H-DDCM: Hadamard Codebooks for Fast Generative Image and Video Compression
H-DDCM uses Hadamard codebooks for faster, lower-memory generative image and video compression.
H-DDCM is a training-free generative compression framework that replaces dense Gaussian codebooks used by methods such as Turbo-DDCM with implicit randomized Hadamard dictionaries. Fast Walsh-Hadamard transforms synthesize full-dimensional dictionaries with O(d log d) projection and reconstruction cost and O(d) memory, plus a table-free lexicographic coder. Adapted to an autoregressive video flow model, a streaming scheme periodically re-encodes recent frames into a fresh KV-cache prefix at no extra bits, enabling long-form near real-time compression on one GPU with competitive rate-distortion-perception results.
- H-DDCM replaces dense Gaussian codebooks with implicit randomized Hadamard dictionaries.
- Walsh-Hadamard transforms need O(d log d) compute and O(d) memory.
- A streaming scheme re-encodes frames into a fresh KV-cache prefix without extra bits.
- It targets near real-time long-form generative video compression on one GPU.
From the original article
1 min read at huggingface.co
Training-free diffusion-based compression methods repurpose the generative priors of a diffusion or a flow model to compress data without retraining them as task-specific codecs.
By Or Ronai, Vladimir Kulikov, Tomer Michaeli · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.