FastWan: Generating a 5-Second Video in 5 Seconds via Sparse Distillation

TL;DR

FastWan is a family of post-trained Wan-series video models that compose Video Sparse Attention (VSA) with Distribution Matching Distillation (DMD) in a single joint training framework — what the team terms sparse distillation. The headline operating points: FastWan2.1-1.3B generates a 5-second 480p clip in 5 s end-to-end on a single H200 (1 s denoising) or 21 s on a single RTX 4090 (2.8 s denoising); FastWan2.2-5B generates a 5-second 720p clip in 16 s on a single H200. The argument is structural: existing sparse-attention methods rely on multi-step denoising redundancy that disappears under aggressive (1–4-step) distillation, so they must instead use a sparse attention mechanism (VSA) that learns data-dependent sparsity at training time and is therefore compatible with step compression. Models, recipe, and synthetic training corpus (882k Wan-generated videos) are Apache-2.0.

Key claims

Sparse distillation is the first method to jointly train sparse attention and step distillation in a unified framework, rather than sequentially [§“Sparse Distillation: Making Video Generation Go Brrrr”].
Existing sparse-attention methods (STA, SVG) prune attention by exploiting redundancy across multi-step denoising — typically sparsifying only late-stage steps — so they collapse when distillation compresses 50 steps into 1–4 [§“Why Existing Sparse Attention Fails Under Distillation”].
Video Sparse Attention (VSA) is a learned sparse attention kernel that dynamically identifies important tokens and can directly replace FlashAttention during training, making it the first sparse attention mechanism compatible with distillation [§“Why Video Sparse Attention is Compatible With Distillation”].
In the sparse-distillation setup, the few-step student uses VSA while both the frozen real-score network and the trainable fake-score network use full attention — runtime acceleration is decoupled from distillation-quality supervision [§“How Sparse Distillation Works”; Fig. 1].
DiT-denoising-time on a single H200 for Wan2.1-1.3B 480p drops from 95.21 s (FA2) → 2.88 s (FA2 + DMD) → 2.14 s (FA3 + DMD) → 1.49 s (FA3 + DMD + torch.compile) → 0.98 s (VSA + DMD + torch.compile) [Table “How Fast is FastWan”].
DiT-denoising-time on a single H200 for Wan2.1-14B 720p drops from 1746.5 s (FA2) → 13 s (VSA + DMD + torch.compile), a ~134× speedup [Table “How Fast is FastWan”].
Wan2.2-TI2V-5B’s ~20K sequence length is too short to benefit from sparse attention; only DMD with full attention is applied (157.21 s → 4.67 s on H200) [Table “How Fast is FastWan”; §“Models and Recipes”].
Attention operations consume more than 85% of total inference time on Wan2.1-14B 720p (50 steps, >80K tokens), motivating the sparse-attention lever [§“Sparse Distillation: Making Video Generation Go Brrrr”].
Sparse distillation is method-agnostic on the distillation side: because VSA only touches the student, the recipe is claimed compatible with consistency distillation, progressive distillation, or GAN-based distillation, not only DMD [§“How Sparse Distillation Works”].
Training uses an in-house synthetic dataset generated by the original Wan models — 600k 480p + 250k 720p videos from Wan2.1-T2V-14B and 32k from Wan2.2-TI2V-5B — released under an open license [§“Distillation Details”].
Memory recipe to fit three 14B models (student trainable, fake-score trainable, real-score frozen) into GPU memory at long sequence length: FSDP2 parameter sharding on all three, activation checkpointing, careful inference-vs-train mode switching across distillation stages, gradient accumulation [§“Distillation Details”].
Reproducibility budget: Wan2.1-T2V-1.3B sparse distillation runs 4k steps on 64 H200 GPUs = 768 GPU-hours ≈ $2,603 at$ 3.39/H200/hour [§“Distillation Details”].

Method

Sparse distillation builds on Distribution Matching Distillation (DMD2) by adding learned sparsity to the student only. Three networks are initialized from Wan2.1 and used jointly during training: (1) a few-step student with Video Sparse Attention (trainable), (2) a frozen real-score network with full attention, and (3) a trainable fake-score network with full attention. Each training step takes a noisy video, the student performs one denoising step with VSA, the output is renoised and passed to both score networks (one denoising step each, full attention); the difference between real and fake scores forms the distribution-matching gradient that updates the student. In parallel, the fake-score model is updated via a diffusion loss on the student’s outputs.

The structural argument is that prior sparse-attention recipes hand-design sparsity patterns based on the assumption that early denoising steps need dense attention and late ones can be pruned. Step distillation collapses this assumption — there are no late steps. VSA sidesteps this by learning its sparsity pattern from data during training, so it can be co-trained with step compression rather than retrofitted afterwards. Because VSA only touches the student, the rest of the distillation pipeline (real-/fake-score networks, the DMD objective itself) is unchanged.

The companion technical detail: at Wan2.2-TI2V-5B’s ~20K sequence length, attention is not yet the bottleneck, so the 5B model gets DMD-only distillation with full attention. The sparse-attention win scales with sequence length — it’s most pronounced on Wan2.1-14B 720p (>80K tokens), where attention is >85% of inference cost.

Results

Wan2.1-1.3B 480p, single H200, full pipeline: 5 s end-to-end (1 s denoising) for a 5-second clip — VSA + DMD + torch.compile recipe [§“How Fast is FastWan”; §“TL;DR”].
Wan2.1-1.3B 480p, single RTX 4090, full pipeline: 21 s end-to-end (2.8 s denoising) for a 5-second clip [§“TL;DR”]. This is the strongest single-consumer-GPU operating point in the post.
Wan2.2-5B 720p, single H200, full pipeline: 16 s end-to-end for a 5-second clip [§“TL;DR”].
Wan2.1-14B 720p DiT-denoising on a single H200: 1746.5 s (FA2 baseline) → 13 s (VSA + DMD + torch.compile), a 134× speedup. With FA2 + DMD alone (no sparsity) the time is 52 s, so the sparse-attention lever buys an additional ~4× on top of step distillation [Table “How Fast is FastWan”].
Wan2.1-1.3B 480p DiT-denoising on a single H200: 95.21 s → 0.98 s, a ~97× speedup. Step-by-step lever decomposition: FA2 → FA2+DMD = 33×, FA3+DMD = 1.35× on top, torch.compile = 1.44× on top, VSA = 1.52× on top [Table “How Fast is FastWan”].
No quantitative quality benchmarks (VBench, human preferences) are reported in the post; the release is framed around throughput and reproducibility, with the implicit quality claim that VSA + DMD does not degrade vs. DMD alone.

Why it’s interesting

This is the system-level companion to the Diffusion Distillation cluster currently on the wiki: most recent distillation papers attack the trajectory parameterization (ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation), role separation across student steps (Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis), the AR-bidirectional injectivity gap (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation), or resolution/chunk cascades (Cross-Resolution Distribution Matching for Diffusion Distillation, Streaming Autoregressive Video Generation via Diagonal Distillation) — but they all assume the underlying attention kernel is dense. FastWan inverts the assumption: at long-sequence (>80K-token) video DiTs, attention itself is the dominant cost at low NFE, and the most underexplored axis is whether the attention pattern can be learned jointly with step compression. The empirical claim — that retrofitted sparse-attention schemes (STA, SVG) which rely on multi-step redundancy fail under few-step distillation — is a useful negative result that bounds the prior approaches in the Diffusion Distillation thread.

It also complements FastVideo: Create a 5s 1080p Video in 4.5s on a Single GPU, which is the same team’s systems-only recipe for accelerating LTX-2.3 (NVFP4, fused kernels, B200-tuned attention) with no algorithmic step or attention changes. FastWan is the algorithmic sibling: same Hao AI Lab, same FastVideo codebase, but the lever is sparse + few-step joint training on Wan2.1/2.2 rather than precision and kernel fusion on LTX-2. Together they give the wiki two clean operating points on the same Pareto from one lab — and per the blog they can be combined.