Research map / Video / Video generation
ViBiDSampler: Enhancing Video Interpolation Using Bidirectional Diffusion Sampler
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 ViBiDSampler, a training-free bidirectional sequential sampler over Stable Video Diffusion (SVD-XT) that interpolates between two keyframes, augmented with DDS last-frame guidance and CFG++.
Problem
Keyframe (start/end-frame) video interpolation with image-to-video diffusion models, where existing time-reversal fusion methods suffer off-manifold artifacts and need multiple re-noising or fine-tuning.
Approach
Replace parallel forward/backward denoising with sequential sampling: denoise along the forward path conditioned on the start frame, re-noise a single step, then denoise the time-flipped path conditioned on the end frame, combined with DDS last-frame guidance and CFG++ so samples stay on the diffusion manifold without fine-tuning.
Key results
- On DAVIS, Ours (Full) reaches LPIPS 0.2355, FID 35.659, FVD 399.15, beating the best baseline Generative Inbetweening (FID 36.273, FVD 490.34).
- The vanilla bidirectional sampler (FVD 717.37 on Pexels) outperforms TRF (FVD 880.97) without any guidance, supporting the sequential-sampling claim.
- Interpolating 25 frames at 1024x576 takes 195 s on a single RTX 3090 with 50 NFE, versus 443 s for TRF and 1,222 s for Generative Inbetweening.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use bidirectional sampling instead of TRF's fusion strategy | ViBiDSampler · DAVIS | FVD: 543.31 vs 622.16 | Helped |
| Use bidirectional sampling instead of TRF's fusion strategy | ViBiDSampler · Pexels | FVD: 717.37 vs 880.97 | Helped |
| Add CFG++ to vanilla bidirectional sampling, without DDS (ablation) | ViBiDSampler · DAVIS | LPIPS: 0.3031 -> 0.2571 | Helped |
| Add CFG++ to vanilla bidirectional sampling, without DDS (ablation) | ViBiDSampler · Pexels | FVD: 717.37 -> 478.35 | Helped |
| Add DDS to bidirectional sampling with CFG++ (ablation) | ViBiDSampler · DAVIS | FVD: 434.41 -> 399.15 | Helped |
| Add DDS to bidirectional sampling with CFG++ (ablation) | ViBiDSampler · Pexels | FVD: 478.35 -> 452.34 | Helped |
| Increase CFG++ guidance scale from 0.6 to 1.0 (ablation) | ViBiDSampler · DAVIS | FID: 52.5059 -> 35.6594 | Helped |
| Increase CFG++ guidance scale from 0.6 to 1.0 (ablation) | ViBiDSampler · DAVIS | LPIPS: 525.36 -> 399.15 | Helped |
| Use SVD-XT instead of SVD as the base video model (ablation) | ViBiDSampler · Pexels | FVD: 608.07 -> 452.34 | Helped |
| Add CFG++ and DDS guidance (ablation) | Generative Inbetweening · Pexels | FVD: 746.26 -> 747.95 | No clear effect |
Surprising findings
- FILM achieves the highest PSNR and SSIM on both datasets, outperforming all diffusion-based methods, which the authors attribute to its PSNR-oriented L1 training.
- On Pexels, FILM also has the best LPIPS and FID, so the proposed method does not win on every metric.
Limitations
- Built only on SVD with image conditioning, so text conditioning is not supported.
- FILM still achieves best PSNR/SSIM and some perceptual metrics on Pexels, so the method is not uniformly best.
- Decoding only 4 frames at a time due to memory limits on a 3090 GPU.
Experimental setup
Datasets: DAVIS Pexels
Benchmarks: DAVIS Pexels
Metrics: LPIPS FID FVD PSNR SSIM inference time
Baselines: FILM TRF DynamiCrafter Generative Inbetweening
Builds on: Time Reversal Fusion Stable Video Diffusion CFG++ DDS EDM
Closest papers by meaning
- Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation (ICLR 2025 · similarity 0.91)
- Video Interpolation with Diffusion Models (CVPR 2024 · similarity 0.90)
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.90)
- IV-mixed Sampler: Leveraging Image Diffusion Models for Enhanced Video Synthesis (ICLR 2025 · similarity 0.90)
- AsymRnR: Video Diffusion Transformers Acceleration with Asymmetric Reduction and Restoration (ICML 2025 · similarity 0.90)
- VideoGuide: Improving Video Diffusion Models without Training Through a Teacher's Guide (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.