ViRDM: Controlling Representation Distribution Matching for Few-Step Causal Video Generation
ViRDM speeds causal video generation via a four-step post-training recipe: 84.87 VBench Total, 20 updates, single-GPU support, code, weights, eval.
Key Points
- ViRDM is an open few-step causal video post-training recipe that controls representation distribution matching by updating only the generator, avoiding the online diffusion teacher and learned critic used in DMD-based pipelines.
- The method adapts image RDM ideas to video, but requires video-specific changes to population size and initialization, because simple image defaults do not transfer well to causal video generation.
- Matching representation features alone is not sufficient for motion quality, so the recipe includes ablated lessons on temporal dynamics and practical settings for few-step video training.
- The reported four-step causal recipe achieves 84.87 VBench Total after only 20 generator updates, with about 2 hours of post-training on 8×A100 and support for a single 80GB A100.
- The release improves reproducibility by providing training and inference code, model weights, precomputed reference artifacts, evaluation scripts, and a single-GPU training launcher.

For teams trying to make causal video models generate faster without a full distillation stack, the neu-vi ViRDM release now provides a concrete open recipe. The shared announcement does not give a calendar date, but it describes a four-step causal video post-training path that reports 84.87 VBench Total after only 20 generator updates, with approximately 2 hours of post-training on 8×A100 and support for a single 80GB A100. It includes training and inference code, model weights, precomputed reference artifacts, reproduction configurations, evaluation scripts, and a single-80GB-GPU training launcher.
Meta description: ViRDM releases a few-step causal video post-training recipe: 84.87 VBench Total, 20 generator updates, 80GB A100 support, code, weights, and evaluation scripts.
Tags: ViRDM, causal video generation, video post-training, VBench, single-GPU training
Why few-step video post-training is harder than images
The recipe is inspired by FD Loss and RDM in image generation. Those results suggested a simpler path for video: match a precomputed representation distribution and update only the generator. That framing removes the online diffusion teacher and learned critic commonly used in DMD-based methods such as Self Forcing and Causal Forcing. For readers, the practical difference is that ViRDM does not require the same teacher-critic loop that many DMD-based post-training pipelines rely on.
The idea is straightforward on paper, but video creates harder engineering constraints. A multi-step causal video generation pipeline has a longer backward path than a one-step image pipeline. Backpropagating through the video VAE is especially memory-heavy, so out-of-memory failures had to be solved before the training recipe could be evaluated. This is why the release emphasizes single-80GB-GPU support: it addresses one of the most common barriers for researchers and smaller teams who cannot reserve large clusters for a new post-training experiment.
What had to change from image RDM
The team reports that several image RDM defaults did not transfer directly to few-step causal video. Population size and initialization had to be reconsidered. Image RDM can use very large generated populations and can start directly from a pretrained bidirectional model, but those choices could not simply be carried over to the video setting. The final recipe therefore treats population size and initialization as decisions that need video-specific ablation, not as fixed defaults.
The second major lesson was that feature matching alone did not close the motion dynamic gap. Even when video representation features were used, matching feature distributions was not enough to fully fix degraded dynamics. That detail matters for anyone adapting RDM-style training to video: a score that looks reasonable after a distribution-matching objective may still hide temporal or motion-quality issues. Through ablated experiments, the authors packaged the working configurations, practical lessons, and pitfalls into a recipe intended to reduce the trial-and-error cost for the community.
What the published results say
The final four-step causal recipe reports 84.87 VBench Total and only 20 generator updates. The paper evaluation also says the recipe achieves better results than the DMD-based baselines evaluated there, with substantially less post-training time and lower peak memory. The reported post-training cost is approximately 2 hours on 8×A100, while the single-80GB-A100 launcher makes the training setup feasible in a configuration that is challenging for common DMD pipelines.
These numbers are not a general benchmark claim across all video models or all hardware; they describe the paper's evaluated setting and the baselines compared there. The more immediate value is reproducibility: the release includes precomputed reference artifacts and evaluation scripts, so readers can inspect not just the headline score but also how the recipe was reproduced.
What readers can do next
The release gives readers four practical entry points. They can run training and inference code, load model weights and precomputed reference artifacts, reproduce the evaluation with provided scripts, or test the single-80GB-GPU launcher. The website and demos, GitHub repository, and paper are linked below:
- Website and demos: https://neu-vi.github.io/ViRDM/
- GitHub: https://github.com/neu-vi/ViRDM
- Paper: https://arxiv.org/abs/2609.28923
The broader impact is mainly in access and iteration speed. Few-step causal video post-training usually forces a choice between expensive distillation infrastructure and long, uncertain experiments. ViRDM does not remove those costs, but it shows that a narrower RDM-style approach can be made workable for video and can fit on a single 80GB A100. For teams building video models, the next question is whether the recipe transfers to their own base model, data, and motion-quality target. For teams without large compute pools, the single-GPU launcher may be the most useful part: it turns a paper claim into a local experiment that can be run, measured, and compared against their current pipeline.
The release also credits the foundational works it builds on, including FD Loss, RDM, Self Forcing, and Causal Forcing.

Comments (0)