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..
Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models
Authors: Hyeonggeun Han, Sehwan Kim, Hyungjun Joo, Sangwoo Hong, Jungwoo Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluations demonstrate that the proposed initial sample adjustment strategies produce more text-aligned, non-memorized outputs than those without adjustment, and outperform existing inference-time mitigation baselines in reducing memorization. ... We employ four complementary metrics to evaluate memorization, image-text alignment, image quality, and image diversity. |
| Researcher Affiliation | Collaboration | 1ECE & 2Next Quantum, Seoul National University 3CSE, Konkuk University 4Hodoo AI Labs EMAIL EMAIL |
| Pseudocode | Yes | The pseudocode for Per-sample mitigation is provided in Appendix A. ... A Pseudocode for Per-sample mitigation |
| Open Source Code | Yes | Code is available at https://github.com/hygnhan/init_noise_diffusion_memorization. |
| Open Datasets | Yes | For Stable Diffusion v1.4, we use 500 memorized prompts extracted from the LAION dataset [38], and for Stable Diffusion v2.0, we use 219 memorized prompts. These prompts are provided by Webster [47] and Ren et al. [34], respectively. |
| Dataset Splits | No | For each diffusion model, we generate 10 images per memorized prompt using identical inference configurations across all baselines and our proposed methods. |
| Hardware Specification | Yes | All experiments were conducted using an NVIDIA A100 GPU. |
| Software Dependencies | Yes | The implementation environment included Hugging Face 0.23.4, CUDA 12.0, Python 3.8.0, and Py Torch 2.3.1 with Torchvision 0.18.1. |
| Experiment Setup | Yes | Detailed configurations for the proposed mitigation strategies and baselines are outlined below. Batch-wise mitigation. For Stable Diffusion v1.4, the adjustment strength is set to γ = 0.7, the sharpness parameter to ρ = 50.0, and the number of adjustments to M = 2. CFG is applied with a guidance scale of 7.0 for timesteps t ≤ τ, and disabled (i.e., set to 0.0) for t > τ, where the CFG application start timestep is set to τ = 900. For Stable Diffusion v2.0, we set γ = 0.7, ρ = 30.0, M = 4, and τ = 917.4490. Per-sample mitigation. For Stable Diffusion v1.4, the initial noise sample is adjusted using the Adam W optimizer [30] with a learning rate of 0.01, while all other hyperparameters remain at their default values. The target threshold is varied as ltarget ∈ {0.7, 0.9, 1.1, 1.3, 1.5}. For Stable Diffusion v2.0, the initial noise sample is adjusted using the same optimizer with a learning rate of 0.1, and the target threshold is varied as ltarget ∈ {5, 8, 10, 15, 20}. |