Research map / Image generation / Fast diffusion sampling
Bidirectional Noise Injection: Enhancing Diffusion Models via Coordinated Input-Output Perturbation
Tianyi Zheng, Jiayang Gao, Peng-Tao Jiang, Fengxiang Yang, Ben Wan, Hao Zhang, Jinwei Chen, Jia Wang, Bo Li
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add synchronized input-output noise injection (CIOP) to DDPM training | ADM/DDPM · CIFAR 32x32 | FID: 3.47 -> 2.46 | Improved |
| add CIOP | ADM/DDPM · FFHQ 128x128 | FID: 14.07 -> 6.65 | Improved |
| add CIOP to latent text-to-image model | latent diffusion · MS-COCO 256x256 | FID: 5.95 -> 5.74 | Improved |
| add CIOP to DiT | DiT-L/2 · ImageNet-256 class-conditional | FID: 11.78 -> 10.26 | Improved |
| apply noise injection every iteration (NaiveCIOP) instead of stochastically (ablation) | DiT-S/2 · ImageNet-256 class-conditional | FID: 49.23 -> 49.98 | Hurt |
| apply stochastic injection (p<1) instead of always injecting (ablation) | CIOP · CelebAHQ latent | FID: 8.93 -> 6.60 | Improved |
| combine CIOP with Min-SNR, P2-Weight, Debias or DDPM-BS | existing improvement methods · FID comparison (Figure 2) | FID | Improved |
| add CIOP to EDM | EDM · AFHQ-D 256x256 | FID: 15.42 -> 13.84 | Improved |
| add CIOP to EDM on CIFAR | EDM · CIFAR 32x32 | FID: 2.70 -> 2.59 | Improved |
| add CIOP to flow matching SiT | SiT · CelebAHQ SDE sampler-150 | FID: 6.93 -> 6.31 | Improved |
| increase injected noise variance from 0.10 to 0.20 (ablation) | CIOP · AFHQ-D 256x256 | FID: 13.02 -> 19.22 | Hurt |
| use uniform instead of Gaussian injected noise (ablation) | CIOP · AFHQ-D 256x256 | FID: 13.02 -> 14.46 | Hurt |
Closest papers by meaning
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.88)
- NoiseCtrl: A Sampling-Algorithm-Agnostic Conditional Generation Method for Diffusion Models (CVPR 2025 · similarity 0.88)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.88)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.88)
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.
