Paper 2512.05150
TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows
- Published
- Dec 2025
- Research lab
- Independent
- Citations
- 16
- GitHub
- 537 stars
01 In brief
Summary
TWINFLOW is a framework for training one-step generative models on large multimodal models, addressing the inefficiency of multi-step diffusion and flow matching models that require 40-100 NFEs.
Existing few-step methods either rely on auxiliary trained models (e.g., GAN discriminators) or frozen teachers, causing instability and memory overhead, or degrade at very few steps.
TWINFLOW extends the time interval to [-1,1], creating twin trajectories: a positive branch maps noise to real data, and a negative branch maps noise to fake data generated by the model itself.
The training objective minimizes the velocity difference between these trajectories, creating a self-adversarial signal without external discriminators or teachers.
This enables 1-step generation with high quality.
On text-to-image tasks, TWINFLOW achieves a GenEval score of 0.83 at 1-NFE, outperforming SANA-Sprint (0.72) and RCGM (0.80).
Applied to Qwen-Image-20B, it matches the original 100-NFE model's performance (GenEval 0.86 vs 0.87, DPG-Bench 86.52 vs 88.32) with 1-NFE, reducing computational cost by 100x.
The method also scales to full-parameter training on 20B models, avoiding OOM issues seen with DMD2 and SANA-Sprint.
Ablations show the importance of the balancing hyperparameter λ and the TwinFlow loss, which improves 1-NFE performance significantly (e.g., Qwen-Image from 59.50 to 86.52 DPG score).
02 From the paper
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 of 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 $100\times$ with minor quality degradation. Project page is available at https://zhenglin-cheng.com/twinflow.