| --- |
| license: apache-2.0 |
| pipeline_tag: text-to-image |
| --- |
| |
| # Transition Models: Rethinking the Generative Learning Objective |
|
|
| This repository contains the official implementation of **Transition Models (TiM)**, a novel generative model presented in the paper "[Transition Models: Rethinking the Generative Learning Objective](https://huggingface.co/papers/2509.04394)". |
|
|
| TiM addresses the dilemma in generative modeling by introducing an exact, continuous-time dynamics equation that analytically defines state transitions across any finite time interval. This enables a novel generative paradigm that adapts to arbitrary-step transitions, seamlessly traversing the generative trajectory from single leaps to fine-grained refinement with more steps. |
|
|
| For more detailed information, code, and usage instructions, please refer to the official [GitHub repository](https://github.com/WZDTHU/TiM). |
|
|
| ## Highlights |
|
|
| * **Arbitrary-Step Generation**: TiM learns to master arbitrary state-to-state transitions, unifying few-step and many-step regimes within a single, powerful model. This approach allows it to learn the entire solution manifold of the generative process. |
| * **State-of-the-Art Performance**: Despite having only 865M parameters, TiM achieves state-of-the-art performance, surpassing leading models such as SD3.5 (8B parameters) and FLUX.1 (12B parameters) across all evaluated step counts on the GenEval benchmark. |
| * **Monotonic Quality Improvement**: Unlike previous few-step generators, TiM demonstrates consistent quality improvement as the sampling budget increases. |
| * **High-Resolution Fidelity**: When employing its native-resolution strategy, TiM delivers exceptional fidelity at resolutions up to 4096x4096. |
|
|
| <p align="center"> |
| <img src="https://github.com/WZDTHU/TiM/raw/main/assets/illustration.png" width="800" alt="TiM Illustration"> |
| </p> |
|
|
| ## Model Zoo |
|
|
| A single TiM model can perform any-step generation (one-step, few-step, and multi-step) and demonstrate monotonic quality improvement as the sampling budget increases. |
|
|
| ### Text-to-Image Generation |
|
|
| | Model | Model Size | VAE | 1-NFE GenEval | 8-NFE GenEval | 128-NFE GenEval | |
| |---------|------------|------------------------------------------------------------------------|---------------|---------------|-----------------| |
| | TiM-T2I | 865M | [DC-AE](https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers) | 0.67 | 0.76 | 0.83 | |
|
|
| ### Class-guided Image Generation |
|
|
| | Model | Model Size | VAE | 2-NFE FID | 500-NFE FID | |
| |-----------|------------|------------------------------------------------------------------------|-----------|-------------| |
| | TiM-C2I-256 | 664M | [SD-VAE](https://huggingface.co/stabilityai/sd-vae-ft-ema) | 6.14 | 1.65 | |
| | TiM-C2I-512 | 664M | [DC-AE](https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers) | 4.79 | 1.69 | |
|
|
| ## Citation |
|
|
| If you find this project useful, please kindly cite: |
|
|
| ```bibtex |
| @article{wang2025transition, |
| title={Transition Models: Rethinking the Generative Learning Objective}, |
| author={Wang, Zidong and Zhang, Yiyuan and Yue, Xiaoyu and Yue, Xiangyu and Li, Yangguang and Ouyang, Wanli and Bai, Lei}, |
| year={2025}, |
| eprint={2509.04394}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.LG} |
| } |
| ``` |
|
|
| ## License |
| This project is licensed under the Apache-2.0 license. |
