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..
Rectified CFG++ for Flow Based Models
Authors: Shreshth Saini, Shashank Gupta, Alan Bovik
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on large-scale text-to-image models (Flux, Stable Diffusion 3/3.5, Lumina) show that Rectified-CFG++ consistently outperforms standard CFG on benchmark datasets such as MS-COCO, LAION-Aesthetic, and T2I-Comp Bench. ... In this section, we present a comprehensive empirical evaluation of Rectified-CFG++ for text-to-image (T2I) generation using large-scale models. Our experiments aim to rigorously demonstrate the effectiveness of our approach at improving text image alignment, color fidelity, and the preservation of fine details, generating high-quality samples while expending comparable inference costs as competing baseline methods. ... Table 1: Comprehensive Quantitative Evaluation of CFG against Rectified-CFG++ when both are integrated into leading T2I Models on MS-COCO 10K validation samples. |
| Researcher Affiliation | Academia | Shreshth Saini Shashank Gupta Alan C. Bovik The University of Texas at Austin EMAIL |
| Pseudocode | Yes | Algorithm 1 Rectified-CFG++ |
| Open Source Code | Yes | Project page: https://rectified-cfgpp.github.io/. ... Reproducibility: Apart from the pseudocode and implementation details provided in the paper, the source code is available on the project page: https://rectified-cfgpp.github.io/. The complete source code will be released upon acceptance. |
| Open Datasets | Yes | Extensive experiments on large-scale text-to-image models (Flux, Stable Diffusion 3/3.5, Lumina) show that Rectified-CFG++ consistently outperforms standard CFG on benchmark datasets such as MS-COCO, LAION-Aesthetic, and T2I-Comp Bench. ... Specifically, we used subsets of the MS-COCO dataset [20, 5] ... We also used a subset of 1,000 image-text pairs from LAION-Aesthetic [35] ... and the 1,000 prompts from Pick-A-Pic [18]. |
| Dataset Splits | Yes | Specifically, we used subsets of the MS-COCO dataset [20, 5], comprising 10,000 and 1,000 image-text pairs (referred to as MS-COCO 10K and MS-COCO 1K, respectively). We also used a subset of 1,000 image-text pairs from LAION-Aesthetic [35] (LAION-Aesthetic 1K) and the 1,000 prompts from Pick-A-Pic [18]. |
| Hardware Specification | Yes | All experiments were performed using a single NVIDIA A100 40GB GPU. |
| Software Dependencies | Yes | Code was written in Python 3.10, using Py Torch 2.0.1 and the latest Hugging Face Diffusers library. |
| Experiment Setup | Yes | When using our proposed method, Rectified-CFG++, we determined a set of effective hyperparameters which were kept consistent across all datasets and when integrated into baseline models. For all the compared methods, we utilized the default settings and configurations as reported in their original publications to ensure fair comparisons. Further detailed information regarding the experimental setup and hyperparameter settings can be found in Appendix D.1. |