A new compression technique cuts the memory AI video generators need by up to 6.2x without the flicker that cheaper compression methods leave behind.
Researchers describe the method, called QuantWM, in a paper on 2-bit KV cache quantization for video generation and world models. They found that existing 2-bit quantization methods look nearly lossless on benchmarks like VBench, yet still cause visible flickering and degraded video quality. Digging into why, they discovered that compressing the attention Key data, even though it produces smaller numerical errors than compressing Value data, does far more damage to output quality because it shifts the attention calculations that determine which frames and tokens the model focuses on. QuantWM fixes this with two techniques: one that picks compression-friendly value ranges for the channels most critical to attention (QSAC), and another that corrects leftover errors using low-rank projections aligned with the query data (PSAC). Tested across five video and world models, including Causal-Forcing, HY-World 1.5, and Longcat-Video, it beat prior 2-bit methods on both memory savings and visual quality.
This matters because memory, not raw compute, is what caps how long or detailed generated video can get. Every frame a world model generates requires caching attention data for all the frames before it, and that cache is what fills up GPU memory first. The paper is also a useful reminder that benchmark scores like VBench can hide real degradation. A method can pass the test and still flicker on screen.
Whether QSAC and PSAC hold up once developers push past benchmark clips toward the minute-long, high-resolution generations everyone actually wants is the question this paper doesn't yet answer.