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..
EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
Authors: Naga Sai Abhiram Kusumba, Maitreya Patel, Kyle Min, Changhoon Kim, Chitta Baral, Yezhou Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical results demonstrate that Erase Flow outperforms existing baselines and achieves an optimal trade-off between performance and prior preservation. 5 Experimental Results Here, we conduct a comprehensive evaluation of Erase Flow by extensively benchmarking it on various erasing tasks. |
| Researcher Affiliation | Collaboration | Abhiram Kusumba2 Maitreya Patel1* Kyle Min3 Changhoon Kim4 Chitta Baral1 Yezhou Yang1 1Arizona State University 2Capital One 3Oracle 4Soongsil University |
| Pseudocode | Yes | Algorithm 1 Erase Flow: Concept Erasure with Anchor-Trajectory Training. Zϕ: Flow partition function, pθ: denoising process, q: noising process, c : anchor prompt, c: target prompt, T: number of diffusion steps, STOP_SAMPLING: epoch at which anchor resampling stops. |
| Open Source Code | Yes | Our implementation of Erase Flow training script along with the evaluation pipeline, is publicly available at: https://github.com/Abhiramkns/Erase Flow |
| Open Datasets | Yes | For the nudity task, we use red-teaming prompts from multiple sources: 142 from I2P [53], 79 from Ring-a-Bell [58], 1000 from MMA-Diffusion [66], and 142 more from I2P extracted using UDAtk [71]. ... We also evaluate image quality using CLIP Score [21] (higher is better) and FID [22] (lower is better) on MSCOCO [36], and we report training time (in minutes) for each method. |
| Dataset Splits | No | For the nudity task, we use red-teaming prompts from multiple sources: 142 from I2P [53], 79 from Ring-a-Bell [58], 1000 from MMA-Diffusion [66], and 142 more from I2P extracted using UDAtk [71]. For artistic style erasure, we use 50 adversarial prompts per target style generated via UDAtk. Fine-grained erasure is evaluated using 10 diverse prompts per concept generated with GPT-4o, with 10 images per prompt, and scored using Gecko [31], inspired by Erase Bench [1]. ... we test Erase Flow and all baselines on 10,000 prompts from the MSCOCO [36] dataset |
| Hardware Specification | Yes | Erase Flow reaches comparable or superior performance with just 3 minutes of training on a single A100 GPU. |
| Software Dependencies | No | Optimization is carried out using the Adam optimizer with hyperparameters β = (0.9, 0.999) and ϵ = 10-8. For all experiments on Erase Flow , we fine-tune the SD v1.4 model using Lo RA, following the procedure described in [6]. Training is conducted with bfloat16 precision. |
| Experiment Setup | Yes | Erase Flow is implemented following Algorithm 1 and trained for 20 iterations, each using a single data batch. We directly set logβ in Eq. (8) to 2.5. The STOP_SAMPLING parameter is set to 21 for nudity erasure, 11 for fine-grained erasure, and 1 for artistic style erasure. A learning rate of 3.0 10-4 is used for nudity and fine-grained tasks, and 5.0 10-4 for artistic style erasure. |