Hugging Face's logo
Roberto Tacconelli's picture
5 2 7

Roberto Tacconelli PRO

robtacconelli
branikita's profile picture OzTianlu's profile picture tylerwsl's profile picture
Ā·

AI & ML interests

None yet

Recent Activity

reacted to their post with šŸ¤Æ about 3 hours ago
šŸ§¬ Midicoth: diffusion-based lossless compression ā€” no neural net, no GPU, no training data What if reverse diffusion could compress text ā€” without a neural network? Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree ā€” 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core. Beats every dictionary compressor we tested: enwik8 (100 MB) ā†’ 1.753 bpb (āˆ’11.9% vs xz, āˆ’15% vs Brotli, āˆ’24.5% vs bzip2) alice29.txt ā†’ 2.119 bpb (āˆ’16.9% vs xz) Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics ā€” no mixer, no gradient descent, just counting. The Tweedie denoising layer adds 2.3ā€“2.7% on every file tested ā€” the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based. No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput. šŸ’» Code: https://github.com/robtacconelli/midicoth šŸ“„ Paper: https://huggingface.co/papers/2603.08771 ā­ Space: https://huggingface.co/spaces/robtacconelli/midicoth If you ever wondered whether diffusion ideas belong in data compression ā€” here's proof they do. ā­ appreciated!
reacted to their post with šŸš€ about 4 hours ago
šŸ§¬ Midicoth: diffusion-based lossless compression ā€” no neural net, no GPU, no training data What if reverse diffusion could compress text ā€” without a neural network? Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree ā€” 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core. Beats every dictionary compressor we tested: enwik8 (100 MB) ā†’ 1.753 bpb (āˆ’11.9% vs xz, āˆ’15% vs Brotli, āˆ’24.5% vs bzip2) alice29.txt ā†’ 2.119 bpb (āˆ’16.9% vs xz) Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics ā€” no mixer, no gradient descent, just counting. The Tweedie denoising layer adds 2.3ā€“2.7% on every file tested ā€” the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based. No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput. šŸ’» Code: https://github.com/robtacconelli/midicoth šŸ“„ Paper: https://huggingface.co/papers/2603.08771 ā­ Space: https://huggingface.co/spaces/robtacconelli/midicoth If you ever wondered whether diffusion ideas belong in data compression ā€” here's proof they do. ā­ appreciated!
posted an update about 4 hours ago
šŸ§¬ Midicoth: diffusion-based lossless compression ā€” no neural net, no GPU, no training data What if reverse diffusion could compress text ā€” without a neural network? Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree ā€” 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core. Beats every dictionary compressor we tested: enwik8 (100 MB) ā†’ 1.753 bpb (āˆ’11.9% vs xz, āˆ’15% vs Brotli, āˆ’24.5% vs bzip2) alice29.txt ā†’ 2.119 bpb (āˆ’16.9% vs xz) Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics ā€” no mixer, no gradient descent, just counting. The Tweedie denoising layer adds 2.3ā€“2.7% on every file tested ā€” the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based. No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput. šŸ’» Code: https://github.com/robtacconelli/midicoth šŸ“„ Paper: https://huggingface.co/papers/2603.08771 ā­ Space: https://huggingface.co/spaces/robtacconelli/midicoth If you ever wondered whether diffusion ideas belong in data compression ā€” here's proof they do. ā­ appreciated!
View all activity

Organizations

None yet