FMT Targets Fixed-Step Parallel Lossless Image Decoding
Ammar Ali’s paper introduces FMT, a learned lossless image format that combines a coarse-to-fine tree with parallel rANS decoding so sequential decoding steps remain fixed as image size grows. The paper reports that its network-free codec beat NVIDIA’s HTJ2K on size and speed, while a large-network version decoded Kodak photos 9 to 61 times faster than tested learned codecs on the same A100 at equal or smaller size. These performance and size claims come from the paper and have not been independently corroborated in the supplied report.
FMT addresses the decoding latency that has limited learned image codecs, potentially making compressed data more practical for AI training and archival pipelines if its reported results hold outside the authors’ tests.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.23108 (cs)
[Submitted on 17 Nov 2026]
Title:FMT: Lossless Image Compression in a Fixed Number of Parallel Steps
Authors:Ammar Ali
View a PDF of the paper titled FMT: Lossless Image Compression in a Fixed Number of Parallel Steps, by Ammar Ali