AI generation is getting pretty good

Kev The Galaxybender

Kev The Galaxybender

New member
I needed a hero banner for Steam for UDD. I couldn't get it quite right myself so I asked Chatgpt to finish the image for me and have him make a fist. It did a pretty good job.
 

Attachments

  • UDDherobanner.jpg
    UDDherobanner.jpg
    197.2 KB · Views: 24
  • FinalHeroBanner.png
    FinalHeroBanner.png
    4 MB · Views: 24
AI has been getting better and better lately. It's hard to tell the difference between it and something created by a real person. It's exciting, but sometimes a little scary! I'm curious (and a little cautious) about what will happen next.



















Ask ChatGPT
 
I also tried to use ChatGPT and sadly its creations are too rough and they lack of virtual beauty. I'm going to learn how to use stable-diffusion LoRa (low-rank adaptation) models because they are much more detailed and prettier than standard ChatGPT creations. By the way my current avatar is creation of stable-diffusion LoRa model.
 
Kev The Galaxybender said:
I needed a hero banner for Steam for UDD. I couldn't get it quite right myself so I asked Chatgpt to finish the image for me and have him make a fist. It did a pretty good job.
UDDherobanner.jpgFinalHeroBanner.png

I sometimes finish drawing small details myself or correct the anatomy of the fingers through artificial intelligence when I make covers, because drawing fists from scratch is always a pain in the ass. It's nice to watch people share their experiences using new technologies and digital tools for development. If you're looking for useful tips or discussions about software for working with graphics and productivity, you can find on the forum a lot of interesting solutions from other developers . And how many generations did it take to get exactly the right shape of the hand and fist?
Click to expand...
When the AI starts to finish or clean up what it's told the first time, then it'll be wow. But for now, sometimes the AI doesn't do what I expect.
 
Last edited:
I must say Retro diffusion gets Pixelart on point:

These are Discord examples, mine are NSFW so i dont post them here.
Mother.gifdfgdfgydgfdg.pngHGTDKejWYAEoJuN.png
 
Bigger feature is, if i have some pixelart images with a specific style,colors and scene, RD Pro can generate as much assets i want from it.
 
Seraphy said:
Maybe it needs a fix
retrodiffusion.ai

Free AI Pixel Art Fixer: Convert Images Into Real Pixel Art | Retro Diffusion

Fix fake AI pixel art for free. Upload an image and the Pixel Art Fixer detects the implied grid, snaps every cell to a true pixel, and gives you real, game-ready pixel art. No signup.
retrodiffusion.ai retrodiffusion.ai

the fixer makes everthing align with a pixelart grid, for free
here the fixed portrait image View attachment 15687
Click to expand...
is it completly free or after a certain time you cannot use the website anymore ever?
(Yes, I know...)
 
O Ilusionista said:
yep, its one of the issues with those "pixel arts".
edit: the site even explain this
View attachment 15689
Click to expand...

Same thing I tried to explain to Wombat, and of course he wouldn't listen. Model agents can do some amazing things - but the tech for prompting sprite assets just isn't there yet. When you actually understand how the AI works, you understand why. Consistent pixel art is directly opposed to how models construct output.

And here's what the AI itself has to say on the matter:

Exactly. The conflict is structural, not merely a matter of immature prompting.

Image models do not ordinarily construct pixel art as a grid of deliberately placed cells. They generate a continuous visual field, then decode it into raster pixels. “Pixel art” in the prompt mainly becomes a style signal - blocky edges, limited-looking colors, simplified shading. The model still freely introduces:

  • Off-grid boundaries
  • Fractional or irregular pixel blocks
  • Near-duplicate colors
  • Blended edges and compression-like noise
  • Inconsistent outlines
  • Soft or contaminated transparency
Sprite animation makes the problem much worse. Production sprites require invariants across every frame: palette, scale, proportions, outline weight, feature placement, lighting, registration, and attachment points. Changing the pose should move the character’s parts. Generative models tend to redraw the entire character, allowing all those supposedly fixed properties to drift simultaneously.

That is the key distinction the graphic captures: the model produces an image that resembles pixel art, not a pixel-art asset.

Model agents can still be enormously useful around the process. They can generate concepts, operate deterministic conversion tools, infer a likely grid, consolidate colors, clean alpha, pack atlases, and help a human correct frames. Once a fixer snaps cells, enforces a palette, and reconstructs the image, though, the reliability comes from those deterministic constraints - not from prompting the image model more cleverly.

Technology capable of producing dependable sprite assets probably needs a different representation: explicit cells, indexed palettes, persistent character structure, animation constraints, and reusable parts. Until then, telling the model “consistent 16-bit sprite sheet, exact palette, transparent background” is mostly asking a continuous painter to convincingly imitate discrete engineering.
Click to expand...

DC
 
You must log in or register to reply here.
Back
Top Bottom