Thoughts on AI image generation

Text to Image

Whenever I look at generated images I want to edit them, especially those that are a little 'off'.

For example these cats were generated using the Black Forest's Schnell model with the prompt: "Angry cat in Japanese style."

The upper lips are too thin, also the toes are off by one. This cat has an extra human-like hand, and its furless foot looks odd.

With Reference image

When I gave a specific reference image the end result still looks off. This is the original

"Treachery of images" by Rene Magritte, 1929

and here's the generated image with a twist. An image I generated with Gemini 2.5 Flash

Perhaps editing the original would've been better, as I only wanted to change the text.

The original painting's text read: "This is not a pipe". It's a representation of a pipe.

I changed it to say: "This is not AI." It's a representation of an AI-generated image.

Prediction

AI tools are affordable now; their prices will rise after a consolidation. Similar to how Uber and Lyft's fares were low to starve off competitors, and now they're high.

In AI's case the big providers eg. Google will drive other providers out of business before raising prices.

How to avoid vendor lock-in?

One way is to use open source models (they're on Huggingface), but it's hard for them to compete against better-funded models.

Another option is to continue to develop core skills (eg. coding/drawing/writing/editing/animation) and avoid being dependent on AI tools. One can still use them to augment existing skills, but they should not be a crutch.