ViRDM Removes Teacher and Critic from Few-Step Video Distillation

The authors of ViRDM propose a post-training method for a video generator without a teacher and a critic, reducing VRAM usage. The work was published on arXiv on September 24, 2026. Instead of three networks — teacher, critic and generator — only the generator remains, which is fitted to a precomputed target distribution of representations. The authors did not release any code.

What it means

This news note concerns the field of few-step causal video generation, which is not yet represented by a separate tool in our video generators section. The closest tool pages on the topic — LTX-Video, CogVideoX and Maestro — describe ordinary diffusion models, not few-step autoregressive generators with post-training via distribution matching. LTX-Video has open source code under Apache-2.0, its model weights require a separate license for companies with revenue from 10 million dollars, and as described by the author the model runs with just 1 GB of VRAM. CogVideoX uses Apache-2.0 for the code and its own model weights license with registration for commercial use; the fact sheet lists the CUDA platform. Maestro is a local studio based on WanGP with a non-commercial license for the code and MiniMax H3 model weights, which need at least 12 GB of VRAM. ViRDM is a research work with no published code and no specified license. Until the method turns into an available tool, there is nothing to compare it with the reference tool pages in terms of hardware requirements or terms of use.