Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Exploring the Design Space of Diffusion Bridge Models
Authors: Shaorong Zhang, Yuanbin Cheng, Greg Ver Steeg
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate our model s state-of-the-art performance in both image quality and sampling speed across diverse I2I tasks, including deblurring, edges-to-handbags translation, and depth-to-RGB conversion. Notably, for handbag generation, our approach yields significantly more diverse outputs with varied colors and textures. |
| Researcher Affiliation | Academia | Shaorong Zhang, Yuanbin Cheng, Greg Ver Steeg Unversity of California Riverside EMAIL |
| Pseudocode | Yes | Algorithm 1 ECSI Sampler |
| Open Source Code | Yes | Code is available at https://github.com/szhan311/ECSI. |
| Open Datasets | Yes | We evaluate on I2I translation tasks on Edges Handbags [16] scaled to 64 64 pixels and DIODE-Outdoor scaled to 256 256 [37], and Deblurring on Image Net dataset [9]. |
| Dataset Splits | No | We evaluate on I2I translation tasks on Edges Handbags [16] scaled to 64 64 pixels and DIODE-Outdoor scaled to 256 256 [37], and Deblurring on Image Net dataset [9]. ... Deblurring on Image Net Dataset. We evaluate our models for Gaussian deblurring applying a Gaussian kernel with σ = 10 and Uniform deblurring, shown in Table 5. The results demonstrates that our ECSI models achieve much lower FID score. Table 5: Deblurring results with respect to different kernels, evaluated by FID on the 10k Image Net (256 256) validation subset. Our results are achieved by 20 NFEs. |
| Hardware Specification | Yes | Table 9: Training settings ... GPU 1 A6000 48G 1 H100 96G 1 H100 96G ... GPU 1 H100 96G 1 H100 96G |
| Software Dependencies | No | Architecture. We maintain the architecture and parameter settings consistent with [18], utilizing the ADM model [10] for 64 64 resolution, modifying the channel dimensions from 192 to 256 and reducing the number of residual blocks from three to two. |
| Experiment Setup | Yes | Table 9: Training settings ... Batch size 32 128 200 ... Learning rate 1 10 5 5 10 5 1 10 4 epochs 2078 2106 1443 ... Batch size 16 16 ... Learning rate 2 10 5 2 10 5 epochs 2617 1745 |