Research map / Video / Video generation
RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers
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 RIFLEx, which lowers the intrinsic RoPE frequency of video diffusion transformers so generated videos extrapolate in length (2x training-free, 3x with light fine-tuning) without temporal repetition or slow motion.
Problem
Length extrapolation for video diffusion transformers beyond training length; existing RoPE extrapolation methods (PE, PI, NTK, YaRN, TASR) cause temporal repetition or motion deceleration.
Approach
Identifies an intrinsic RoPE frequency whose period matches the first observed repetition frame and reduces only that frequency so its period exceeds the extrapolated length, suppressing repetition while preserving motion speed.
Key results
- On CogVideoX-5B 2x training-free, RIFLEx achieves NoRepeat Score 54.2 and Dynamic Degree 59.4 versus PE 46.6/58.6 and PI 59.0/5.0, balancing repetition and slow motion.
- On HunyuanVideo 2.3x training-free, RIFLEx reaches NoRepeat Score 54.0 versus NTK 20.0.
- With fine-tuning on HunyuanVideo 2x, RIFLEx reaches NoRepeat Score 89.0 and Dynamic Degree 82.0; users preferred training length over RIFLEx in 56.0% of comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RIFLEx instead of the strongest listed baselines in training-free 2× extrapolation | CogVideoX-5B · 2× extrapolation, training-free | Dynamic Degree: 58.6 -> 59.4 | Helped |
| Use RIFLEx instead of YaRN | CogVideoX-5B · 2× extrapolation, training-free | NoRepeat Score: 59.4 -> 54.2 | Hurt |
| Use RIFLEx instead of NTK | CogVideoX-5B · 2× extrapolation, training-free | Imaging Quality: 55.3 -> 56.9 | Helped |
| Fine-tune RIFLEx rather than PE | CogVideoX-5B · 2× extrapolation, fine-tuning | NoRepeat Score: 13.2 -> 61.3 | Helped |
| Fine-tune RIFLEx rather than PE | CogVideoX-5B · 2× extrapolation, fine-tuning | Imaging Quality: 56.6 -> 60.4 | Helped |
| Use RIFLEx instead of the best listed baseline | HunyuanVideo · 2× extrapolation, training-free | NoRepeat Score: 86.0 -> 72.0 | Hurt |
| Use RIFLEx instead of PE | HunyuanVideo · 2× extrapolation, training-free | Dynamic Degree: 63.0 -> 57.0 | Hurt |
| Use RIFLEx instead of NTK | HunyuanVideo · 2× extrapolation, training-free | Overall Aspects: 1.6 -> 1.4 | Helped |
| Use RIFLEx instead of NTK | HunyuanVideo · 2.3× extrapolation, training-free | NoRepeat Score: 20.0 -> 54.0 | Helped |
| Use RIFLEx instead of NTK | HunyuanVideo · 2.3× extrapolation, training-free | Imaging Quality: 65.5 -> 65.0 | Hurt |
| Fine-tune RIFLEx rather than PE | HunyuanVideo · 2× extrapolation, fine-tuning | NoRepeat Score: 40.0 -> 89.0 | Helped |
| Fine-tune RIFLEx rather than PE | HunyuanVideo · 2× extrapolation, fine-tuning | Dynamic Degree: 74.0 -> 82.0 | Helped |
Surprising findings
- NTK performs well on HunyuanVideo at 2x extrapolation only incidentally, because it unintentionally reduces the intrinsic frequency, and fails at 2.3x.
- Reducing low-frequency components has negligible impact, while reducing high-frequency components slows motion.
Limitations
- Intrinsic frequency is identified empirically by visual inspection rather than a theoretical criterion.
- Quality degrades beyond 3x extrapolation, and 3x requires fine-tuning due to training-testing mismatch.
- Fine-tuning uses a private dataset that is not released.
Experimental setup
Datasets: private 20,000-video fine-tuning set VBench prompts
Benchmarks: VBench
Metrics: NoRepeat Score Dynamic Degree Imaging Quality Overall Consistency user study ranking
Baselines: PE PI NTK YaRN TASR No extrapolation
Builds on: RoPE Position Interpolation NTK-aware scaled RoPE YaRN
Closest papers by meaning
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.88)
- Video Motion Transfer with Diffusion Transformers (CVPR 2025 · similarity 0.88)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.88)
- ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation (ICLR 2025 · similarity 0.88)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.88)
- Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think (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.