Video generation's efficiency week: 50% less compute, generative compression, and a model that skips denoising

Spend compute where it matters

In a preprint posted this week, ForcingBooster takes aim at the most obvious waste in autoregressive video diffusion: uniform computation. Existing methods typically allocate the same number of denoising steps to every patch of every video chunk, which means easy regions get as much refinement as hard ones. ForcingBooster instead assigns denoising steps adaptively by patch complexity — easy patches stop early, refinement concentrates on difficult regions — and reports cutting the total number of denoising steps across video patches by about 50%. It is compatible with most existing few-step autoregressive diffusion models and needs only a lightweight routing module. m:…2141

The design is a small lesson in where efficiency gains live now: not in bigger models but in smarter scheduling. If half the compute in video diffusion is spent re-refining regions that were already fine, the fix is a scheduler, not a shrink. As with all three papers here, these are the authors' own reported results in a preprint — promising, unreviewed, and worth treating as a direction rather than a done deal. m:…2141

Compress like you generate

In another preprint this week, H-DDCM attacks the memory side of the same problem. Generative compression methods like the recent Turbo-DDCM rely on dense Gaussian codebooks — big, explicit dictionaries that are expensive to store and search. H-DDCM swaps them for implicit randomized Hadamard dictionaries synthesized on the fly with the fast Walsh-Hadamard transform, claiming O(d log d) compute and O(d) memory efficiency plus a table-free lexicographic coding scheme. The payoff the authors emphasize is practical: a streaming compression scheme for video that can run in real time on a single GPU. m:…7325

Compression research rarely gets the spotlight, but it is the quiet enabler of everything downstream — cheaper storage, cheaper streaming, cheaper training data pipelines. A generative compressor that runs in real time on one GPU changes what a small team can afford to build on top of video. m:…7325

Skip the denoiser

The boldest of the three comes from Gen4P, also a preprint this week. Instead of making denoising cheaper, it asks whether you need it at all: the framework repurposes a pretrained video generator into a feed-forward 4D perceiver that recovers geometry and motion from monocular video without iterative denoising. On-demand cross-attention gives flexible spatial sampling, continuous-time queries give temporal flexibility, and a pretrained video-depth LoRA serves as a frozen geometry prior. The reported result is state-of-the-art performance on 3D point tracking, video depth, and camera pose estimation — from a model whose training was about generation, not perception. m:…2975

If the result holds up, it is a neat inversion of the field's assumptions: the best perceiver of 4D structure might be a generator that learned the world's dynamics by learning to fake them. Watch for whether this stays a clever trick or becomes the standard way to get geometry out of video models. m:…2975

Why it matters

Three preprints, one obsession: video is the most expensive modality to generate, and the field is learning to spend compute only where it pays. Adaptive denoising, generative compression, and denoiser-free perception are all moves toward the same goal — video AI that runs at interactive cost on hardware a startup can rent. The practical upshot for builders: the video generation stack is getting cheaper faster than it is getting better, which means the moat moves to data, evals, and product. Budget accordingly. m:…2141 m:…7325 m:…2975