FastGen-PDD: NVIDIA's 4-8 Step Distillation for Video and Image Models

Video diffusion models are slow because sampling is iterative — hundreds of small denoising steps, hundreds of network evaluations, before a single clip exists. The current standard fix, variational score distillation (VSD) with adversarial losses, works but is notoriously unstable to train and prone to mode collapse: fewer distinct outputs, less motion, samples that start to look alike.

NVIDIA and the Weizmann Institute of Science published FastGen-PDD: Parallel Decoding Distillation for Image and Video Generation, a trajectory-based distillation method that sidesteps VSD and adversarial training entirely. Authors Neta Shaul, Chao Liu, Arash Vahdat, and Julius Berner report state-of-the-art results at 4-8 network evaluations (NFE) on LTX-2.3, Wan2.1 14B, and Qwen-Image — with meaningfully better output diversity than adversarial baselines like DMD2 and AnyFlow.

This lands in the same fast-generation research thread explainx.ai has tracked with Mercury 2's 1,000+ tok/sec diffusion LLM and Google's DiffusionGemma — trading network evaluations for wall-clock speed without retraining from scratch.

TL;DR

The Problem With Current Few-Step Distillation

Sampling a diffusion or flow-matching model with a classical ODE solver takes many small steps — hundreds of network evaluations for a single generation. Trajectory-based distillation compresses this by training a student model to predict a large-step update that merges several small teacher steps into one.

The dominant approach for doing this well is variational score distillation (VSD), usually paired with an adversarial loss. It produces high-quality few-step video, but the paper's authors call the training "notoriously hard to optimize" — it suffers from mode collapse, meaning the distilled model converges toward a narrower set of outputs, sacrificing video diversity and motion range compared to the full teacher.

That instability is the gap PDD targets.

The PDD Idea: Predict, Don't Merge

Instead of merging multiple small teacher updates into one big student update, Parallel Decoding Distillation predicts multiple teacher updates from a single network evaluation.

Time is discretized into a fixed sequence of $N$ intervals, grouped into blocks of size $L$. During training, the student learns to predict the teacher's ODE-solver step for every interval inside a block in one forward pass. At generation, each student evaluation then advances $L$ intervals at once instead of one.

Conceptually, PDD learns a representation of the mean velocity without regressing its derivative through JVPs or finite-difference approximations — the technique some competing consistency-style methods use. It decomposes the mean-velocity prediction over a full interval into parallel sub-interval predictions, then optimizes over a single sub-interval per training iteration to supply the learning signal for the full-interval mean velocity.

Architecture: One Backbone, Replicated Output Heads

PDD's student architecture starts from the pretrained teacher backbone — same weights, same structure — and replicates the final linear layer once per interval in the discretization grid, giving the student a distinct output head per interval $t_i$.

Predicting all $N$ grid-level outputs (not just the final block-level output) is what enables variable NFE at inference time: the model trains with variable block sizes, then generation adjusts the block size to whatever NFE budget you want.

The efficiency trick: since a weighted sum of linear transformations is itself linear, the replicated per-interval layers can be fused into a single linear layer per block at generation time. Inference only needs to evaluate and store one fused layer per block, not $L$ separate ones. This architecture also means student weights initialize directly from the teacher — no cold start.

Training: On-Policy Trajectory Matching

PDD trains with an on-policy algorithm that aligns each student output with the corresponding teacher ODE-solver step:

1. Sample a noisy state $x_t$ at some timestep.
2. Roll the student forward across the next block of intervals, starting from $x_t$.
3. Sample a single state $x_$ from that student rollout.
4. Match the student's $i$-th output against the teacher's ODE-solver step evaluated at $x_$.

No discriminator, no adversarial min-max game — just a direct regression target computed from the teacher at a state the student itself produced. The paper includes a toy-problem simulation showing the student trajectory converging onto the teacher's "optimal trajectory" as training progresses, with the largest-error batch sample plotted at each iteration.

Results: SOTA at 4-8 NFE

NVIDIA benchmarked PDD against teacher models and existing few-step baselines across three model families:

On Wan2.1 14B, side-by-side comparisons across multiple noise seeds for the same prompt show PDD holding motion and subject variation that DMD2 tends to flatten. Against AnyFlow at both 4 and 8 NFE, the same pattern holds — PDD's headline claim is improved video diversity, not just matched quality, at equal step budgets.

On LTX-2.3, PDD at 8 NFE (with the official upscaler and refiner) is compared directly against the 4×30 NFE teacher and the officially distilled LTX-2.3 model, spanning prompts with dialogue, ambient sound, and camera motion — since LTX-2.3 generates audio alongside video.

Why This Matters for Production Video Generation

Every NFE is a full forward pass through a large video diffusion transformer — often billions of parameters, run at every one of dozens to hundreds of steps, sometimes multiple times per step for classifier-free guidance. Going from 50+ NFE with 2-4x guidance overhead down to 4-8 NFE is roughly a 10-25x reduction in network evaluations for the same generation.

That kind of reduction is what separates a video model that runs as an interactive product from one that only works as a batch job — the same tradeoff explainx.ai has covered across Runway's real-time video agents and Google's Gemini Omni video model. Distillation quality at low NFE is the bottleneck standing between "video model" and "interactive video product," and mode collapse has historically been the tax teams paid to get there with adversarial methods.

For teams building on Qwen-Image or Wan-family video backbones specifically, PDD's approach — compatible with any pretrained diffusion or flow model, without requiring adversarial retraining — is a lower-risk path to a fast, deployable variant than building a new GAN-based distillation pipeline in-house.

Limitations to Watch

• Code is not yet released — the project page lists code as "coming soon" at time of publication; there's no public repo to verify claims independently yet.
• Diversity gains are relative, not absolute — PDD improves diversity versus DMD2/AnyFlow at the same NFE; it isn't claimed to fully match the full multi-hundred-step teacher trajectory's diversity.
• NFE parity isn't wall-clock parity — 8 NFE with an upscaler and refiner (LTX-2.3) still involves extra passes beyond the core denoiser; compare end-to-end latency, not just NFE count, when evaluating for production.
• Benchmarked on specific backbones — LTX-2.3, Wan2.1 14B, and Qwen-Image are the tested models; generalization to other architectures (e.g., different guidance schemes or 3D VAEs) isn't demonstrated in this paper.

Related Reading

• MiniMax-H3: Open Video Model Ranked #1 on Video Editing With Audio
• Mercury 2: 1,009 Tokens/Sec Diffusion LLM and What It Changes for Agent Loops
• Google DiffusionGemma: 4× Faster Text Generation
• How Diffusion Models Work: Complete Guide to AI Image Generation
• AI Video Generation in 2026: Complete Guide to Sora, Runway, Kling
• Qwen-Image-3.0: Dense Layouts and 10px Text
• NVIDIA Nemotron-Labs-TwoTower: 2.42× Faster Diffusion Generation
• ViMax: Agentic Video Generation Complete Guide

Sources: FastGen-PDD project page and paper, NVIDIA Research and Weizmann Institute of Science, authors Neta Shaul, Chao Liu, Arash Vahdat, and Julius Berner. Code listed as "coming soon" as of publication. Stats and comparisons accurate as of July 2026.