Research map / Video / Video segmentation and tracking
IQ-VFI: Implicit Quadratic Motion Estimation for Video Frame Interpolation
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 IQ-VFI, a video frame interpolation framework that models quadratic (accelerated) motion by distilling latent acceleration priors and task-oriented intermediate flows from a teacher that sees the ground-truth frame into a student that sees only two input frames.
Problem
Video frame interpolation methods based on linear motion assumptions fail on complex curvilinear motions because they ignore latent acceleration between input frames, and supervising predicted intermediate motions with off-the-shelf optical flow pseudo-labels is sub-optimal.
Approach
An implicit acceleration estimation network (IANet) predicts a latent acceleration prior from two frames, which modulates linear flows from a two-stage VFI backbone into quadratic flows via an implicit motion modulation module; a teacher trained with the ground-truth frame supervises the student through implicit acceleration distillation and implicit motion distillation losses.
Key results
- IQ-VFI reaches 36.60 dB PSNR on Vimeo90K, outperforming AMT (36.53) and VFIFormer (36.50) among compared motion-based methods.
- On the ablation model, adding the teacher-student distillation raises Vimeo90K PSNR from 35.33 (no distillation, no IANet) to 35.52.
- Implicit motion distillation loss alone adds 0.10 dB and acceleration distillation alone adds 0.11 dB over reconstruction-only training in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the IQ-VFI quadratic-motion method | VFI · Vimeo90K | PSNR: 36.53 -> 36.60 | Helped |
| add the IQ-VFI quadratic-motion method | VFI · UCF101 | PSNR: 35.45 -> 35.48 | Helped |
| add the IQ-VFI quadratic-motion method | VFI · SNU-FILM Extreme | SSIM: 0.865 -> 0.863 | Hurt |
| add the IQ-VFI quadratic-motion method | VFI · Xiph 2K | PSNR: 36.74 -> 36.68 | Hurt |
| add the IQ-VFI quadratic-motion method | VFI · Xiph 4K | PSNR: 34.63 -> 34.72 | Helped |
| use a two-stage backbone rather than the linear scheme (ablation) | IQ-VFI · Vimeo90K | PSNR: 32.59 -> 35.35 | Helped |
| add task-oriented RefineNet to the linear scheme (ablation) | IQ-VFI · Vimeo90K | PSNR: 32.59 -> 34.98 | Helped |
| add IANet without knowledge distillation (ablation) | IQ-VFI · Vimeo90K | PSNR | No clear effect |
| add knowledge distillation to IANet and the two-stage backbone (ablation) | IQ-VFI · Vimeo90K | PSNR: 35.33 -> 35.52; 0.19dB | Helped |
| add implicit acceleration distillation loss LIA (ablation) | IQ-VFI student · Vimeo90K | PSNR: 35.33 -> 35.44; 0.11dB | Helped |
| add implicit motion distillation loss LIM (ablation) | IQ-VFI student · Vimeo90K | PSNR: 35.33 -> 35.43; 0.10dB | Helped |
| add both LIA and LIM losses (ablation) | IQ-VFI student · Vimeo90K | PSNR: 35.33 -> 35.52 | Helped |
Surprising findings
- Adding IANet without distillation gave no gain (35.33 vs 35.35 for two-stage), suggesting acceleration is hard to learn from two frames without supervision.
- Distilling off-the-shelf optical flow pseudo-labels from the ground-truth frame decreased performance by 0.09 dB, and a fine-tuned flow teacher overfit to ground truth.
Limitations
- Explicitly solving accurate acceleration from two frames remains difficult; the paper relies on distillation from a teacher that sees the ground-truth frame.
- Ablations use a reduced-channel model on Vimeo90K, and gains are small (about 0.1 dB per component).
Experimental setup
Datasets: Vimeo90K
Benchmarks: Vimeo90K UCF101 SNU-FILM Xiph
Metrics: PSNR SSIM
Baselines: ToFlow DAIN CAIN BMBC AdaCoF ABME RIFE M2M-VFI VFIFormer IFRNet EMA-VFI AMT
Builds on: AMT IFRNet RIFE PWC-Net Laplacian loss
Closest papers by meaning
- BiM-VFI: Bidirectional Motion Field-Guided Frame Interpolation for Video with Non-uniform Motions (CVPR 2025 · similarity 0.89)
- Generalizable Implicit Motion Modeling for Video Frame Interpolation (NeurIPS 2024 · similarity 0.87)
- Disentangled Motion Modeling for Video Frame Interpolation (AAAI 2025 · similarity 0.87)
- VFIMamba: Video Frame Interpolation with State Space Models (NeurIPS 2024 · similarity 0.86)
- Sparse Global Matching for Video Frame Interpolation with Large Motion (CVPR 2024 · similarity 0.86)
- Repurposing Pre-trained Video Diffusion Models for Event-based Video Interpolation (CVPR 2025 · similarity 0.85)
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.