H-DDCM: Hadamard Codebooks Make Diffusion-Based Image and Video Compression Scalable
Technion · hf · 2026-10-16
Technion researchers introduce H-DDCM, a scalable generative compression framework addressing bottlenecks of codebook-based training-free diffusion compression like Turbo-DDCM: dense Gaussian codebooks are costly in memory and computation and scale poorly with the large latent dimensions of modern image/video generators, while combinatorial bitstream encoding becomes computationally heavy at scale.
- Replaces Gaussian codebooks with implicit randomized Hadamard dictionaries, using the fast Walsh-Hadamard transform to synthesize full-dimensional dictionaries with O(d log d) compute and O(d) memory, without materializing the full dictionary.
- Introduces a table-free lexicographic coding scheme making combinatorial encoding practical at larger dimensions.
- Works for images and extends to video; project page at h-ddcm.github.io.
More from Multimodal
- Flux.3 klein vs Qwen Image 2.1 turbo on RTX 4070: 10s vs 3s per 1MP edit — Ant_6431 · 2026-10-16
- Kling 4.0 fully available, shows markedly better motion coherence in fast-action scenes — alifcoder · 2026-10-16
- MJev-Spatial: Single-Pass Classification Beats Autoregressive VLMs on Spatial Reasoning — mindlab-research · 2026-10-16
- ComfyUI v0.41.0 ships with Flux3 image model and MiniMax-H3 LightVAE support — Fcking_Chuck · 2026-10-16
- AI-generated short film 'The Dragon Rider' glides cinematically above the clouds — jijo6686 · 2026-10-16
- Realistic candid Midjourney portrait prompt: natural skin, lived-in bedroom light — tisch_eins · 2026-10-16