dKV Forcing Separates Generation and Memory in Video Generators via Differentiable KV Cache

The authors of dKV Forcing: Separating Generation and Memory Representations through a Differentiable KV Cache (arXiv, November 1, 2026) propose a method for causal video generators that learns a separate memory representation for the KV cache. In a pretrained generator, the KV state used when denoising a segment already differs noticeably from what is written to the cache for future segments. The authors introduce a writer-specific low-rank residual on top of the shared backbone: writing to memory is separated from generation and is trained through future reads of what was written. To do this, the training cache is replaced with persistent differentiable pages with local rematerialization, which allows computing in a single pass the gradient that two-pass reconstruction computes in two. The full memory gradient costs 7% overhead relative to the generator update, while resident memory is 28% smaller and training cycles are 20.8% shorter than with two-pass reconstruction. Compared with Self Gradient Forcing (SGF), writer adaptation raises VBench-Long aesthetic quality by 6.6 points at 60 seconds and by 8.5 points at 240 seconds, with the highest subject and background consistency. The authors have not released the code.

What it means

The work concerns long videos and frame consistency — the topic around which our video generators section is built. The reference currently lists three open-source tools, and none of them mentions a mechanism similar to dKV Forcing. Maestro by Blizaine is built on the WanGP pipeline and uses the MiniMax H3 model. The code is distributed under the WanGP Non-Commercial Evaluation License 1.1 — non-commercial use of the program itself only, although the generated results can be sold with attribution. The MiniMax H3 weights are available under the MiniMax H3 Community License with a commercial revenue threshold of 20 million dollars, but the license does not apply in the EU, the UK, Korea and the US. As described by the author, it needs 6 to 24 GB of VRAM, and the largest weight file is 9.7 GB. The fact sheet says nothing about a differentiable cache or separating writing from generation. Imagination-Studio-by-keys-Wasserman is an AI film studio for Mac and NVIDIA DGX Spark on TensorFold, using MiniMax H3 and FastH3 for video. The code is under Apache-2.0, the drama-3 model weights are under the Fish Audio Community License with a revenue threshold of 500 thousand dollars and separate hosting licensing. It needs 8.3 GB of VRAM and 128 GB of unified memory as system RAM. The fact sheet has no mechanisms for working with the KV cache. LTX-Video by Lightricks is an image-to-video generator, with code under Apache-2.0 and weights under the LTXV Open Weights License with a threshold of 10 million dollars for free commercial use. It claims just 1 GB of VRAM, and the largest weight file is 26.6 GB. The description says nothing about the KV cache or long-term memory. For now, dKV Forcing remains a research paper without published code, so the method cannot be tested on the tools in the reference. If the authors release an implementation, it could be compared with how Maestro, Imagination Studio and LTX-Video maintain consistency on long clips.