FMT: Lossless Image Compression in a Fixed Number of Parallel Steps
FMT reports decoding Kodak photos 9–61 times faster than tested learned codecs, at the same file size or smaller.
The comparison used the same NVIDIA A100 GPU. FMT arranges image data in a coarse-to-fine tree, letting a GPU predict independent values together and decode them through parallel entropy-coding lanes. The number of sequential decoding steps stays fixed as image size grows. A version with no neural network also beat NVIDIA’s HTJ2K on both compressed size and speed in the reported tests.
HF Daily Papers' note
The comparison used the same NVIDIA A100 GPU. FMT arranges image data in a coarse-to-fine tree, letting a GPU predict independent values together and decode them through parallel entropy-coding lanes. The number of sequential decoding steps stays fixed as image size grows. A version with no neural network also beat NVIDIA’s HTJ2K on both compressed size and speed in the reported tests.
HF Daily Papers' note
score 5