Accelerating Wan 2.2 video generation by 16.8× via 4-step diffusion distillation

Generating a single 5-second video clip ($480 \times 832$ resolution, 81 frames) using the default Wan 2.2 A14B pipeline requires 809 seconds ($\sim 13.5$ minutes) of processing time on an NVIDIA A100 GPU. This latency stems from executing 80 forward passes through a 14-billion-parameter Transformer architecture.

Applying distillationTraining a fast student model to reproduce a slow teacher's output in far fewer steps. Here the student lands a clip in 4 steps instead of the teacher's 40. LoRA adapters (Wan2.2-Lightning) compresses the 40-step diffusion schedule into a 4-step trajectory while eliminating Classifier-Free Guidance (CFG) evaluations. This reduces network evaluations from 80 to 4 per clip, lowering generation latency to 48 seconds ($16.8\times$ speedup) without requiring model retraining or custom CUDA kernels.

40 steps · 809 s

4 steps · 48 s16.8×

Profiling Latency in Denoising Schedules and Guidance Passes

Wan 2.2 A14B is a dual-expert Mixture-of-Experts (MoEA model split into specialist sub-networks, only some active at a time. Wan 2.2 has two: a high-noise layout expert and a low-noise detail expert, ~14B active per step out of ~27B on disk.) diffusionA generative model that starts from noise and removes a little on each step until an image or clip appears. Video diffusion runs this over all frames at once. transformer containing $\sim 27$ billion parameters on disk ($\sim 14$ billion active per step):

• High-Noise Layout Expert: Processes timesteps 1–26 to generate global scene structure, composition, and motion dynamics.
• Low-Noise Detail Expert: Processes timesteps 27–40 to refine surface textures and high-frequency details.

Generating $480 \times 832$ resolution across 81 frames produces 32,760 tokens per frame representation. Sequence self-attention accounts for $73\%$ of per-step compute time.

high-noise layout expertsteps 1–26

low-noise detail expertsteps 27–40

distilled: 4 steps span the same trajectory

Total generation latency depends on the total count of network forward passes ($N_{\text{evals}}$):

$$N_{\text{evals}} = N_{\text{steps}} \times N_{\text{passes}}$$

The baseline pipeline executes 40 denoising stepsThe number of times the sampler calls the network to walk from pure noise to the final clip. Wan 2.2's default schedule is 40.. Standard Classifier-Free GuidanceRuns the network twice per step, once with the prompt and once without, then pushes the result away from the unconditional prediction to sharpen prompt adherence. It doubles the cost per step. (CFG) evaluates the network twice per step (conditional and unconditional passes):

$$N_{\text{evals}} = 40 \times 2 = 80 \text{ passes}$$

Baseline (40 steps × CFG)80 evaluations

Distilled (4 steps, no CFG)4 evaluations · 20× evaluation reduction

Mathematical function of Classifier-Free Guidance (CFG)

CFG adjusts latent prediction updates by extrapolating away from un-conditioned estimates:

$$\epsilon_{\text{guided}} = \epsilon_{\text{uc}} + s \cdot (\epsilon_{\text{c}} - \epsilon_{\text{uc}})$$

where $s$ is the guidance scale factor. Because CFG requires evaluating $\epsilon_{\text{c}}$ and $\epsilon_{\text{uc}}$ separately, each step incurs two full forward passes. Distillation embeds guidance directions directly into student weights, allowing $s=1.0$ (disabling unconditional passes).

Distillation and Guidance Elimination Mechanics

Reducing $N_{\text{evals}}$ by running the baseline model at lower step counts (e.g., 12 steps) introduces severe discretization errors because the baseline noise schedule is not optimized for large step sizes.

Instead, we integrate Phased-DMD distillation adapters (Wan2.2-Lightning):

1. Step Reduction ($10\times$): Compresses the 40-step trajectory into 4 discrete UniPC solver steps.
2. Guidance Elimination ($2\times$): Folds CFG trajectory vectors into low-rank LoRALow-rank adapter. A small set of extra weights that modifies a frozen base model without retraining it. Here, one rank-64 LoRA per expert, a few hundred MB rather than a new 14B network. weights (rank-64 per expert), enabling $s=1.0$.

$$\text{Speedup} = (40 \cdot 2) / (4 \cdot 1) = 20$$

Hyperparameter configuration for distilled inference

Distilled inference uses three primary hyperparameter updates:

• Scheduler: UniPC multistep ODE solver (maintains stability at step counts $\le 4$).
• Flow Shift: Set to 5 (down from baseline 12), allocating higher step density to early high-noise layout resolution.
• Guidance Scale: Set to 1.0 (disables unconditional pass execution).

Multi-Scene Empirical Quality Benchmarks

To verify performance stability across diverse visual distributions, we benchmarked 13 test scenes across lighting, motion, and subject conditions using identical random seeds.

40-step

4-step

Breaking ocean wave, slow motion

40-step

4-step

Campfire, embers and sparks

40-step

4-step

Sports car on a neon street in the rain

40-step

4-step

Aerial over a city skyline at dusk

40-step

4-step

Elderly fisherman, wind in his beard

40-step

4-step

Ballet dancer under a spotlight

40-step

4-step

Golden eagle over misty peaks

40-step

4-step

Tokyo pedestrian crossing at night

40-step

4-step

Syrup poured over pancakes

40-step

4-step

Hummingbird at a flower, macro

40-step

4-step

Sea turtle over a coral reef

40-step

4-step

Walking through heavy snowfall

40-step

4-step

Camel caravan over dunes at sunset

Color Degradation and Checkpoint Iteration Analysis

Initial distilled checkpoints (Lightning V1.1) exhibited dynamic range compression, producing desaturated output frames on high-contrast scenes (e.g., campfire ember lighting).

We evaluated three corrective strategies:

1. Luma Matching (tone_match): Adjusted linear brightness and contrast against baseline references. (Inadequate for restoring lost color saturation).
2. Post-Processing Color Grading: Applied per-channel CLAHE histogram equalization and HSV saturation scaling. (Restored dynamic range but introduced post-processing overhead).
3. Model Checkpoint Upgrade (Lightning V2.0): Retrained LoRA weights with improved loss weighting. V2.0 restored baseline color warmth directly during 4-step generation without post-processing steps.

Base · 40 stepsBaseline color reference

V1.1 · 4 steps, rawDesaturated dynamic range

V1.1 · 4 steps + rich gradePost-processed color grade

V2.0 · 4 steps, no grade · pickRestored baseline saturation

Methodological limits of LPIPS/PSNR metrics on generative trajectories

Standard perceptual distance metrics (LPIPS, PSNR) rely on exact pixel alignment between baseline and target samples. Because UniPC 4-step trajectories follow distinct ODE paths relative to 40-step DDIM baselines, pixel positions diverge naturally while maintaining visual coherence. Consequently, quality verification relies on visual inspection across benchmark suites rather than LPIPS reconstruction error.

Step Count Trajectory and Speed-Fidelity Pareto Frontiers

Sweeping step counts across baseline, distilled, and hybrid pipelines establishes the trade-off between generation latency and visual fidelity:

Base · 40 steps719 s · 1.0× speedup

Base · 20 steps362 s · 2.0× speedup

Base · 12 steps221 s · 3.3× speedup

Hybrid: draft + refine155 s · 4.6× speedup

V2.0 · 8 steps82 s · 8.8× speedup

V2.0 · 4 steps · pick44 s · 16.0× speedup

V1.1 · 4 steps44 s · 16.0× speedup

Base · 40 steps1.0× · 719 s

Base · 20 steps2.0× · 362 s

Base · 12 steps3.3× · 221 s

Hybrid (4-step draft + base refine)4.6× · 155 s

V2.0 · 8 steps8.8× · 82 s

V2.0 · 4 steps16.0× · 44 s · Optimal balance

Architecture of the Hybrid Draft-and-Refine pipeline

The hybrid pipeline generates a 4-step distilled draft, then passes the output tensor into WanVideoToVideoPipeline using the baseline low-noise detail expert at denoising strength $0.4$. This configuration executes 4 draft steps + 6 refinement steps ($155\text{s}$ total, $4.6\times$ speedup), providing exact texture matching to the 40-step baseline.

Residual Fixed Overhead in VAE Decoding and Text Encoding

At 4 steps, Transformer inference latency drops from 803 seconds to 42.5 seconds ($88.4\%$ of total clip execution time). Non-iterative fixed overheads account for the remaining latency:

• Text Encoding (T5-XXL / CLIP): $\sim 1.2$ seconds.
• VAEVariational autoencoder. The decoder that turns the model's compressed latent into actual pixels. It runs once per clip regardless of step count. Spatial/Temporal Decoding: $\sim 4.4$ seconds.

Because fixed overheads total $\sim 5.6$ seconds, further step reductions (e.g., to 2 steps) yield diminishing returns unless accompanied by lightweight VAE decoders (e.g., TAEHV).

Cross-Workload Acceleration Hierarchy

Comparing step distillation against lower-level kernel optimizations highlights the primacy of evaluation count reduction:

Wan 2.2 Two-Expert Step Cache (Kernel)1.78× · Inter-step feature caching

Gemma-4 Decoding Stack (FP8 + MTP)2.34× · Quantization & speculative decoding

Wan 2.2 Step Distillation (This Work)16.8× · Evaluation count reduction (80 → 4)

Conclusion

Distilling Wan 2.2 A14B from 80 evaluations to 4 steps reduces video generation latency from 809 seconds to 48 seconds ($16.8\times$ speedup) on NVIDIA A100 GPUs. Updating adapter checkpoints to Lightning V2.0 resolves color dynamic range loss without post-processing steps.

Glossary

Glossary — every term, defined

Diffusion

A generative model that starts from noise and removes a little on each step until an image or clip appears. Video diffusion runs this over all frames at once.

Denoising step

One call to the network on the way from pure noise to the final clip. Wan 2.2 defaults to 40; the distilled model uses 4.

Classifier-free guidance (CFG)

Runs the network twice per step, once with the prompt and once without, then extrapolates away from the unconditional prediction to sharpen prompt adherence. It doubles the per-step cost. guidance_scale=1.0 turns it off.

Distillation

Training a fast student model to match a slow teacher's output in far fewer steps. The Lightning student lands a clip in 4 steps instead of the teacher's 40.

LoRA (low-rank adapter)

A small set of extra weights that modifies a frozen base model without retraining it. Wan2.2-Lightning is one rank-64 LoRA per expert, a few hundred MB rather than a new 14B network.

Mixture of experts

A model split into specialist sub-networks, only some active at a time. Wan 2.2 A14B has two experts, high-noise and low-noise, ~14B active per step out of ~27B on disk.

Flow shift

A knob that redistributes where the sampling steps land along the noise schedule. The base recipe uses 12; the distilled run uses 5, so more of its 4 steps go to the high-noise layout phase.

VAE decoder

The variational-autoencoder decoder that turns the model latent into pixels. It runs once per clip regardless of step count, which is why it is now a fixed cost rather than a per-step one.

UniPC

A multi-step ODE solver used as the sampler. It stays stable at the very low step counts where a plain DDIM or Euler schedule breaks down.

Phased-DMD

The distribution-matching distillation method (DMD family) that trained Lightning. It matches the teacher’s output distribution in phases across the noise schedule, so one student covers the whole trajectory in a few steps.

Network evaluation

One full forward pass of the ~14B-active transformer. The base recipe runs 80 per clip (40 steps × 2 for CFG); the distilled recipe runs 4.

Hybrid (draft + refine)

A fast 4-step distilled draft, then a base low-noise refine as a video-to-video pass. 155 s at denoising strength 0.4 (~4.6×); the quality-first point when a clip must match the base’s exact texture.