Research map / Video / Video generation
SF-V: Single Forward Video Generation Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts Stable Video Diffusion into a single-step image-to-video generator by latent adversarial training with a frozen-backbone discriminator carrying spatial and temporal heads.
Problem
Video diffusion models need many denoising steps, making sampling expensive; single-step video generation had not been achieved.
Approach
Extends latent adversarial distillation (LADD-style) to video by initializing the generator and a frozen-backbone discriminator from SVD, adding parallel spatial and temporal discriminator heads, sampling discriminator noise from a lognormal (EDM-style) distribution, and adding a Pseudo-Huber reconstruction loss, trained on real video data.
Key results
- Single-step SF-V achieves FVD 180.9 on UCF-101 versus 153.4 for 25-step SVD and 1917.2 for UFOGen at one step.
- Latency of 0.30 s for one UNet pass versus 10.79 s for 25-step SVD, about 23x speedup over SVD (comparable to 16-step SVD with FVD 194.4).
- Discriminator ablation: spatial plus temporal heads give FVD 180.9 vs 514.7 (spatial only) and 539.2 (temporal only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use spatial and temporal discriminator heads together instead of spatial heads only (ablation) | SF-V · UCF-101 | FVD: 514.7 -> 180.9 | Helped |
| Use spatial and temporal discriminator heads together instead of temporal heads only (ablation) | SF-V · UCF-101 | FVD: 539.2 -> 180.9 | Helped |
| Set P_mean=-2.0 and P_std=-1.0 instead of the default -1.0 and -1.0 (ablation) | SF-V discriminator noise distribution · UCF-101 | FVD: 180.9 -> 3370.4 | Hurt |
| Set P_mean=0.0 and P_std=1.0 instead of the default -1.0 and -1.0 (ablation) | SF-V discriminator noise distribution · UCF-101 | FVD: 180.9 -> 416.7 | Hurt |
| Set P_mean=1.0 and P_std=1.0 instead of the default -1.0 and -1.0 (ablation) | SF-V discriminator noise distribution · UCF-101 | FVD: 180.9 -> 632.9 | Hurt |
| Use SF-V instead of SVD at 16 sampling steps | SVD · UCF-101, 14 frames | FVD: 194.4 -> 180.9 | Helped |
| Use SF-V instead of AnimateLCM at 8 steps | AnimateLCM · UCF-101, 14 frames | FVD: 281.0 -> 180.9 | Helped |
| Use SF-V instead of AnimateLCM* at 8 steps | AnimateLCM* · UCF-101, 14 frames | FVD: 321.1 -> 180.9 | Helped |
| Use SF-V instead of UFOGen at 1 step | UFOGen · UCF-101, 14 frames | FVD: 1917.2 -> 180.9 | Helped |
| Use SF-V instead of LADD at 1 step | LADD · UCF-101, 14 frames | FVD: 1893.8 -> 180.9 | Helped |
| Use SF-V at 1 step instead of SVD at 25 steps | SVD · UCF-101, 14 frames | Latency (s): 10.79 -> 0.30 | Helped |
| Use SF-V at 1 step instead of SVD at 25 steps | SVD · UCF-101, 14 frames | FVD: 153.4 -> 180.9 | Hurt |
Surprising findings
- Using only spatial or only temporal discriminator heads leads to much worse FVD, showing both are needed for quality and motion.
- Lognormal sampling of discriminator noise with too-low noise levels (P_mean=-2) collapses performance (FVD 3370.4).
Limitations
- Complex motion (e.g., running) yields blurry or unsatisfactory frames, inherited from SVD.
- Temporal VAE decoder and image encoder dominate runtime at single step and are not accelerated.
- Training relies on an internal dataset of about one million videos that is not released.
Experimental setup
Datasets: internal video dataset (~1M videos) UCF-101
Benchmarks: UCF-101 FVD
Metrics: FVD latency (s)
Baselines: SVD AnimateLCM UFOGen LADD
Builds on: Stable Video Diffusion LADD EDM UFOGen
Closest papers by meaning
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.93)
- Diffusion Adversarial Post-Training for One-Step Video Generation (ICML 2025 · similarity 0.93)
- BIVDiff: A Training-Free Framework for General-Purpose Video Synthesis via Bridging Image and Video Diffusion Models (CVPR 2024 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.89)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.