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..

RespoDiff: Dual-Module Bottleneck Transformation for Responsible & Faithful T2I Generation

Authors: Silpa Vadakkeeveetil Sreelatha, Sauradip Nag, Muhammad Awais, Serge Belongie, Anjan Dutta

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate the effectiveness of our approach for responsible T2I generation, with a focus on fairness and safety. Respo Diff surpasses existing fair-generation baselines and effectively generalizes to unseen prompts, including profession-specific scenarios, without requiring any profession-specific training or fine-tuning. Our approach further ensures semantic alignment with prompts while maintaining the visual quality of diffusion models.
Researcher Affiliation Academia Silpa Vadakkeeveetil Sreelatha University of Surrey Sauradip Nag Simon Fraser University Muhammad Awais University of Surrey Serge Belongie University of Copenhagen Anjan Dutta University of Surrey
Pseudocode Yes We provide a pseudocode for the training and inference of Respo Diff in Algorithm 1 and Algorithm 2 respectively. We also provide an inference diagram in Fig. 5.
Open Source Code No The project page is available at https://vssilpa.github.io/respodiff_project_page. (This is a project page, not a direct code repository link, so it does not meet the criteria for "Yes" according to the instructions.)
Open Datasets Yes We evaluate our method on the Winobias benchmark (Zhao et al., 2018), following the approaches in (Gandikota et al., 2024; Li et al., 2024; Orgad et al., 2023)...
Dataset Splits Yes We evaluate our method on the Winobias benchmark (Zhao et al., 2018)... Five prompts per profession are used, including templates like A photo of a profession . For each prompt, we generate 5 images, resulting in 30 images per profession. In total, we evaluate on 5400 images.
Hardware Specification Yes We conduct all our training and inference experiments for Stable Diffusion v1.4 on a single NVIDIA RTX 3090 with 24 GB of VRAM while we use single A100 with 80GB memory for experiments with SDXL .
Software Dependencies No The paper mentions using Stable Diffusion v1.4 and SDXL models, but does not provide specific version numbers for software dependencies like programming languages or libraries (e.g., Python, PyTorch, CUDA).
Experiment Setup Yes We learn transformations as constant functions linearly added to bottleneck activations, optimizing over 5000 iterations with batch size 1, as in Section 4.2.