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..
Aligning Text to Image in Diffusion Models is Easier Than You Think
Authors: Jaa-Yeon Lee, ByungHee Cha, Jeongsol Kim, Jong Chul Ye
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results across text-to-image generation and text-guided image editing tasks validate the effectiveness of our approach in improving the semantic consistency of T2I generative models. Project Page: https://softrepa.github.io/. 1 Introduction ... We conducted text-to-image generation experiments on SD1.5, SDXL, and SD3, training soft tokens using the text-image paired COCO dataset [23]. ... We conducted an ablation study on SD3 using the COCO val dataset [23] to evaluate the effects of the number of layers using soft text tokens and token length on generation quality (fig. 6). |
| Researcher Affiliation | Academia | Jaa-Yeon Lee 1 Byunghee Cha 1 Jeongsol Kim2 Jong Chul Ye1 1Kim Jaechul Graduate School of AI, KAIST 2Department of Bio and Brain Engineering, KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Image Generation with Soft Tokens in MM-Di T Require: Gaussian noise z N(0, I) Require: Text Y pdata Require: Soft token s Embedding(k, t) Require: Number of layers N, Threshold layer L Require: Time steps {t T , t T 1, ..., t0} 1: Initialize H(0,T ) img z 2: Initialize H(0,T ) text Text Encoder(Y ) 3: n = |H(0,T ) text | 4: for t in {t T , t T 1, ..., t0} do 5: for l = 1 to N do 6: if k L then 7: s(k,t) Embedding(k, t) 8: ˆ H(k 1,t) text [s(k,t); H(k 1,t) text ] 9: else 10: ˆ H(k 1,t) text H(k 1,t) text 11: end if 12: H(k,t) img , ˆ H(k,t) text Layerl(H(k 1,t) img , ˆ H(k 1,t) text ) 13: H(k,t) text ˆ H(k,t) text [ n :, :] Drop soft tokens 14: end for 15: end for 16: return ˆX = Decoder(H(N,t0) img ) |
| Open Source Code | Yes | Project Page: https://softrepa.github.io/. ... The code is publicly available at https://github.com/softrepa/Soft REPA. |
| Open Datasets | Yes | We conducted text-to-image generation experiments on SD1.5, SDXL, and SD3, training soft tokens using the text-image paired COCO dataset [23]. The implementation details are provided in appendix B. To evaluate the generated images, we assessed human preference scores [45, 17], text-image alignment [48, 44], and image quality [47, 34] on COCO-val 5K dataset. Furthermore, we conducted evaluation on Gen Eval [9], compared with Rank DPO [16]... For evaluation, we used PIEBench [14], DIV2K [1], and Cat2Dog datasets. |
| Dataset Splits | Yes | To evaluate the generated images, we assessed human preference scores [45, 17], text-image alignment [48, 44], and image quality [47, 34] on COCO-val 5K dataset. ... For DIV2K editing dataset, we selected 800 high-quality images and generated detailed source and target captions using LLa VA [25]; dataset construction details are provided in appendix C. |
| Hardware Specification | Yes | For the training of softs token, the batch size was set to 16 with less than 30,000 iterations using two A100 GPUs. ... For the memory efficiency metrics reported in table 1, values were measured by averaging results over 50 runs with an A100 GPU. |
| Software Dependencies | No | Specifically, we employ a mean squared error (MSE) loss between the predicted object count, which is derived from denoised images, and groundtruth labels using the lightweight object detection module, YOLOv8 [40]. |
| Experiment Setup | Yes | For the training of softs token, the batch size was set to 16 with less than 30,000 iterations using two A100 GPUs. For Stable Diffusion 3, we set the length of the soft token to 4 and used 5 layers to attach the soft tokens for further experiments. ... Table 5: The implementation details for training. Models lr wd batch size (positive, negative) iterations token init optimizer lr scheduler SD1.5 1e-3 1e-4 32(4, 28) 26,000 Adam W Cosine Annealing Warm Restarts SDXL 1e-3 1e-4 16(1, 15) 30,000 N(0, 0.02) Adam W Cosine Annealing Warm Restarts SD3 1e-3 1e-4 16(4, 12) 30,000 N(0, 0.02) Adam W Cosine Annealing Warm Restarts |