TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows

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.

• The paper introduces TwinFlow, a self-adversarial twin-trajectory framework that enables efficient 1-step generation on ultra-large models with nearly 100x sampling acceleration.
• The method leverages a unified any-step RCGM paradigm to map noise to data directly, eliminating the need for auxiliary networks, frozen teachers, or external discriminators.
• Empirical results on Qwen-Image-20B show that TwinFlow nearly matches the original model's 100-NFE performance at 1-NFE, maintaining high visual quality and sample diversity.

TwinFlow: One-step Generation on Large Models with Self-adversarial Flows

Introduction

The proliferation of large-scale generative models for visual synthesis has enabled highly capable multimodal systems but imposed prohibitively high inference costs due to the iterative nature of diffusion and flow-based methods. While prior acceleration approaches—such as progressive and consistency distillation, or adversarial distribution matching—can reduce the number of function evaluations (NFEs) required for sampling, they are hindered by compounding trade-offs between sampling quality, training complexity, and scalability for ultra-large models. TwinFlow addresses this core bottleneck by proposing a self-adversarial, twin-trajectory flow matching framework that realizes practical 1-step (and few-step) generative sampling on models of 20B+ parameters, entirely without auxiliary networks, frozen large teachers, or adversarial discriminators (2512.05150). This section overviews the method's conceptual core and its relevance for efficient high-fidelity generation.

Methodology

Unified Any-Step Framework

TwinFlow adopts and extends the any-step recursive consistent generative modeling (RCGM) framework, allowing both multi-step and few-step generation to be cast in a probabilistic flow-ODE paradigm. The generic prediction function $\mf(\xx_t, r) := \xx_r - \xx_t$ supports integrating diverse training objectives, where the core distinction of TwinFlow is its novel use of twin trajectories in time: for each sample, positive time $(t > 0)$ corresponds to mapping noise to real data, while negative time $(t < 0)$ maps noise to self-generated "fake" samples.

Self-adversarial Twin Trajectories

Rather than introducing auxiliary discriminators—as in GAN or DMD-style frameworks—TwinFlow creates a self-contained adversarial supervision by constructing a 'fake' trajectory alongside the real one. The model is tasked with matching the velocity fields of these forward (real) and backward (fake) trajectories. The key learning signal is derived by minimizing the discrepancy between the velocity fields at symmetric time points, which, under continuous-time probability flow, is theoretically equivalent to minimizing KL-divergence between the distributions defined along each trajectory.

Figure 1: The core TwinFlow framework, where the standard flow is contrasted with its twin; minimizing velocity field discrepancies eliminates the need for external discriminators or teacher models.

Rectification Loss

A tractable rectification loss operationalizes this idea: for each perturbed "fake" sample, the gradient of the loss aligns the model's velocity prediction with the stop-gradient twin velocity, strongly encouraging straight, direct transport from noise to data in a single step. This formulation directly supports efficient 1-step or few-step inference while also allowing for standard multi-step sampling; both goals are unified under the same model.

Training Simplicity and Scalability

TwinFlow is architecturally simple by design. It requires neither frozen teachers nor auxiliary networks, eliminating the compute and instability issues endemic to adversarial or distillation-based pipelines. The model's loss is a mixture of conventional flow-matching and the proposed self-adversarial objective, with balance controlled by a simple hyperparameter.

Empirical Results

Image Generation on Large Multimodal Models

Extensive experiments validate TwinFlow's performance across unified multimodal models (notably Qwen-Image-20B) and dedicated text-to-image generators. On Qwen-Image-20B, TwinFlow achieves 0.86 GenEval and 86.5 DPG-Bench scores at 1-NFE, nearly matching the original model's 100-NFE performance and surpassing prior 1-step and few-step methods. Notably, Qwen-Image-Lightning—a prior public 1-step baseline—suffers from severe mode collapse and insensitivity to stochasticity, whereas TwinFlow yields diverse, high-quality generations.

Figure 2: 2-NFE results from Qwen-Image-20B-TwinFlow, demonstrating prompt diversity and generation quality at an extremely low function evaluation count.

Figure 3: Qwen-Image-TwinFlow generates higher quality images at 1-NFE than the original Qwen-Image at 16-NFE; at 2-NFE, superior visual fidelity compared to 32-NFE sampling is observed.

A key practical claim is that TwinFlow enables $100\times$ sampling acceleration with minor quality degradation on ultra-large models, opening the door for real-time, high-throughput large model deployments.

Ablation, Diversity, and Scalability

Systematic ablation studies reveal that TwinFlow's efficacy hinges on properly weighted self-adversarial loss; over- or underweighting degrades quality. Critically, in direct head-to-head tests, TwinFlow-trained models exhibit robust sample diversity—confirmed both qualitatively and via LPIPS distance—whereas methods that bypass external adversarial or teacher signals consistently degenerate under 1-step settings.

Figure 4: Compared to Qwen-Image-Lightning, TwinFlow maintains output diversity without sacrificing quality in the 1-NFE setting, ameliorating mode collapse.

Qualitative Progress Across Training

Analysis across training steps highlights rapid initial convergence (200-400 steps), with significant refinement achievable with further optimization (800–6400 steps), demonstrating both efficiency and scalability in practical training regimes.

Figure 5: Visualization of Qwen-Image-TwinFlow 1-NFE outputs at increasing training steps: rapid convergence is followed by systematic improvement in detail and realism.

High-Resolution and Specialized Tasks

The method generalizes naturally to high-resolution image synthesis ($1328\times1328$) and image editing tasks, albeit the latter remains an area for further investigation. For image editing, even with minimal data and basic tuning, TwinFlow achieves competitive results using only 2–4 NFEs.

Figure 6: Representative high-resolution generations (NFE=4) from Qwen-Image-TwinFlow, underscoring preservation of fine details and compositionality.

Implications and Future Directions

Practical Impact

TwinFlow removes major barriers to the scalable deployment of large diffusive and flow-matching generative models by directly eliminating external teachers, discriminators, and auxiliary objectives. This results in dramatic improvements to throughput and latency (orders of magnitude in acceleration), decreases memory pressure (no duplicated model components), and provides robust generative diversity—all critical properties for interactive applications and model democratization.

Theoretical Considerations

By reframing adversarial regularization as internal self-play of velocity fields—rather than external discriminator competition—TwinFlow conceptually extends consistency and flow-matching theory towards maximal efficiency. This may offer new insights into the tractability of accelerating continuous generative processes, as well as inspire further unified frameworks where self-adversarial structures supplant externally-induced regularization.

Prospective Developments

Future developments should explore extending TwinFlow's framework to other modalities, especially video and cross-modal synthesis (audio, text), improved curriculum schedules for the twin loss balance, and integration with state augmentation or memory-efficient architectures. Additionally, bridging TwinFlow's trajectory-matching approach with emerging autoregressive or hybrid paradigms may further compact inference without quality compromises.

Conclusion

TwinFlow provides a rigorously validated, conceptually parsimonious solution for one-step and few-step high-fidelity synthesis on modern ultra-large generative models. Its self-adversarial flow-matching design dispenses with auxiliary models and external teacher dependencies, yielding a framework that is both scalable and simple. The strong empirical results and diversity preservation at scale underscore TwinFlow's significance for the practical advancement and future research of efficient generative modeling.