Long video generation is bottlenecked not by model weights but by the KV cache. Video diffusion transformers reuse cached keys and values across denoising steps, and the cache grows linearly with generation history; the paper, posted 19 Nov 2026, says it can reach tens of gigabytes, more than commodity GPUs can hold (https://arxiv.org/abs/2611.26334). An independent team behind QVG gives a sharper number: autoregressive video KV caches often exceed 30 GB, blocking deployment (https://arxiv.org/abs/2602.02958). A three-author group from MIT's Operations Research Center now pushes that cache to 2 bits while lifting subject consistency from 90.38 to 92.02.
Why LLM recipes do not transfer
KV-cache quantization is mature on the LLM side. The QuaRot line of work rotates first and quantizes second, mixing outlier channels into benign ones so the distribution flattens before bits are cut (https://arxiv.org/abs/2404.00456). The paper argues those methods were built for the structure of LLM caches and lose substantial accuracy when applied directly to video diffusion transformers. Video caches have their own structure: the cached vectors of one image patch across adjacent frames share a slowly varying per-channel mean. The paper quantifies it - with 2-frame blocks, 91.22% of the squared norm in Key caches sits in the block mean; even at 16-frame blocks the mean still holds 76.88%. Quantize that to 2 bits naively and the mean destroys everything.
Two-bit angles, fitted online per video
TRQ gives each temporal block (default 8 frames) a tree of 2x2 planar rotations. Each angle is chosen from just four values - 0, pi/4, pi/2, 3pi/4 - so it costs 2 bits of metadata; a fixed Hadamard transform shared by all blocks sits behind the tree. The goal is mean flattening: rotate the block's per-channel mean into a near-constant vector, so the quantizer offset absorbs the shared component. The fit is greedy, and the paper proves the greedy fit is globally optimal for this objective. Everything happens online for each new video, with no gradients and no weight changes. A block of 8 frames and 128 channels needs only 2048 bits of rotation metadata.
What the numbers say
The comparison spans LongCat-Video-13B, Self-Forcing-Wan-1.3B and HY-WorldPlay-8B, with the main arena on LongCat: 100 matched videos at 480p. At matched compression (TRQ 6.89x, QVG 6.88x), subject consistency reads 93.02 for the BF16 baseline, 90.38 for QVG, 92.02 for TRQ; PSNR rises from 24.06 to 26.53. On 25 matched 720p videos the gap widens: 93.46 versus 91.06. On the 8B HY-WorldPlay the paper's abstract reports 86.33 up to 87.61, and the paper claims leads on all seven evaluated metrics for LongCat and Self-Forcing, six of seven for HY-WorldPlay. The ablation tells the story: Hadamard alone leaves PSNR at 25.90, random angles sink to 23.30, versus 27.69 for the full method - the online-fitted angles are where the gain lives. On an H100, fused Triton kernels cut the per-pair rotation cost from 6.638 ms to 1.785 ms, and whole-cache processing from 10.196 s to 2.743 s.
The caveats
First, the paper itself flags weak temporal redundancy - fast motion, scene cuts - as an open problem, with adaptive block lengths listed as future work. Second, TRQ (rotations) and QVG (centroids) exploit complementary structure; the paper says combining them could beat either alone, but does not do it. Third, the paper says fused Triton kernels are released, yet no repository link appears in the text. Fourth, every comparison number is self-reported, and QVG's standing is likewise the paper's framing, not an independent replication. The mechanism layer is on firmer ground: QuaRot showed rotations remove outliers without changing the output (LLaMa2-70B at 4 bits loses at most 0.47 WikiText-2 perplexity and keeps 99% of zero-shot performance), and QVG independently confirmed the 30 GB wall.
So the next phase of long video generation may be decided not by whose model is bigger but by whose cache is cheaper. Per-video, two-bit rotations are now a working path; fast motion and scene cuts are the holes to fill.