage.

My recommendation is to decide where you're going to store your models and finished videos before downloading everything.

In ComfyUI Desktop, look for the storage settings and configure your model and output directories.

A separate SSD with plenty of free space is ideal. A hard drive can also store models, although loading may be slower.

Step 3: Choose a Video Model

ComfyUI offers example workflows and templates that make getting started much easier.

I've been experimenting with two models:

LTX-2.3

This was my first major video-generation experiment. It produced some visually impressive results, although longer scenes with multiple characters had continuity problems.

Wan 2.2

I started testing Wan because I wanted to compare character movement, visual quality, and consistency.

Each model has different strengths, hardware requirements, and generation times.

Don't assume the largest model is automatically the best option for your computer.

Step 4: Start Small

Seriously, don't start by asking it to generate a 30-second cinematic masterpiece.

Start with a short video.

For example, try a simple prompt:

A cinematic shot of a black sports car driving through a rain-soaked city at night. Red neon lights reflect across the pavement. The camera tracks alongside the vehicle. Realistic movement, dramatic lighting, detailed reflections.

Try approximately five seconds first, using the workflow's default settings.

Once that works, start experimenting with more complex scenes.

Step 5: Generate and Review

Submit the workflow and let your computer do its thing.

Depending on your hardware and model, this could take minutes or considerably longer.

When it's finished, watch the result and decide what you'd change.

That's where things get interesting.

The Results Have Been… Interesting

One of my first experiments involved a couple having an argument in an apartment.

The characters looked surprisingly believable.

Until they started turning into each other.

Literally.

The woman's appearance would morph into the man's, and vice versa.

So I adjusted the prompt to describe each character more clearly.

I also wanted a more attention-grabbing opening.

Apparently, I went a little too far with that instruction.

The next video featured a woman practically flying into the scene with a chair.

Not exactly what I had in mind.

But that's part of experimenting with these models.

Sometimes they follow your instructions.

Sometimes they interpret them in ways you never imagined.

And sometimes your computer decides it's had enough and freezes.

Yes, that happened too.

What I'm Learning About Making Better Videos

The biggest lesson so far is that typing more instructions doesn't necessarily produce a better video.

If you're generating everything from text, the AI gets to decide what the scene looks like, where characters appear, and how the camera is positioned.

I'm now exploring a different approach:

Start with an image, then use text to animate it.

For example:

  1. Create an image with the exact characters, setting, lighting, and composition you want.
  2. Load that image into an image-to-video workflow.
  3. Describe what should move and how the camera should behave.
  4. Generate a short clip.
  5. Combine multiple clips into a finished social media video.

This gives you greater control over the starting shot, although the model can still introduce unexpected changes.

Eventually, I'd like to build reusable workflows that generate music videos, social content, and other creative projects without rebuilding everything each time.

Is It Worth Doing?

That depends on what you're looking for.

If you want a polished video immediately, a cloud service might be the easier choice.

If you enjoy experimenting, already have a capable gaming PC, or want to learn how AI generation actually works, local AI is worth exploring.

There are real tradeoffs:

  • Generations can be slow.
  • Large models require significant memory.
  • You'll need storage space for model files.
  • Not every generation will turn out well.
  • Some workflows require troubleshooting.
  • Commercial use depends on the license of the model you're using.

But being able to generate another video without purchasing more credits is pretty appealing.

And for me, learning what I can build with the hardware I already own is half the fun.

What's Next?

I'm still experimenting with ComfyUI, LTX, Wan, and other tools.

My goal is to build a workflow where I can choose images, control a scene's composition, describe the motion, and produce short videos for social media.

I'll share what works, what doesn't, and probably a few more ridiculous failures along the way.

If you're trying local AI video generation yourself, send me a message.

I'd love to hear what hardware you're using, which models you've tried, and what kind of results you're getting.

And if you're coming here from TikTok or Instagram after commenting VIDEO, welcome! This is the guide I wish I'd had when I started experimenting.

Building things. Learning things. Occasionally breaking things.

Dealing with something like this?

Tell us how it shows up in your business. We'll help you find the simplest practical next step—no pressure, no obligation.