Nvidia Research has published Parallel Decoding Distillation (PDD), a technique that lets a video or image diffusion model skip most of its own generation steps without being retrained from zero. Step count is the main cost driver in video generation, so a large cut there changes what a video product can afford to ship. On Wan2.1 14B, a text-to-video model, PDD needs four network evaluations to produce a clip where the unmodified model needed 100. On Nvidia’s own LTX-2.3 model, PDD needs eight where the unmodified model needed 120.
Diffusion and flow-matching generators build a clip by starting from random noise and refining it toward a finished frame across many small steps, and each step is a full pass through the network. A normal Wan2.1 run chains 50 of those steps, and because it also uses classifier guidance, each step costs two passes, not one, for 100 total. LTX-2.3’s ordinary run chains 30 steps at four passes each for guidance, for 120 total. Every one of those passes has to finish before the clip is ready, which is why longer, higher-quality video generation gets slow and costly fast.
PDD collapses that chain by training a second network to predict several steps ahead in one pass instead of one step at a time. Nvidia’s researchers split the sampling trajectory into a fixed run of intervals and group consecutive intervals into blocks. A student model, built on the teacher’s own architecture and seeded with the teacher’s starting weights, learns to predict where the trajectory lands after an entire block, matching what the teacher’s own step-by-step solver would have produced. Once trained, one network call advances the whole block at once, so an eight-step generation needs exactly eight evaluations rather than the dozens a standard solver would run. Because the model keeps a distinct output layer for every individual interval, and those layers combine into a single layer before running, the same trained model can serve different step counts, four or eight in Nvidia’s demos, without separate training runs for each.
The diversity claim is the part worth scrutinizing, since squeezing many steps into a handful typically costs a model its variety. Nvidia’s own framing blames two families of shortcut used elsewhere, score-based distillation and adversarial losses, for that exact failure: they tend to converge on a narrow set of outputs, a problem generally known as mode collapse. PDD instead trains against points sampled from its own predicted path, matched to the teacher’s real solver output at that point, which Nvidia positions as a steadier target than an adversarial signal. As evidence, the project page runs identical prompts through PDD and through DMD2, a rival few-step method built on that adversarial approach, using three different random noise seeds side by side for each prompt. The DMD2 clips look closer to one another across seeds; the paired PDD clips vary more in composition and motion. Nvidia does not attach a numeric diversity score to that comparison, so the case rests on the video grid rather than a benchmarked statistic.
Nvidia also ran PDD against AnyFlow, another few-step method, at both four and eight evaluations on Wan2.1, and against its own already-shipped distilled LTX-2.3 model at eight evaluations, using the same upscaler and refiner across every method in each comparison. Nvidia states that PDD reaches state-of-the-art results at four to eight evaluations across every generator the team tested: LTX-2.3 for video and audio, the 14-billion-parameter Wan2.1 for video, and Qwen-Image, Nvidia’s text-to-image system.
For anyone shipping a video-generation feature, cost per generated second is the number that decides whether the feature is viable, and that cost tracks network evaluations per clip almost directly. Taking a 100-to-120-call generation down to four or eight evaluations is a twelve to thirty times cut in the compute a provider burns before a customer sees a finished video. If PDD’s quality and diversity hold up on prompts outside Nvidia’s own demo set, it moves AI video generation closer to a cost structure that supports free or high-volume tiers rather than a metered, premium-only feature.
This account is drawn from NVIDIA Research’s Parallel Decoding Distillation project page, which does not list a publication date.
