Research map / Video / Video generation
VISION-XL: High Definition Video Inverse Problem Solver using Latent Image Diffusion Models
Taesung Kwon, Jong Chul Ye
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize with pseudo-batch inversion of the measurement instead of batch-consistent noise (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 707.7 -> 184.8 | Improved |
| initialize with pseudo-batch inversion instead of random noise (ablation) | VISION-XL · Pexels landscape SR+ | PSNR: 29.43 -> 30.74 | Improved |
| use pseudo-batch inversion to cut sampling time (ablation) | VISION-XL · Pexels landscape SR+ | sampling time (min): 8.3 -> 2.5 | Improved |
| raise inversion timestep from 0.15T to 0.30T (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 229.5 -> 184.8 | Improved |
| raise inversion timestep from 0.30T to 0.45T (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 184.8 -> 288.7 | Hurt |
| increase CG update steps from 1 to 10 (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 1150 -> 184.8 | Improved |
| increase CG update steps from 10 to 20 (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 184.8 -> 486.1 | Hurt |
| add scheduled low-pass filtering before re-encoding (ablation) | VISION-XL · Pexels landscape SR+ | FVD: 209.7 -> 184.8 | Improved |
| use latent SDXL pseudo-batch method instead of pixel-space SVI for pixel-wise random inpainting | VISION-XL · DAVIS Inpaint+ | PSNR: 29.60 -> 28.81 | Hurt |
| use VISION-XL instead of SVI for deblurring with frame averaging | VISION-XL · DAVIS Deblur+ | FVD: 638.9 -> 228.6 | Improved |
| use pseudo-batch sampling instead of 16-frame batch sampling | VISION-XL · 25-frame 768x1280 video | memory (GB): 18.5 -> 12.7 | Improved |
Later papers that compared with this method
4 later accepted papers report VISION-XL as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- LVTINO: LAtent Video consisTency INverse sOlver for High Definition Video Restoration (ICLR 2026 · better on 13 of 18 shared results)
- InstantViR: Real-Time Video Inverse Problem Solver with Distilled Diffusion Prior (CVPR 2026 · better on 8 of 9 shared results)
- Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models (NeurIPS 2026 · better on 5 of 6 shared results)
- Temporal-Consistent Video Restoration with Pre-trained Diffusion Models (AAAI 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Solving Video Inverse Problems Using Image Diffusion Models (ICLR 2025 · similarity 0.92)
- Warped Diffusion: Solving Video Inverse Problems with Image Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Prompt-tuning Latent Diffusion Models for Inverse Problems (ICML 2024 · similarity 0.88)
- Solving Inverse Problems with Latent Diffusion Models via Hard Data Consistency (ICLR 2024 · similarity 0.88)
- Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (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.
