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..
ReDi: Rectified Discrete Flow
Authors: Jaehoon Yoo, Wonjung Kim, Seunghoon Hong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluated our method on class conditional image generation and text generation. On image generation, Re Di shows comparable few-step generation performance against existing distillation methods, and significantly outperforms in one-step generation, due to direct rectification of the couplings contributing to the factorization error. On text generation, we observe that iteratively applying rectification improves the sampling efficiency and that Re Di can also be applied with existing distillation methods. 4 Experiments |
| Researcher Affiliation | Academia | Jaehoon Yoo KAIST EMAIL Wonjung Kim KAIST EMAIL Seunghoon Hong KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Coupling Update (Standard) 1: Input: p(X0), pθ, dataset size N 2: for i = 1 to N do 3: x(i) 0 p(x0) 4: x(i) 1 pθ(x1|x(i) 0 ) 5: end for 6: Return: {(x(i) 0 , x(i) 1 )}N i=1 |
| Open Source Code | Yes | Code is available at https://github.com/Ugness/Re Di_discrete. |
| Open Datasets | Yes | Datasets We conduct experiments on two benchmark datasets: Image Net [9] for class-conditional image generation and Open Web Text dataset [14] for text generation. |
| Dataset Splits | Yes | Datasets We conduct experiments on two benchmark datasets: Image Net [9] for class-conditional image generation and Open Web Text dataset [14] for text generation. For the image generation task... The model is trained for 13 epochs on the original Image Net dataset [9], with each epoch consisting of 1.3M images... For image generation, we generate 50k images and measure Fréchet Inception Distance (FID) [18], Inception Score (IS) [33] following the procedure of prior works [8, 28]. |
| Hardware Specification | Yes | The global batch size is set to 512, and the training take 96 GPU hours with A6000 GPUs. Each of Re Di variants take 16 GPU hours with A6000 GPUs. Each training takes 12 GPU hours with H100 GPUs. |
| Software Dependencies | No | No specific software versions like Python, PyTorch, or TensorFlow are mentioned, only specific models (VQGAN, GPT-2 tokenizer, LLaMa 3.1-8B) and optimizers (Adam W). |
| Experiment Setup | Yes | For finetuning Mask GIT [3, 8] with stochastic initial states, we use the Adam W optimizer with a learning rate of 1e-4 and a weight decay of 1e-5. The model is trained for 13 epochs on the original Image Net dataset [9], with each epoch consisting of 1.3M images. The global batch size is set to 512... For training the Re Di variants, we set the global batch size of 512 and applied a cosine learning rate scheduler with decay over 100 epochs... For Re Di1 and Re Di2, pairs for the coupling were generated with a CFG level of 1 and 16-step decoding. For the Re Di3-distill model, pairs were generated with a CFG level of 8 and 16-step decoding. |