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..
Precise Diffusion Inversion: Towards Novel Samples and Few-Step Models
Authors: Jing Zuo, Luoping Cui, Chuang Zhu, Yonggang Qi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on COCO 2017, Dark Face, and a stylized cartoon dataset show that Precise Inv achieves state-of-the-art performance in both reconstruction quality and inference speed. |
| Researcher Affiliation | Academia | Jing Zuo Luoping Cui Chuang Zhu Yonggang Qi # School of Artificial Intelligence, Beijing University of Posts and Telecommunications EMAIL, EMAIL # Corresponding author |
| Pseudocode | Yes | Algorithm 1 Precise Inv for Diffusion Models with DDIM Sampler 1: Input: Real image x0, diffusion model ϵθ, convergence threshold η, number of inference steps T 2: Output: Inverted latent x T 3: for t = 1 to T do 4: Initialize ϵ t N(0, I) 5: if t > 1 then 6: x t 1 αt 1 x0 + 1 αt 1 ϵ t 1 7: else 8: x t 1 x0 9: end if 10: // Apply a single DDIM Inversion step 11: ϵ t 1 1 αt µ(x t 1) αtx0 (µ defined in Eq. (16)) 12: x t αt x0 + 1 αt ϵ t 13: Lrec µ(x t ) x t 1 2 14: while Lrec < η do 15: x t αt x0 + 1 αt ϵ t 16: Lrec µ(x t ) x t 1 2 17: ϵ t ϵ t Lrec 18: end while 19: end for 20: return x T |
| Open Source Code | Yes | Code is available at https://github. com/panda7777777/Precise Inv |
| Open Datasets | Yes | Datasets. We evaluate Precise Inv on three domains: LAION-aligned, low-light, and stylized. (i) The COCO 2017 [24] validation set contains 5,000 natural images with diverse everyday scenes. It serves as a close proxy to the LAION distribution used in training most text-to-image diffusion models. (ii) The Dark Face [41] validation set includes 6,089 nighttime images captured in real-world low-light conditions, exhibiting extreme visibility degradation. |
| Dataset Splits | Yes | Datasets. We evaluate Precise Inv on three domains: LAION-aligned, low-light, and stylized. (i) The COCO 2017 [24] validation set contains 5,000 natural images with diverse everyday scenes. It serves as a close proxy to the LAION distribution used in training most text-to-image diffusion models. |
| Hardware Specification | Yes | We run all experiments on a single RTX 4090 24GB GPU, except for SD3, which requires a single A100 40GB GPU due to its higher memory and compute demands. |
| Software Dependencies | No | We implement our method using Py Torch and the Hugging Face diffusers library. |
| Experiment Setup | Yes | We use the Adam W optimizer with β1 = 0.9, β2 = 0.999, and a weight decay of 0.01. The learning rate is set to 0.1 for SD v1.4, 0.05 for both LCM-SD v1.5 and SDXL, and 0.025 for SD3. We use the DDIM sampler unless otherwise specified; for SD3, the Euler discrete sampler is adopted due to architectural compatibility. We keep the number of inversion steps equal to the number of sampling steps, denoted as T. All experiments are conducted under mixed precision: float16 is used for SD v1.4, LCM-SD v1.5, and SDXL, while bfloat16 is used for SD3 to ensure numerical stability. |