TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows
Generation Speed Comparison (1328×1328)
Abstract
Recent advances in large multi-modal generative models have demonstrated impressive capabilities in multi-modal generation, including image and video generation. These models are typically built upon multi-step frameworks like diffusion and flow matching, which inherently limits their inference efficiency, requiring 40-100 Number of Function Evaluations (NFEs). While various few-step methods aim to accelerate the inference, existing solutions have clear limitations. Prominent distillation-based methods, such as progressive and consistency distillation, either require an iterative distillation procedure or show significant degradation at very few steps (< 4-NFE). Meanwhile, integrating adversarial training into distillation (e.g., DMD/DMD2 and SANA-Sprint) to enhance performance introduces training instability, added complexity, and high GPU memory overhead due to the auxiliary trained models. To this end, we propose TwinFlow, a simple yet effective framework for training 1-step generative models that bypasses the need for fixed pretrained teacher models and avoids standard adversarial networks during training, making it ideal for building large-scale, efficient models. On text-to-image tasks, our method achieves a GenEval score of 0.83 in 1-NFE, outperforming strong baselines like SANA-Sprint (a GAN loss-based framework) and RCGM (a consistency-based framework). Notably, we demonstrate the scalability of TwinFlow by full-parameter training on Qwen-Image-20B and transform it into an efficient few-step generator. With just 1-NFE, our approach matches the performance of the original 100-NFE model on both the GenEval and DPG-Bench benchmarks, reducing computational cost by 100× with minor quality degradation.
News
- We release TwinFlow-Qwen-Image-v1.0! And we are also working on Z-Image-Turbo to make it more faster!
Overview
The Trilemma of Few-Step Generation
Recent large multi-modal models have achieved stunning generative capabilities, but they come at a steep cost: inference efficiency. Standard diffusion and flow-matching models typically require 50-100 NFEs to generate a single image.
While researchers have raced to solve this with few-step distillation, existing solutions force a compromise between complexity, stability, and quality. As shown below, current SoTA methods rely heavily on "baggage"—auxiliary discriminators/fake scores or frozen teacher models, which inflate memory costs and destabilize training.
| Method | Generation Type | Requires Auxiliary Trained Model? (e.g., Discriminator, fake score) | Requires Frozen Teacher Model? | Major Drawback |
|---|---|---|---|---|
| GANs | 1-step | 1 | 0 | Unstable training dynamics |
| Consistency Models | 1-step / Few-step | 0 | 1 | Quality degradation at < 4 steps |
| DMD / DMD2 | 1-step / Few-step | 1-2 | 1 | High GPU Memory, Complex pipeline |
| TwinFlow (Ours) | 1-step / Few-step | 0 | 0 | - |
TwinFlow: Simplicity via Self-Adversarial Flows
We introduce TwinFlow, a framework that realizes high-quality 1-step and few-step generation without the pipeline bloat.
Instead of relying on external discriminators or frozen teachers, TwinFlow creates an internal "twin trajectory". By extending the time interval to , we utilize the negative time branch to map noise to "fake" data, creating a self-adversarial signal directly within the model.
Then, the model can rectify itself by minimizing the difference of the velocity fields between real trajectory and fake trajectory, i.e. the . The rectification performs distribution matching as velocity matching, which gradually transforms the model into a 1-step/few-step generator.
Key Advantages:
- One-model Simplicity. We eliminate the need for any auxiliary networks. The model learns to rectify its own flow field, acting as the generator, fake/real score. No extra GPU memory is wasted on frozen teachers or discriminators during training.
- Scalability on Large Models. TwinFlow is easy to scale on 20B full-parameter training due to the one-model simplicity. In contrast, methods like VSD, SiD, and DMD/DMD2 require maintaining three separate models for distillation, which not only significantly increases memory consumption—often leading OOM, but also introduces substantial complexity when scaling to large-scale training regimes.
Scalability: Unlocking Full-Parameter Training on Qwen-Image-20B
The true power of TwinFlow's "One-model" efficiency is demonstrated by its scalability. Prior methods like VSD/SiD/DMD, DMD2 is hard to scale to massive models because loading multiple model (Fake score + Teacher (Real score) + Generator, Discriminator (Optional)) causes OOM.
TwinFlow is the first framework to successfully enable full-parameter 1-step/few-step training on the massive Qwen-Image-20B model, reducing inference cost by nearly 100× while maintaining strong generation quality.
| Method | Scale | Memory Tradeoffs | NFE ⬇️ | GenEval ⬆️ | DPG-Bench ⬆️ | WISE ⬆️ |
|---|---|---|---|---|---|---|
| Qwen-Image (Original) | 20B | - | 50×2 | 0.87 | 88.32 | 0.62 |
| VSD / DMD / SiD | Generator (20B), Teacher (20B), Fake score (20B) | OOM (>80GB) | - | - | - | - |
| VSD | Generator (20B), Teacher (20B), Fake score (LoRA, ~420M) | Fake score as LoRA & Small batch size | 1 | 0.67 | 84.44 | 0.22 |
| SiD | Generator (20B), Teacher (20B), Fake score (LoRA, ~420M) | Fake score as LoRA & Small batch size | 1 | 0.77 | 87.05 | 0.42 |
| DMD | Generator (20B), Teacher (20B), Fake score (LoRA, ~420M) | Fake score as LoRA & Small batch size | 1 | 0.81 | 84.31 | 0.47 |
| sCM (JVP-free) | 20B | JVP through finite differences / Special JVP kernels | 8 | 0.60 | 85.54 | 0.45 |
| MeanFlow (JVP-free) | 20B | JVP through finite differences / Special JVP kernels | 8 | 0.49 | 83.81 | 0.37 |
| TwinFlow | 20B | No tradeoffs | 1 | 0.85 | 85.44 | 0.51 |
| TwinFlow | 20B | No tradeoffs | 2 | 0.86 | 86.35 | 0.55 |
| TwinFlow (longer training) | 20B | No tradeoffs | 1 | 0.89 | 87.54 | 0.57 |
| TwinFlow (longer training) | 20B | No tradeoffs | 2 | 0.90 | 87.80 | 0.59 |
Citation
@article{cheng2025twinflow,
title={TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows},
author={Cheng, Zhenglin and Sun, Peng and Li, Jianguo and Lin, Tao},
journal={arXiv preprint arXiv:2512.05150},
year={2025}
}
