Picture in, game out
From a toy robot drawing to a 3D N64 game: Bolt Bot Rush
We gave img2n64 one drawing of a toy robot and two sentences, and got back Bolt Bot Rush: a 3D arena game on our N64 engine where that robot, as a solid 3D figure with the drawing wrapped onto it, collects bolts while boulders roll after it. The agent built it, played it and checked it in 2 minutes 39 seconds, for 10 gems of a new account's 30. The figure itself cost nothing.
Also in: Español · Deutsch · Français · Português · Italiano · Nederlands · Svenska · Norsk · Bahasa Indonesia · Türkçe · Polski · Tiếng Việt · Русский · 日本語 · 한국어 · 繁體中文

On this page
Here is everything that went in, what came out, and what the AI did in between, from the session's own record.
The picture
A simple drawing of a toy robot: orange head, round cyan eyes, an antenna, a blue body with a yellow button, standing on a white background. Nothing special was done to it. One character, seen from the front, with a plain border, which is the kind of picture that gives the cleanest figure.

The prompt
On the start card we chose The player and Solid 3D figure (free), and typed:

Because the border of the drawing was plain, the page cut the robot out itself and sent the agent a transparent picture (96 x 133 pixels), and it added the site's role line: import the picture, make it a solid 3D figure, put it in the cartridge as the hero's mesh, light it, give it a shadow, and do not redraw it or show it as a flat cut-out.
What the AI did, step by step
This was a new account, started from the site, on a local copy of our stack (the same app and agent as img2n64.com). The session took 26 agent steps. In order:
- Imported the picture free as an image asset and named the game Bolt Bot Rush.
- Made the 3D figure:
create_assetwith the freepicture-3dmodel. Back came a closed, rounded body 628 triangles strong, with a 64 x 64 texture in 16 colors that takes 2,080 bytes of the console's 4 KB texture memory. 0 gems. - Put it in the cartridge as
MESH_hero; its texture came along asTEX_hero. It also added two of the built-in sounds, a blip and a sting. - Read the engine's reference for the N64 engine's API and its sample game, then wrote the whole game in one go in
main.sg64: a title screen with an auto-playing demo, the arena, five bolts, three boulders, lives, a score and a best score saved to the cartridge's save chip. - Built and looked. The cartridge compiled and booted, but the agent noticed the robot drew zero triangles. It searched the docs, found that a model is scaled so its widest half spans 256 object units, and changed the hero's scale and height. The next build drew the robot.
- Tuned the size once more so the robot read well against the arena, and played: START, then the d-pad and stick, with screenshots between.
- Was asked to prove it. Its first report said it had not played a round to the end, so the app sent it back to check. It played until the boulders took all three lives and the game-over screen came up, confirmed START restarts a round, and, since its own steering never reached a bolt, let the title demo (which runs the real game code) play 600 frames: the score read 6 bolts.
- Reported honestly what was still unconfirmed: the best score saving to the cartridge, because restarting the cartridge during the check had cleared the save.
Its closing message, word for word: "Press START to begin, then steer with the control stick to collect bolts while dodging the boulders." Below it, the chat showed "Worked for 2m 39s · 10.0 gems" and offered a next idea: "Make the boulders bigger and add a speed-boost pickup".

The figure
Here is the figure the agent made, open in the studio's model view with its wireframe showing. It is the drawing, but with depth: the head and body are puffed into rounded shapes, and the edges curve away the way a plush toy's do.

In the game it is lit by the arena's sun, turns to face where it rolls, bobs while it moves and has a soft shadow under it.

What we would ask for next
The first build is a starting point, and the chat stays open. If this were our game, the next turns would be:
- "Make the boulders bigger and add a speed-boost pickup" (the agent's own suggestion).
- "Bring the camera lower and closer behind the robot", so the figure fills more of the 320 x 240 screen.
- "Save the best score and show it on the title screen", then ask the agent to prove it by playing two rounds.
Each of these is a normal agent turn. Code edits you make yourself in the Code view cost nothing.
How to make your own
- Open img2n64.com and drop your picture, or use Generate to make one from words, or Transform to change one you have.
- Pick what it becomes (the player, an enemy, the prize, or the sky and colors) and keep Solid 3D figure, which is free.
- Write one or two sentences about the game and press Make my N64 game.
The full walkthrough is in how to turn a picture into an N64 game.
FAQ
Was this a real session or a mock-up?
Real. Every screenshot here is from the session, and the steps above come from the agent's own transcript. It ran on a local copy of our stack with a new account, the same app and agent that img2n64.com uses; only the account was a test account.
Why did the robot first draw as nothing?
The figure arrives scaled to the console's object units, and the agent's first guess at its size was far too large for the camera. It saw the problem in its own screenshot, looked up the scaling rule in the docs and fixed it on the next build. This kind of check is why the agent plays its own builds.
Why does the agent say some things are unconfirmed?
The app asks the agent to prove what it claims by playing, and to say what it could not check. In this session that was the high-score save, and it said so.
Could the robot be the enemy instead?
Yes. Choose An enemy before you start, and describe who the player is. The agent then puts your figure in the arena as the thing that turns toward you and chases you.
Can I play Bolt Bot Rush?
It was made on a test account and is not published. Making your own takes the same three steps above and about the same few minutes.
Further reading
Last updated: November 17, 2026


