Research map / Video / Video generation
FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention
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.
Training-free method (FreeLong) that extends 16-frame text-to-video diffusion models (LaVie, VideoCrafter2) to 128-frame videos by blending low-frequency global and high-frequency local temporal attention features in the 3D FFT domain.
Problem
Short-trained video diffusion models degrade when directly generating long videos (e.g., 128 frames), with high-frequency spatial loss and temporal high-frequency increase causing blur and flicker, and existing sliding-window approaches lack global consistency.
Approach
Decouple temporal attention into a windowed local branch and a full global branch, then fuse them in the frequency domain by taking low-frequency components from global features and high-frequency components from local features, applied during early denoising steps.
Key results
- On 200 VBench prompts, FreeLong reaches subject consistency 95.16 vs 92.30 for FreeNoise and 88.95 for direct sampling.
- FreeLong attains the highest imaging quality (67.55) and motion smoothness (96.85) among compared training-free methods.
- FreeLong's inference time is 2.2s, faster than FreeNoise and sliding window (2.6s each), since it uses single-pass temporal attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Subject consistency: 92.30 -> 95.16 | Helped |
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Background consistency: 95.87 -> 96.80 | Helped |
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Motion smoothness: 96.32 -> 96.85 | Helped |
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Temporal flickering: 94.94 -> 96.04 | Helped |
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Imaging quality: 67.14 -> 67.55 | Helped |
| Use FreeLong instead of FreeNoise | LaVie / VideoCrafter long-video generation · VBench (200 prompts), 128-frame video | Inference time: 2.6s -> 2.2s | Helped |
| Use global features only instead of global-low/local-high spectral blending (ablation) | FreeLong · Long-video generation | Fidelity | Hurt |
| Use local features only instead of global-low/local-high spectral blending (ablation) | FreeLong · Long-video generation | Temporal consistency | Hurt |
| Directly combine global and local features without spectral blending (ablation) | FreeLong · Long-video generation | Fidelity | Hurt |
| Blend low-frequency global features with high-frequency local features (ablation) | FreeLong · Long-video generation | Video fidelity | Helped |
| Blend low-frequency global features with high-frequency local features (ablation) | FreeLong · Long-video generation | Temporal consistency | Helped |
Surprising findings
- Directly sampling 128 frames from a 16-frame model lowers high-frequency SNR in space (0.68) while temporal high-frequency SNR rises (1.5), explaining flicker and blur.
Limitations
- Global attention features still contain high-frequency degradation, so fusion is restricted to early denoising steps (tau=25).
- Evaluation uses only 200 VBench prompts and two base models.
Experimental setup
Datasets: VBench
Benchmarks: VBench
Metrics: subject consistency background consistency motion smoothness temporal flickering imaging quality inference time
Baselines: Direct sampling Sliding window FreeNoise Gen-L-Video StreamingT2V Nuwa-XL
Builds on: FreeNoise latent diffusion models LaVie VideoCrafter2
Closest papers by meaning
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.93)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.92)
- Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion (AAAI 2025 · similarity 0.91)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.90)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.90)
- FreePCA: Integrating Consistency Information across Long-short Frames in Training-free Long Video Generation via Principal Component Analysis (CVPR 2025 · similarity 0.89)
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.