Video post-training recipe drops the teacher and critic models
single source· 1 articles · confidence: medium · first seen 2026-09-23 20:00 UTC
What this means for you
Nothing to build on yet: no code, weights or release date appear, and the 84.87 carries no evaluation date. If you post-train few-step video generators, the number to check is the cost — 16 A100 GPU-hours, with no teacher or critic to host.
A paper listed on Hugging Face Daily Papers and posted to arXiv, ViRDM, describes post-training a few-step autoregressive video generator without the teacher-and-critic stack that distribution-matching distillation normally requires — the large pretrained model and the separate scoring network used to train the small one. It trains only the generator against a precomputed target distribution, using a lightweight VAE decoder (the component that turns compressed latents back into pixels) and staged vector–Jacobian products to keep memory tractable. After 20 generator updates it reports 84.87 on the official VBench evaluation, 0.36 above the previous best few-step causal baseline, at 16 A100 GPU-hours. No evaluation date is given and the work is not peer reviewed.
Key facts
- ·ViRDM is described as a teacher- and critic-free post-training recipe for few-step causal video generation, posted as arXiv 2609.28923. source
- ·It reports 84.87 on the official VBench evaluation after 20 generator updates. source
- ·The authors state that score is 0.36 above the previous best few-step causal baseline. source
- ·Stated training cost is 16 A100 GPU-hours. source
- ·Exploratory results are reported for one-, two- and four-step bidirectional generation and for lower causal sampling budgets. source
- ·No evaluation date, code release or peer-review status is given in the listing. source
What the sources say
- Hugging Face Daily Papers (research) — Abstract-only listing of a video post-training method that removes the teacher and critic.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation2026-09-23