UnStep is a training-free wrapper that speeds up few-step causal video diffusion models. It reduces the number of steps used at inference and limits the temporal window in the attention KV cache. To recover quality lost from these reductions, UnStep uses two inference-only techniques: renoising generated latent frames and applying truncated SVD to attention projections. It also includes a runtime stack optimization for the DiT and VAE decoder. The result is a substantial speedup, achieving 50 FPS on a single H100 and 77 FPS on a GB200, with no quality loss compared to the original model.
Training-free wrapper accelerates causal video diffusion to 50 FPS without quality loss
UnStep reduces inference steps and attention window size, then recovers quality with renoising and truncated SVD.
Industry
Youssef Mansour · Enis Simsar · Fadime Sener · Markos Georgopoulos · Albert Pumarola · Ali Thabet · +1 more
Meta Superintelligence Labs
Research Digest··3 min read
In one line
UnStep accelerates causal video diffusion models to 50 FPS on an H100 without quality loss and without retraining.
Before you read this
Three ideas this paper builds ondiffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.75 papers →test-time adaptationAdapting a deployed model during inference using new data or interactions, without a separate training phase.84 papers →kv cache compressionReducing the size of the key-value cache in transformers via selective retention, low-rank updates, or adaptive compression.82 papers →
Read it as
The authors present UnStep, a training-free wrapper that accelerates already-distilled causal video diffusion models by running them with fewer denoising steps and a limited temporal KV cache.
Mansour et al. start with a few-step causal video diffusion model (e.g., CausVid, distilled from Wan) and run it with even fewer DiT steps (e.g., 1–2 instead of 4) and a limited temporal KV attention window (e.g., 4–8 frames).
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code