strong-followup
Parallel Decoding Distillation for Fast Image and Video Generation
Neta Shaul, Arash Vahdat · NVIDIA · 2026-07 · arXiv:2607.26004
Why it matters
Trajectory-based few-step distillation that avoids the Jacobian-vector-product and finite-difference machinery MeanFlow-style objectives rely on, reporting state of the art at four to eight function evaluations.
What this paper does
read: full textBefore this
State-of-the-art few-step distillation for video diffusion and flow models relied on variational score distillation and adversarial losses. These distribution-based losses are notoriously hard to optimize and prone to mode collapse, which shows up as reduced sample diversity and a lack of motion in generated video. Trajectory-based alternatives in the MeanFlow family avoided that instability but required Jacobian-vector products or finite-difference approximations to regress the derivative of the mean velocity, machinery that is expensive or unstable to scale to large video models.
The problem
No trajectory-based distillation method had successfully scaled to few-step, high-resolution video generation while both avoiding the JVP or finite-difference machinery of MeanFlow-style training and preserving the sample diversity that distribution-based methods sacrifice.
The idea
Instead of distilling a single large multi-step-to-one-step mapping, train a parallel decoder that predicts multiple denoising steps within a fixed time block from a single network evaluation. The decoder learns a representation of mean velocity across each interval in the block directly by regression against teacher-estimated velocities, without differentiating through the mean velocity via JVPs or finite differences.
How it works
PDD operates in the native latent space of each pretrained teacher model (video, audio, or image VAE latents) and keeps the teacher flow model frozen throughout. It trains a shared backbone plus N parallel linear output heads, one per discretized time step within a block, on top of the pretrained architecture. During training the student generates intermediate states via its own parallelized process, the frozen teacher estimates ground-truth mean velocities at those same states using a Runge-Kutta solver (Euler or midpoint), and a stop-gradient is applied to the teacher's estimate so gradients do not flow through it. The student's parallel decoding loss then regresses its own predicted mean velocities onto these teacher-estimated targets, so what is actually trained is a direct regression onto mean velocity rather than a derivative of it, and what is predicted at each step is that mean velocity over an interval rather than instantaneous noise, clean data, or a single-point velocity.
Evidence
On LTX-2.3 text-to-video/audio (22B parameters, 10-second 720p), PDD at 8 NFE matches or beats the official distilled model after only 250 training iterations, against a teacher that itself required 4x30 NFE. On Wan 14B text-to-video at 4 NFE, PDD reaches VBench overall 84.92 against the teacher's 83.90, with diversity metrics well above distribution-based baselines DMD2 and AnyFlow, including a V-JEPA 2 cosine distance of 0.0791 versus DMD2's 0.0568. On Qwen-Image text-to-image (20B) at 8 NFE, PDD reports OneIG 0.541, DPG-Bench 88.46, and GenEval 0.86, the best across all three benchmarks among compared methods, with OneIG diversity of 0.181 against DMD2's 0.109 while staying competitive on human preference (HPSv2 31.56). On ImageNet-256 at 1 NFE with the midpoint solver, PDD reaches FID 2.69, competitive with Pi-Flow's 2.85 though behind FreeFlow's 1.45. Sampling at 1, 2, 4, or 8 NFE is supported from a single trained model without retraining.
Limitations
The authors concede that their large-scale text-to-image and video experiments rely on data-free training, and that investigating PDD in data-dependent settings beyond ImageNet-256 remains future work. They also note that motion in generated videos increases later in training, a convergence dynamic they flag as needing further investigation, and identify adaptive block-size selection as an unexplored direction.
Why it matters
It is presented as the first trajectory-based distillation method to scale to few-step, high-resolution video generation while keeping the diversity that distribution-based methods like DMD2 lose to mode collapse, and it does so with a single regression objective that sidesteps the JVP and finite-difference machinery MeanFlow-style training needs, at 4-8 NFE across 14B-22B-parameter production-scale models.
Abstract, in the authors' own words
Generation in video diffusion or flow models is computationally expensive due to the slow and iterative sampling process. Current state-of-the-art (SOTA) acceleration methods heavily rely on variational score distillation (VSD) and adversarial losses to distill diffusion models into few-step generators. Albeit achieving high-quality video generation, these training losses are notoriously hard to optimize and suffer from mode collapse, leading to loss of video diversity and lack of motion. In this paper, we introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for fast inference of diffusion and flow matching models. Our architecture and training procedure are compatible with any pre-trained model and support sampling with a varying number of function evaluations (NFE). PDD accelerates generation by predicting multiple denoising steps per network evaluation. Conceptually, it learns a representation of the mean velocity without regressing its derivative using JVPs or finite-difference approximations. Our method achieves SOTA performance with 4-8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image. Moreover, PDD presents a significant improvement in generated video diversity.
Research line
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.