Code: https://github.com/kulikov0/micro-doom-world Weights: https://huggingface.co/kulikov0/micro-doom-world
A 39M-parameter diffusion transformer predicts the next 128x96 frame from the previous 31 frames and the keys you press. It runs in the browser on WebGPU (it was tested on chrome and safari, I'm not sure how it runs in other browsers). Trained it on 1.5M frames of a scripted bot playing freedoom E1M1-E1M3.
Note: it won't run well on every device. On my M4 Max it's smooth, but on a Pixel I get only 1-2 fps because of the hardware. Without WebGPU it falls back to CPU and it's very slow
Comments URL: https://news.ycombinator.com/item?id=50533073
Points: 1
# Comments: 0

