Research map / Image generation / Image restoration
I$^2$SB: Image-to-Image Schrödinger Bridge
Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos Theodorou, Weili Nie, Anima Anandkumar
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 |
|---|---|---|---|
| Use Schrödinger bridge from degraded image to clean image (I2SB) instead of Gaussian-noise start (Palette-style conditional diffusion) | I2SB vs Palette · JPEG QF=5 and QF=10, inpainting; ImageNet 256 | FID: e.g. 4.6 vs 8.3 (QF=5) | Improved |
| Use I2SB bridge (no corruption operator knowledge) instead of inverse-problem DIM methods | I2SB vs DDRM/DDNM/PiGDM · nine restoration tasks, ImageNet 256 | FID | Mixed |
| Replace Schrödinger bridge baseline CDSB with I2SB | I2SB vs CDSB · all restoration tasks | FID | Improved |
| Use I2SB with DDPM-style sampling and very few steps (NFE 2-10) instead of Palette with noise prior | inpainting Freeform 20-30% · NFE reduction | FID | Improved |
| Deterministic OT-ODE sampling (remove stochasticity from posterior) instead of stochastic posterior sampling (ablation) | I2SB on JPEG restoration QF=5 · JPEG QF=5 | FID difference: +5.3 | Hurt |
| Deterministic OT-ODE sampling instead of stochastic posterior sampling (ablation) | I2SB on deblurring · Deblurring Uniform and Gaussian | FID difference: -0.3 (Uniform), -0.6 (Gaussian) | Improved |
| Deterministic OT-ODE sampling instead of stochastic posterior sampling (ablation) | I2SB on inpainting with noiseless masks · Inpainting Center and Freeform 20-30% | FID difference: +50.9 (Center), +13.0 (Freeform 20-30%) | Hurt |
| Inject Gaussian noise into masked regions of inputs when using OT-ODE (ablation) | I2SB OT-ODE on inpainting · Inpainting Center and Freeform 20-30% | FID difference: -0.1 (both) | No clear effect |
| Increase mixing of posterior q(Xt|X0,X1) vs forward-SDE marginal as sampling proposal (toward posterior) (ablation) | I2SB on JPEG restoration QF=5 · JPEG QF=5 | FID and CA | Improved |
| Use I2SB with a single network call (NFE=1) instead of Palette with NFE=1 (ablation) | inpainting Freeform 20-30% · NFE=1 | FID-10k: 6.7 vs 19.8 | Improved |
| Add conditioning on degraded input X1 to I2SB for general image-to-image translation | edges2shoes · Pix2pix tasks, NFE=1000 | FID: 37.8 vs Pix2pix 73.9 | Improved |
Later papers that compared with this method
7 later accepted papers report I2SB as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- Resolving Endpoint Underfitting in Diffusion Bridges via Noise Alignment (CVPR 2026 · better on 13 of 14 shared results)
- Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution (NeurIPS 2026 · better on 3 of 5 shared results)
- MixBridge: Heterogeneous Image-to-Image Backdoor Attack through Mixture of Schrödinger Bridges (ICML 2025 · better on 3 of 4 shared results)
- System-Embedded Diffusion Bridge Models (NeurIPS 2025 · better on 1 of 2 shared results)
- Feature out! Let Raw Image as Your Condition for Blind Face Restoration (ICML 2025 · better on 0 of 1 shared result)
- Stochastic Interpolants with Data-Dependent Couplings (ICML 2024 · better on 1 of 1 shared result)
- ResShift: Efficient Diffusion Model for Image Super-resolution by Residual Shifting (NeurIPS 2023 · better on 1 of 2 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
- Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network (CVPR 2025 · similarity 0.92)
- Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge (ICML 2024 · similarity 0.91)
- Unpaired Image-to-Image Translation via Neural Schrödinger Bridge (ICLR 2024 · similarity 0.90)
- Feature out! Let Raw Image as Your Condition for Blind Face Restoration (ICML 2025 · similarity 0.90)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.89)
- Diffusion Bridge Implicit Models (ICLR 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.
