Parallel Decoding Distillation for Fast Image and Video Generation
Neta Shaul, Chao Liu, Arash Vahdat, Julius Berner
cs.CV, cs.LG
2026-07-29
PDD adds a parallel decoder that predicts multiple denoising steps per forward pass and trains by pure regression, reaching SOTA at 4-8 NFE on Wan, Qwen-Image and LTX with far more video diversity than DMD2.
Generating with diffusion or flow-matching video models means running the denoising loop dozens to hundreds of times, with a full network forward pass each step (Network Function Evaluations, NFE). The standard fix is to distill the model into a few-step generator.
The bottleneck is the distillation loss. The strongest methods, led by DMD2, rely on variational score distillation (VSD) plus adversarial losses. These losses are hard to tune, memory-hungry, and prone to mode collapse: the generator plays it safe and collapses to a few samples, so the videos look clean but are near-identical, static, and short on motion. Trajectory-based distillation holds up on images but has not carried its weight on video.
PDD targets fast, simple, and diverse at once.
PDD's trick is "multiple steps per evaluation." The pretrained model predicts one mean velocity per forward pass; PDD repeats the final linear layer N times, one copy per denoising interval, so a single forward pass outputs the velocities for L consecutive steps. Sampling then advances block by block: the current state plus the sum of each step's velocity times its time step length, covering L intervals in one move.
Training uses a plain regression loss that fits the student's parallel predictions to the teacher's mean velocity at each step. The teacher's velocity comes from a cheap ODE solver (Euler or midpoint), with no Jacobian-vector products or finite differences. That is the root of "simple and stable": no adversarial term, no VSD, just regression.
A few design choices matter. Variable block size (Lmin to Lmax) lets one set of weights serve several NFE counts without extra time conditioning. At inference, the N copied linear layers fuse into a single weighted-average layer. Training is data-free and on-policy, alternating solver steps with training steps. The architecture and pipeline are compatible with any pretrained model, which is why it drops straight onto Wan, Qwen-Image and LTX.
Images (ImageNet-256, teacher SiT-XL+REPA, 1-step): PDD-Midpoint FID is 2.69, on par with Pi-Flow (2.85) but behind FreeFlow (1.45). At one step it is not the strongest.
The text-to-image (Qwen-Image) and quality-vs-diversity picture is more telling:
| Method | NFE | OneIG-EN up | DPG up | GenEval up | OneIG diversity up |
| DMD2 | 4 | 0.524 | 88.25 | 0.85 | 0.095 |
| Pi-Flow | 4 | 0.533 | 88.11 | 0.85 | 0.182 |
| PDD-Midpoint | 4 | 0.538 | 88.66 | 0.86 | 0.174 |
On quality, PDD roughly matches or edges DMD2. On diversity it is almost double (0.174 vs 0.095), which is exactly the pain point it targets.
Text-to-video (Wan2.1, VBench, 4 steps):
| Model | Method | VBench total up | V-JEPA2 diversity up |
| Wan 1.3B | AnyFlow | 84.45 | 0.0704 |
| Wan 1.3B | DMD2 | 84.69 | 0.0833 |
| Wan 1.3B | PDD-Midpoint | 84.94 | 0.1032 |
| Wan 14B | AnyFlow | 84.95 | 0.0786 |
| Wan 14B | PDD-short | 84.92 | 0.0791 |
On 1.3B, PDD wins on both quality and diversity. On 14B its total (84.92) just trails AnyFlow (84.95) while still scoring higher on diversity, a gap at the 0.03-point scale. On LTX-2.3, the teacher's 4x30 steps are matched by PDD in 8 steps, on par with the official distilled model, and the audio track is covered too.
For generation teams, the practical pull is simplicity. Regression-only training means no adversarial tuning hell and no VSD instability, so reproduction and iteration are cheap. The same parallel-decoder head attaches to any pretrained model, already shown on Wan, Qwen-Image and LTX. Crucially it preserves diversity, the known weakness of DMD2-style methods, and does so without giving up quality.
The cold water: on the 14B model PDD's total does not cleanly beat AnyFlow (84.92 vs 84.95); its edge is diversity, not absolute quality. For a single-image-quality-at-all-costs setting that does not care about sample diversity, it may not be the best pick, and at one-step image generation it is still not a match for FreeFlow.
The authors list several. Large-scale text-to-image and text-to-video rely on data-free training; data-dependent settings were tried only on small-scale ImageNet, so generalization to data-driven video distillation is unverified. On ImageNet the 8-step FID is worse than 4-step (mitigable with a lower guidance scale). The architecture must match the teacher exactly, so changing the backbone means retraining. Variable NFE only works within preset block-size ranges, not arbitrary interpolation.
One extra doubt: fewer NFE means heavier reliance on an accurate teacher mean-velocity estimate, and the paper does not isolate the error of the Euler/midpoint solver at very few steps. The quality ceiling may be capped by solver precision.