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..
Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models
Authors: Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, Yi-Zhe Song
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experimentation, we demonstrate consistent improvements in text alignment (CLIP Score), fidelity (FID, PFID), and human-perceived quality (Image Reward). Our ablation studies validate each design component, while user studies confirm significant preference for NAG-guided outputs. |
| Researcher Affiliation | Collaboration | Dar-Yen Chen1,2 Hmrishav Bandyopadhyay1 Kai Zou2 Yi-Zhe Song1 1Sketch X, CVSSP, University of Surrey 2Net Mind.AI EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 provides PyTorch-style pseudocode for integrating NAG into cross-attention layers. |
| Open Source Code | Yes | We ensure open access by releasing the full code on our project website. |
| Open Datasets | Yes | We evaluate NAG on the COCO-5K dataset [61] using CLIP Score [16], Fréchet Inception Distance (FID) [17], Patch FID (PFID) [7], and Image Reward [18] |
| Dataset Splits | No | We evaluate NAG on the COCO-5K dataset [61] using CLIP Score [16], Fréchet Inception Distance (FID) [17], Patch FID (PFID) [7], and Image Reward [18], which reflects human aesthetic preference. Following NASA [13], we use Low resolution, blurry as a universal negative prompt for quantitative comparison. The paper mentions using the COCO-5K dataset but does not specify any training/test/validation splits, percentages, or methodology for splitting. |
| Hardware Specification | Yes | Unless otherwise specified, experiments are conducted on Flux-Schnell [1] with 4-step sampling, using an NVIDIA A100 GPU. |
| Software Dependencies | No | Algorithm 1 provides PyTorch-style pseudocode for integrating NAG into cross-attention layers. While this implies the use of PyTorch, no specific version number for PyTorch or any other software dependency is provided. |
| Experiment Setup | Yes | We provide the default NAG hyperparameters for different model families in Table 5. Algorithm 1 provides PyTorch-style pseudocode for integrating NAG into cross-attention layers. |