Doomformer Brings Integer Math to DOOM

Developer mottopanikeiku released the open-source project Doomformer on GitHub on October 20, 2026. It performs integer calculations using mechanics from the shooter DOOM. In one demo, the player draws a line on the floor while holding the fire button, sketches a “7,” then presses a read button to light up the result: “7.”

Doomformer converts a neural network that performs integer calculations into a level for ViZDoom, a DOOM-based game environment. Inside the level, player duplicates ride a conveyor belt and add the value assigned to each line they cross. The calculation result appears through the brightness of lights in different sections and the height of the floor. The developer also released a demo that uses a small Transformer to generate DOOM level names.

In the digit-recognition demo DrawPad, 1,355 player duplicates calculate inside a level containing drawn digits. When a duplicate corresponding to the recognized result crosses a line, that digit lights up. The developer evaluated the demo on 10,000 test images from the handwritten-digit dataset MNIST and correctly classified 9,716. The calculations were divided across four maps, with Python passing values between them.

For each image, the game returned 248 values. Across all 10,000 images, those values exactly matched the results of integer calculations using NumPy. An update dated November 2, 2026, labeled v0.7.2, added an option to export DrawPad as a map file. The exported map can run in the DOOM-based environment GZDoom without Python. The reported evaluation was conducted by the developer using 10,000 MNIST test images, a four-map setup, and 248 values per image.

Developer mottopanikeiku released the open-source project Doomformer on GitHub on October 20, 2026. It performs integer calculations using mechanics from the shooter DOOM. In one demo, the player draws a line on the floor while holding the fire button, sketches a “7,” then presses a read button to light up the result: “7.”

Doomformer converts a neural network that performs integer calculations into a level for ViZDoom, a DOOM-based game environment. Inside the level, player duplicates ride a conveyor belt and add the value assigned to each line they cross. The calculation result appears through the brightness of lights in different sections and the height of the floor. The developer also released a demo that uses a small Transformer to generate DOOM level names.

In the digit-recognition demo DrawPad, 1,355 player duplicates calculate inside a level containing drawn digits. When a duplicate corresponding to the recognized result crosses a line, that digit lights up. The developer evaluated the demo on 10,000 test images from the handwritten-digit dataset MNIST and correctly classified 9,716. The calculations were divided across four maps, with Python passing values between them.

For each image, the game returned 248 values. Across all 10,000 images, those values exactly matched the results of integer calculations using NumPy. An update dated November 2, 2026, labeled v0.7.2, added an option to export DrawPad as a map file. The exported map can run in the DOOM-based environment GZDoom without Python. The reported evaluation was conducted by the developer using 10,000 MNIST test images, a four-map setup, and 248 values per image.