Dicefire Update - Sprite Progress, and A Small Rant About Generative Art.

Where are we?

Another week down, another small step closer to the finish line. The main focus of this week has been figuring out the art problem. The quick rundown is that so far, this game has over 100+ unique assets to build. As stated in the last update, these were AI generated assets and they’re bad. More detail about that in a moment. So because we could not use the generated assets in any tangible way, I created a tool I nicknamed “Pixominable”. Thanks to this program and a lot of prompt engineering, we are now making usable assets!

---

AI Slop

Shown off as an image in the last update, we basically created a pretty but useless piece of art. The reason it’s so dang useless is because at normal zoom, it looks fantastic, but once you try to scale it down, you’re left with a garbled, broken mess of an asset. I’ll list out the grievances…

1. What looks like a solid color at a distance is actually thousands once you zoom in on the asset. Because of this behavior, shrinking the asset to size breaks pixels, distorts, and otherwise mangles the image, creating an imperfect sprite.
2. The asset is not uniform. If you placed it on a pixel grid, you would have weird, wobbly lines that make no sense.
3. Checkered patterns are nearly impossible to make without looking as if someone smudged the tile with their finger...
4. If I needed to animate an asset as is, it creates non matching sprites every time. Zero consistency.
5. Person opinion, I don't want to look at a sprite and know it's AI made.

---

The Fix: Pixominable

Maybe I’m too picky. Granted. But I wanted a clean asset, not something I’d have to grit my teeth and apply to a lazy, janky map. So I thought of Aesprite and the map builder we assembled. I thought, why not just give the AI a brush? So after some dice rolls, I decided on the name Pixominable and got to work.

Pixominable is a lightweight series of scripts that does several things. For one, it’s a tool specifically for AI to use and operate. Pixominable creates a grid of whatever size you need, and a limited palette of colors each assigned a number. The program then works one of two ways:

1. It takes an AI generated image, maps it from corner to corner, places it on the grid, and redraws the asset with only approved colors.
2. It creates a grid and can be used by AI as a canvas and brush to plot out the image.

The best part is, it’s not generated, it’s assembled. Each plot on the grid has to have a corresponding number in relation to the palette. All AI is doing is painting by numbers. Once it runs pixominable, the application takes the map and builds the image. This has created some headache here in there, but otherwise things are working out.

To the right you’ll find one of the first assets to get touched: a keyboard. Don’t let the image fool you, that top image is thousands of pixels. the bottom one is more around 60-80 pixels and is a proper SNES-like sprite.

---

The New Problem

Detail. AI struggles with abstraction and nuance. The biggest hurdle was a VOIP phone and for one good reason: The AI, between Claude and GPT, would not use the sample photos of a VOIP phone to build the receiver. What they did, and admitted to doing, was building based off their IDEA of a phone. What resulted was a series of barbells and bananas on top of an otherwise good asset. We eventually got it right, but man…it took 2 days to do so.

The fix, though tedious, was prompt engineering fine tuning. Forcing the agents to focus on depth, connection, color, shape, continuity, etc. This has resulted in much better assets which I am reserving to show. I want to do more testing and honing on these prompts. The idea is to get this rolling out without the need for human intervention.

---

What's Next?

The Dicefire work continues. I’m hoping to be halfway done with assets by end of next week if at all possible. This has become HT’s strongest focus at this time. Hoping to have a positive update next round