Records · kv:2610.00044v1

Reproduction of cloneofsimo/minRF: rectified-flow DiT on MNIST

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-24

Independent reproduction · Benchmark or leaderboard entry · Machine learning

Reproduction of: original code

Abstract

Ran rf.py at the pinned commit, unchanged, in its default MNIST setting: a class-conditional rectified-flow model with a small DiT (Llama-style, dim 64, 6 layers, 0.61M parameters), Adam at learning rate 5e-4, batch size 256, 100 epochs, sampling 16 images after each epoch. The reported values are the per-epoch training losses rf.py prints for ten bins of the flow time t. wandb was disabled and the sample GIFs are withheld as hash-only entries (the last-frame PNGs are kept). The repository's CIFAR-10 setting (--cifar, 13.3M parameters) was attempted but ran out of memory on its first step at its fixed batch size of 256 on this 16 GB GPU, so it is not part of this record. PyTorch 2.11 (CUDA 12.8). One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000044v1,
  title={Reproduction of cloneofsimo/minRF: rectified-flow DiT on MNIST},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00044v1},
  url={https://kveritas.org/records/2610.00044v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.