Research map / Video / Video generation
SNED: Superposition Network Architecture Search for Efficient Video Diffusion 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.
Proposes SNED, a weight-sharing supernet NAS for video diffusion models that yields subnets of varying size and resolution options for pixel-space and latent-space models.
Problem
Video diffusion models are large and computationally expensive for training and inference, and different size and resolution targets normally require separately designed architectures.
Approach
Applies one-shot weight-sharing supernet architecture search to video diffusion models, combining dynamic channel and fine-grained dynamic block search spaces with superposition training across resolutions and a sampling warmup that gradually shrinks the search space.
Key results
- Pixel-space SNED-S (0.64B) achieves FVD12 472.3 and KVD12 16.8 at 16.0 s latency versus supernet SNED-B (1.6B) at FVD12 544.4 and 24.4 s, a 1.53x speedup.
- Latent-space SNED 548M matches LVDM at FVD16 298.3 vs 295.1 and KVD16 20.8 vs 20.9 on Sky Timelapse 256x256.
- Smaller latent subnets (411M, 274M) reach FVD16 348.2 and 472.3 respectively, with latency reduced from 86.8 s to 74.2 s and 66.7 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SNED with 548M parameters instead of the released LVDM model | LVDM · Sky Timelapse, 256×256 | FVD16: 295.1 -> 298.3 | Hurt |
| Use SNED with 548M parameters instead of the released LVDM model | LVDM · Sky Timelapse, 256×256 | KVD16: 20.9 -> 20.8 | Helped |
| Use SNED with 548M parameters instead of the released LVDM model | LVDM · Sky Timelapse, 256×256 | Time: 86.8 -> 86.8 | No clear effect |
| Reduce model size from 548M to 411M parameters (ablation) | SNED · Sky Timelapse, 256×256 | FVD16: 298.3 -> 348.2 | Hurt |
| Reduce model size from 548M to 411M parameters (ablation) | SNED · Sky Timelapse, 256×256 | Time: 86.8 -> 74.2 | Helped |
| Reduce model size from 548M to 274M parameters (ablation) | SNED · Sky Timelapse, 256×256 | FVD16: 298.3 -> 472.3 | Hurt |
| Reduce model size from 548M to 274M parameters (ablation) | SNED · Sky Timelapse, 256×256 | Time: 86.8 -> 66.7 | Helped |
| Reduce pixel-space model size from 1.60B to 0.96B parameters (ablation) | SNED pixel-space base model · Pixel-space video diffusion base model | FVD12: 544.4 -> 452.2 | Helped |
| Reduce pixel-space model size from 1.60B to 0.96B parameters (ablation) | SNED pixel-space base model · Pixel-space video diffusion base model | Time: 24.4 -> 18.1 | Helped |
Surprising findings
- Pixel-space subnets smaller than the supernet sometimes obtain better FVD and KVD scores than the full supernet.
Limitations
- Pixel-space experiments use an internal 19M-pair dataset that is not released, limiting comparison.
- Latent-space comparison is limited to a single dataset (Sky Timelapse) and a basic LVDM baseline rather than the long-video version.
- Evaluation is limited to FVD/KVD and latency with no human or text-alignment evaluation.
Experimental setup
Datasets: internal 19M video-text pairs Sky Timelapse
Benchmarks: Sky Timelapse
Metrics: FVD KVD latency (s)
Baselines: LVDM
Builds on: Imagen Video LVDM Autoformer
Closest papers by meaning
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution (CVPR 2024 · similarity 0.88)
- Hierarchical Patch Diffusion Models for High-Resolution Video Generation (CVPR 2024 · similarity 0.87)
- DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution (NeurIPS 2025 · similarity 0.87)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.87)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.87)
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.