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..
Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models
Authors: Byeonghu Na, Minsang Park, Gyuwon Sim, Donghyeok Shin, HeeSun Bae, Mina Kang, Se Jung Kwon, Wanmo Kang, Il-chul Moon
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through theoretical analysis and empirical results, we show that DATE maintains the generative capability of the model while providing superior text-image alignment over fixed text embeddings across various tasks, including multi-concept generation and text-guided image editing. Our code is available at https://github.com/aailab-kaist/DATE. ... Our theoretical and empirical results demonstrate that DATE improves text-image alignment while preserving the model s original generative capabilities. ... Experiments demonstrate that DATE consistently outperforms fixed embeddings across various tasks and methods involving text-to-image diffusion models... |
| Researcher Affiliation | Collaboration | 1KAIST, 2NAVER Cloud, 3summary.ai EMAIL, EMAIL , EMAIL |
| Pseudocode | Yes | Algorithm 1 Diffusion Sampling with DATE 1: x T p T ( ) 2: corg Iϕ(y) 3: for t = T to 1 do 4: if t {text update steps} then 5: c corg + ρ cht(xt,corg;y,θ) || cht(xt,corg;y,θ)||2 6: end if 7: xt 1 xt + 1 2βt(xt + sθ(xt, c, t)) 8: end for |
| Open Source Code | Yes | Our code is available at https://github.com/aailab-kaist/DATE. |
| Open Datasets | Yes | Following previous evaluation protocols [38, 39, 58], we generate 5,000 images from randomly sampled captions in the COCO [30] validation set. ... We evaluate DATE on the An E dataset [4] with 1) base SD and 2) SD with CONFORM [35]... We apply DATE to DDPM inversion [21], a diffusion-based image editing model; and evaluate it on 30 source images from Image Net-R-TI2I [54]... |
| Dataset Splits | Yes | Following previous evaluation protocols [38, 39, 58], we generate 5,000 images from randomly sampled captions in the COCO [30] validation set. ... For each prompt, we generate images using 64 random seeds and evaluate performance based on text-image similarity and text-text similarity in CLIP space. |
| Hardware Specification | Yes | We conducted most experiments on a single NVIDIA A100 GPU with CUDA 11.4, and some ablation studies were performed on a single Intel Gaudi v2 using Synapse AI 1.18.0. |
| Software Dependencies | Yes | We conducted most experiments on a single NVIDIA A100 GPU with CUDA 11.4, and some ablation studies were performed on a single Intel Gaudi v2 using Synapse AI 1.18.0. |
| Experiment Setup | Yes | Unless stated otherwise, we set the text-conditioned evaluation function h to CLIP score, the scale hyperparameter ρ to 0.5, and use the embedding from the previous update as the original text embedding corg for each update step. ... We use DDIM [51] with 50 sampling steps as the default sampler using classifier-free guidance [19], and experiments with the DDPM [18] sampler are provided in Appendix D.1. We set the guidance scale to 8 by default... |