One night, two GPUs, no artist: how far can original game art go? Everything here was made from scratch on 30 October 2026 by local AI models, code and Blender. Tap Mimsy, then any picture to see it full size.
Every pixel and sound in this scene was generated tonight: background, hero, skeleton, Mimsy, effects and sounds.
Background: Krea 2. Characters: FLUX.3 pictures animated by MiniMax H3, then turned into pixel art. Effects: Wan-Alpha. Sounds: Stable Audio 3 and MiniMax H3. Mixed pixel sizes are left as they came, on purpose.
A solo developer can now make a whole game's worth of consistent art locally, if a human picks the look and the keepers.
Each row is one method I tried. "Time" is for one finished asset on the GPU box (RTX 5090 or RTX PRO 6000).
| Method | Best for | Quality | Time | Ready to use? |
|---|---|---|---|---|
| FLUX.3 Klein (image model) | Everything; fastest; best at edits | High | 5 s | Yes |
| Qwen-Image 2.1 | Icons with real transparency; side views | High | 13 s | Yes |
| Qwen-Image 2512 | The richest painted look | High | 51 s | Yes |
| Krea 2 Turbo | The most convincing pixel art; side-view backgrounds | High | 5 s | Yes |
| MOSS | Cute, vivid characters; side-view backgrounds | High | 13 s | Yes |
| Anima | Anime and chibi, with tag prompts | Medium | 8 s | Only with tag-style prompts |
| Editing one approved picture | Gear tiers, skins, same face every time | High | 9 s | Yes |
| A whole icon set in one picture | Item icons in bulk, all in one style | High | 10 s per 16 to 20 | Yes |
| Snap to a pixel grid + cut out | Turning AI "pixel style" into real pixel art | High | 2 s | Yes |
| Video model MiniMax H3 | Idle, chomp, attack and hit loops, with sound | High | 30 to 45 s per loop | Yes (licence: see Animation) |
| Video model Wan 2.2 | Simple loops | Good; fast moves warp | 93 s per loop | For gentle loops |
| Effects with transparency (Wan-Alpha) | Slashes, coin bursts, magic, smoke | High | 90 s per effect | Yes |
| One picture to a 3D model | Same character from every angle | High | 3.5 min per model | Yes, for key characters |
| Blender, built by code | Rock-steady animation | Plain | About 1 s a frame | For simple shapes |
| Sprites from code only | Placeholders | Low | Instant | No |
| Stable Audio 3 | Sound effects, short music | Needs your ears | 3 s | You decide |
| ACE-Step 1.5 | Music at an exact tempo, key and length | Needs your ears | 12 s per track | You decide |
Twelve art directions for the same scene, so you can point at one.
Each look shows key art of the hero with Mimsy and a mock phone screen. Made with FLUX.3 Klein, about 5 seconds each. The screens are mood boards, not real UI: the text is gibberish and small icons smear.
FLUX.3 Klein wins overall: top quality, 5 seconds a picture, and it can edit. Qwen-Image 2.1 is the best second tool.
Every model got the same 10 Chompchest subjects in the same 4 styles: pixel, painted, chibi 3D and inked.
5 s per picture. Follows prompts closely; styles stay clearly different; Mimsy is great in all four.
13 s. Best at "side view, facing right"; real pixel style; strong app icons. Lost the chest on dragonbone Mimsy.
51 s. The most finished, painterly look, but the four styles look alike.
5 s. Bold and clean; the most convincing real pixel art; draws true side-view backgrounds.
13 s. Vivid and cute; side-view backgrounds; Mimsy gets an odd orange creature inside.
8 s. Built for anime tags, so plain sentences confuse it: odd heroes and a headless golem.
Editing one approved picture keeps the face and pose while the gear climbs tiers. This is the core of Chompchest.
Each picture is an edit of the first one, never a fresh drawing, so nothing drifts. About half the tries are keepers. Edits are bad at new poses (those come from video, below) but great at skins.
Ask for a whole set in one picture and every icon shares one style. 10 seconds makes 16 to 20 icons.
A script then cuts the set apart and turns each icon into real pixel art. The bottom picture is 40 of them in a mock inventory.
AI "pixel art" is fake: pixels of every size, hundreds of colours. Snapping it to a grid makes it real.
A matting model (BiRefNet) cuts the character out, then Pixel Snapper fits one grid for every asset, so the whole game shares one pixel size. It beat my own converter and two other tools in a side-by-side test.
A video model animates an approved picture into a loop. MiniMax H3 is the clear winner: on-model, real attacks, and it adds sound.
Each clip gets the same picture as its first and last frame, so it loops. Every frame is cut out and snapped onto one grid, so the pixel versions don't shimmer. Wan 2.2, the open-licence option, warps fast moves. Licence note for later: MiniMax H3's licence excludes the US, EU, UK and South Korea, so shipping its output needs a commercial deal.
| Sound made by the video model with the clip | MiniMax H3 |
|---|---|
| Mimsy chomp |
Wan-Alpha makes glowing, see-through effects in 90 seconds each: drop them straight over any scene.
Each effect has true soft transparency (glows and smoke fade into whatever is behind them), which normal image models can't do.
Pixal3D + TRELLIS.2 turned the approved Mimsy picture into a textured 3D model in 3.5 minutes, faithful from every side.
Blender then renders it from any angle, so every sprite matches exactly; here an 8-direction turnaround, also as pixel art. Next step for this route: split off the lid and rig it, for chomps that never drift.
Code alone gives tidy placeholders, not a game's look.
Mirrored random shapes for monsters and parts for items. Useful for testing, but generic next to everything above.
Sound effects take 3 seconds and music 3 to 12; I can measure them but not hear them, so your ears decide.
Effects are trimmed and levelled. Music is set to a game-friendly loudness and lasts exactly 16 bars. Whether the music loops without a bump is still untested.
| Effect | Stable Audio 3 Small | Stable Audio 3 Medium |
|---|---|---|
| Chest chomp | ||
| Burp | ||
| Gold drop | ||
| Legendary find | ||
| Level up | ||
| Sword hit | ||
| Critical hit | ||
| Slime pops | ||
| Dragon roar | ||
| Ice cracks | ||
| Fire whoosh | ||
| Button tap |
| Music | ACE-Step 1.5 | Stable Audio 3 Medium |
|---|---|---|
| Title theme | ||
| Crypt zone | ||
| Ice caves | ||
| Ember depths | ||
| Dragon's hoard | ||
| Boss fight |
No game has hit big because of AI art; the ones that did well used it quietly, under a look a human set.