Research map / Video / Video generation
VFRTok: Variable Frame Rates Video Tokenizer with Duration-Proportional Information Assumption
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 VFRTok, a query-based Transformer video tokenizer whose latent length scales with video duration (not frame count), trained with asymmetric encoder/decoder frame rates and Partial RoPE, achieving near-baseline reconstruction and better generation with 1/8 tokens.
Problem
Video latent diffusion tokenizers use fixed temporal compression, so the diffusion transformer's cost grows linearly with frame rate; high-FPS video needs proportionally more tokens.
Approach
A Duration-Proportional Information Assumption (information capacity scales with duration, not frame count) realized by a query-based ViT tokenizer with fixed-length latents, timestamp-based temporal RoPE, asymmetric encoder/decoder frame-rate training, and Partial RoPE (RoPE on only a subset of heads) to improve content-awareness.
Key results
- VFRTok-L (512 tokens) achieves reconstruction comparable to Cosmos-M and OmniTokenizer on K600 (PSNR 31.63 vs 31.66 and 29.35 respectively) with a quarter of Cosmos-M's latent capacity.
- VFRTok-L reaches the best gFVD on UCF101 and K600 while using about 8.4% of the compute of OmniTokenizer and Cosmos-L, with up to 21.6x faster convergence than OmniTokenizer.
- Asymmetric training improves 30→120 fps asymmetric reconstruction PSNR from 21.60 to 23.56 over the symmetric variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use asymmetric encoder-decoder frame-rate training instead of symmetric training (ablation) | VFRTok · Adobe240fps asymmetric reconstruction | PSNR: 22.36 -> 24.05; 21.60 -> 23.56; 23.33 -> 23.90 | Helped |
| Use asymmetric encoder-decoder frame-rate training instead of symmetric training (ablation) | VFRTok · Adobe240fps symmetric reconstruction | PSNR: 25.92 -> 25.87; 24.71 -> 24.53; 23.89 -> 23.95 | Mixed |
| Set Partial RoPE ratio to 0.5 rather than other tested ratios (ablation) | VFRTok · Video reconstruction and generation | quality | Helped |
| Set Partial RoPE ratio to 0, removing positional encoding (ablation) | VFRTok · Video reconstruction and generation | quality | Hurt |
| Replace the decoder's shared learnable patch token with a fixed-length set of independent tokens (ablation) | VFRTok · Reconstruction and variable-frame-rate decoding | reconstruction fidelity | Mixed |
| Use VFRTok rather than FLAVR for video frame interpolation | Video interpolation · UVG interpolation | PSNR: 26.05 vs 22.93; 34.22 vs 24.06; 22.17 vs 22.08; 27.59 vs 23.68; 21.70 vs 19.58 | Hurt |
Surprising findings
- Setting Partial RoPE ratio to 0 (no positional encoding in heads) makes the ViT unable to reconstruct or generate videos.
- VFRTok is substantially worse than FLAVR on video frame interpolation benchmarks despite being able to interpolate.
Limitations
- Dense attention limits scalability to long videos.
- Reconstruction PSNR lags LTX-VAE and video frame interpolation is weaker than dedicated FLAVR.
Experimental setup
Datasets: ImageNet-1K Kinetics-600 BVI-HFR UCF101 LAVIB
Benchmarks: K600 UCF101 LAVIB Adobe240fps UVG
Metrics: PSNR SSIM LPIPS rFVD gFVD TFLOPs
Baselines: OmniTokenizer Cosmos Tokenizer (L/M/S) LTX-VAE LARP FLAVR
Builds on: TiTok LightningDiT RoPE MAETok
Closest papers by meaning
- Efficient Long Video Tokenization via Coordinate-based Patch Reconstruction (CVPR 2025 · similarity 0.85)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.85)
- Language Model Beats Diffusion - Tokenizer is key to visual generation (ICLR 2024 · similarity 0.85)
- OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation (NeurIPS 2024 · similarity 0.84)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.84)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.84)
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.