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..
AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models
Authors: Seunghoon Lee, Jeongwoo Choi, Byunggwan Son, JaeHyeon Moon, Jeimin Jeon, Bumsub Ham
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the efficacy and efficiency of Accu Quant across various tasks and diffusion models on standard benchmarks. ... We show in Tables 1-3 quantitative comparisons of our method and the state of the art [25, 19, 54] for unconditional image generation (i.e., CIFAR-10 [23], LSUN-Bedrooms, and LSUN-Churches [56]), class-conditional image generation (i.e., Image Net [6]) and text-to-image generation (i.e. MS-COCO [28]), respectively. |
| Researcher Affiliation | Academia | Seunghoon Lee Jeongwoo Choi Byunggwan Son Jaehyeon Moon Jeimin Jeon Bumsub Ham School of Electrical and Electronic Engineering, Yonsei University |
| Pseudocode | Yes | A Detailed algorithm for Accu Quant Algorithm 1 Pseudo code of Accu Quant. |
| Open Source Code | No | Answer: [No] Justification: We do not release our code due to copyright restrictions. However, we include detailed pseudocode of our algorithm, dataset descriptions, and experimental guidelines in Sec. 4, enabling straightforward reproduction. |
| Open Datasets | Yes | Datasets and models. We apply Accu Quant to various diffusion models and perform extensive experiments on standard benchmarks for unconditional, class-conditional, and text-to-image generation tasks. For the unconditional generation task, we exploit DDIM [49] on CIFAR-10 [23], and Latent Diffusion Model (LDM) [41] on LSUN-Bedrooms and LSUN-Churches [56]. For class-conditional generation, we perform experiments using LDM [41] on Image Net [6]. We use Stable Diffusion (SD) v1.4 [41] on MS-COCO [28] for text-to-image generation. |
| Dataset Splits | Yes | Following the work of [25], we generate 256 calibration samples for each group with full-precision models, maintaining the total number of the samples consistent with that of [25] across all experiments. |
| Hardware Specification | Yes | We also retain the default settings [41] during the sampling phase, ensuring that we can conduct all experiments on a single A100 (80 GB) GPU. ... Since the official PyTorch quantization API does not support bit-widths lower than 8, we quantize both weights and activations to 8 bits and measure the memory usage and runtime latency using ONNX Runtime with the Intel Xeon Gold 6226R CPU. |
| Software Dependencies | No | The paper mentions 'PyTorch quantization API', 'ONNX Runtime', 'torch-fidelity library', and 'official Guided Diffusion [7] codebase' but does not specify version numbers for any of these software components. |
| Experiment Setup | Yes | We employ adaptive rounding [37, 26] for weight quantizers, following prior approaches [45, 25, 19]. For activation quantization, we split a denoising process into 20 groups for unconditional image generation, 10 groups for class-conditional image generation and 25 groups for text-to-image generation. We perform the calibration process for 50, 20, and 10 epochs on DDIM [49], LDM, and SD [41], respectively, using the Adam optimizer [22]. ... For the CIFAR-10 [23], ... setting learning rate in {1 10 3, 4 10 4}. For the LSUN-Bedroom [56], ... setting learning rate into 4 10 5. For the LSUN-Church [56], ... setting learning rate into 4 10 5. For the Image Net [6], ... setting learning rate into 1 10 3. For text-to-image generation, ... setting learning rate into 1 10 5. We set the calibration batch size to 8 for DDIM [49] and LDMs [41], and to 1 for Stable Diffusion [41]. |