PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhenglin Cheng, Peng Sun, Jianguo Li, Tao Lin

arXiv:2512.05150v2cs.CV

TL;DR

Large-model generation makes efficient sampling important, while existing few-step approaches face trade-offs in simplicity, stability, and quality. TWINFLOW trains with twin trajectories and no auxiliary trained networks or frozen teachers. On Qwen-Image-20B, its 1-NFE scores are highly competitive with the original 100-NFE model.

  • Problem

    Existing few-step methods can require auxiliary or frozen models and face instability, complexity, and memory overhead, while methods without adversarial guidance often lose quality below 4 NFEs.

    Source: §1 · p. 2

  • Method

    TWINFLOW extends time across twin trajectories and minimizes their velocity-field discrepancy, without auxiliary trained networks or frozen teacher models.

    Source: §1 · pp. 2–3

  • Results

    At 1-NFE, Qwen-Image-TWINFLOW scores 0.86 GenEval and 86.52 DPG, versus 0.87 and 88.32 for the original 100-NFE model.

    Source: §1 · p. 3

  • Takeaways & Limitations

    TWINFLOW provides a simple framework for one-step or few-step training on large models, with high-quality text-to-image generation demonstrated at different scales.

    Source: §5 · p. 11

Abstract

from arXiv · show

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.

1 INTRODUCTION

Large generative models make efficient sampling important, but existing few-step methods trade off simplicity, stability, and quality. TWINFLOW uses twin trajectories for simple one-step training and reports strong text-to-image results, including competitive 1-NFE performance on Qwen-Image-20B.

1 INTRODUCTION

  • Large-scale generative systems make efficient sampling economically important because continuous inference costs can exceed one-time training costs.Source: §1 · p. 1
  • Existing few-step methods either add auxiliary or frozen models, instability, complexity, and memory overhead, or often lose quality below 4 NFEs.The paper identifies these trade-offs as obstacles to simplicity and scalability for large models.Source: §1 · p. 2
  • TWINFLOW extends time from [0, 1] to [-1, 1], learning noise-to-data and noise-to-“fake” trajectories by minimizing their velocity-field discrepancy.The framework requires neither auxiliary trained networks nor frozen teacher models.Source: §1 · pp. 2–3
  • At 1-NFE, TWINFLOW achieves 0.83 GenEval, surpassing SANA-Sprint (0.72) and RCGM (0.80).Source: §1 · p. 3
  • At 1-NFE, Qwen-Image-TWINFLOW scores 0.86 GenEval and 86.52 DPG, versus the original 100-NFE model’s 0.87 and 88.32.The paper describes these scores as highly competitive with the original model.Source: §1 · p. 3

2 PRELIMINARIES

The preliminaries present an any-step framework that encompasses multi-step and few-step generative models. Its objective uses a prediction function along a probability-flow ODE trajectory, with specific settings recovering established model objectives.

2 PRELIMINARIES

  • The any-step framework defines f(xt, r) := xr − xt as the prediction from the current point xt to target point xr along a PF-ODE trajectory.Source: §2 · p. 4
  • The unified objective treats multi-step and few-step models as instances of one framework, with xt formed from noise z and data x.The passage defines xt = α(t)z + γ(t)x and identifies Fθ as a neural network, with Fθ− its no-grad version.Source: §2 · p. 4
  • Setting N = 0 recovers diffusion and flow-matching training objectives; sampling integrates the PF-ODE from noise at t = 1 to data at t = 0.Source: §2 · p. 4
  • For N = 1, particular time-step choices recover consistency-model, shortcut-model, and MeanFlow objectives.Source: §2 · p. 4

3 METHODOLOGY

TwinFlow introduces symmetric real and fake trajectories to create a self-adversarial training signal, then rectifies the real trajectory through velocity matching. It combines these objectives with the any-step framework to train one model for both multi-step and few-step generation.

3.1 TWIN TRAJECTORY FOR SELF-ADVERSARIAL TRAINING

  • TwinFlow extends time to [-1, 1], learning a real-data path at positive times and a model-generated fake-data path at negative times.The fake trajectory uses a different noise sample and is trained with a flow-matching objective on negative-time inputs.Source: §3.1 · pp. 4–5

3.2 RECTIFYING REAL TRAJECTORY VIA VELOCITY MATCHING

  • Rectification treats the fake and real trajectory distributions as a matching problem, using their velocity difference to form a tractable loss that straightens noise-to-data trajectories.The resulting straightening allows integration to be approximated with large steps for few-step or one-step generation.Source: §3.2 · pp. 5–6

3.3 THE TWINFLOW OBJECTIVE WITH PRACTICAL DESIGNS

  • TwinFlow combines the base objective, self-adversarial loss, and rectification loss to serve both multi-step fidelity and few-step efficiency.The method uses the N = 2 any-step formulation to enhance training stability.Source: §3.3 · p. 6
  • Training partitions each mini-batch between TwinFlow loss at r = 0 and base loss at randomly sampled r ∈ [0, 1], with λ controlling their relative batch sizes and gradient influence.Setting λ = 0 disables the TwinFlow term; larger values increase its contribution.Source: §3.3 · p. 6

4 EXPERIMENTS

Experiments evaluate TWINFLOW on unified multimodal and dedicated text-to-image models, then examine how its performance varies with model, training steps, and sampling settings. The reported results include competitive few-step benchmark performance and 1-NFE generation on Qwen-Image-20B.

4.1 EXPERIMENTAL SETUP

  • Evaluations cover unified multimodal models on GenEval, DPG-Bench, and WISE, and dedicated SANA-0.6B/1.6B models on GenEval and DPG-Bench.The Qwen-Image experiments include LoRA and full-parameter training; OpenUni-512 is also evaluated with full-parameter training.Source: §4.1 · p. 7

4.2 IMAGE GENERATION ON MULTIMODAL GENERATIVE MODELS

  • TWINFLOW’s unified design removes redundant parameters and supports full-parameter training at 20B; it reports >0.86 GenEval at 1 NFE without auxiliary components or architectural modifications.The paper reports that the 1-NFE result surpasses most multi-step models using 40–100 NFEs.Source: §4.2 · p. 7
  • At 1 NFE, Qwen-Image-TWINFLOW scores 0.86 on GenEval and 86.52% on DPG-Bench, closely matching the original model at 100 NFE and outperforming 4-step Qwen-Image-Lightning on GenEval and WISE.At 1 or 2 NFE, it consistently surpasses MetaQuery-XL, BLIP3-o-8B, and Bagel across all evaluation metrics, and outperforms Qwen-Image-RCGM on GenEval and DPG-Bench under both settings.Source: §4.2 · p. 7

4.3 IMAGE GENERATION ON DEDICATED TEXT-TO-IMAGE MODELS

  • Figure 3 reports higher-quality 1-NFE Qwen-Image-TWINFLOW outputs than original Qwen-Image at 16 NFE, and better visual detail at 2 NFE than original outputs at 32 NFE.The caption identifies prompts in Appendix D.3.Source: Figure 3 · p. 9
  • At 1 NFE, TWINFLOW-0.6B and 1.6B score 0.83 and 0.81 on GenEval, respectively; the 0.6B model exceeds 40-NFE SANA-1.5-4.8B.At 2 NFE, TWINFLOW-0.6B scores 0.84 on GenEval and 79.7 on DPG-Bench; the 1.6B variant scores 79.6 on DPG-Bench.Source: Table 4 · p. 10, §4.3 · p. 10

4.4 ABLATION STUDY AND ANALYSIS

  • DPG-Bench performance as λ increases from zero rises and then falls, peaking at approximately λ = 1/3.The authors interpret this as evidence that balancing samples in the local batch helps model performance.Source: §4.4 · p. 11
  • Adding LTwinFlow improves 1-NFE text-to-image performance across OpenUni, SANA, and Qwen-Image, with Qwen-Image increasing from 59.50 to 86.52.The model comparison uses the same dataset.Source: §4.4 · p. 11
  • As training steps increase, the sampling-step “comfort regime” shifts, while GenEval performance improves in both 1-step and few-step settings.The authors report significant gains as training progresses.Source: §4.4 · p. 11

5 CONCLUSION AND LIMITATIONS

TWINFLOW is presented as a simple framework for large-scale few-step generation, with text-to-image experiments across scales. The authors identify broader task and modality validation as open challenges.

5 CONCLUSION AND LIMITATIONS

  • TWINFLOW is presented as a simple framework for large-scale few-step continuous generation that avoids auxiliary trained components and frozen teacher models.The authors report text-to-image experiments across scales and tasks.Source: §5 · p. 11
  • The authors identify image editing as insufficiently explored and video and audio generation as modalities requiring further validation.They frame broader task and modality coverage as remaining limitations.Source: §5 · p. 11

A RELATED WORK

  • Multi-step diffusion and flow-matching systems generate by iterative transport from noise, requiring numerous sequential evaluations at inference.The passage describes reversing a noising process or integrating a probability-flow ODE.Source: A RELATED WORK · p. 18
  • Few-step distillation uses a frozen teacher, while quality often drops sharply below 4 NFEs; adversarial losses can improve sharpness and alignment but add complexity and instability.These frameworks typically add auxiliary modules and retain teacher dependence, increasing memory overhead and hyperparameter sensitivity.Source: A RELATED WORK · p. 18
  • For 1–2 NFEs on 1024×1024 text-to-image with 10B–20B backbones, existing pipelines face teacher dependence, extra networks, instability, and prohibitive memory costs.The passage identifies these barriers as hindering straightforward scaling.Source: A RELATED WORK · p. 18
  • Recent few-step work explores anchored consistency training, solution consistency, policy optimization, data-free distillation, self-evaluation, and practical evaluations of sCM and MeanFlow.FACM anchors to instantaneous velocity; SoFlow bypasses expensive JVP calculations; π-Flow uses imitation learning termed π-ID; other approaches include prior sampling and model self-estimates.Source: A RELATED WORK · p. 18
  • Recent distribution-matching work reframes DMD acceleration or mode coverage, while video methods combine consistency and score-distillation or use recurrent flow heads.Decoupled-DMD identifies CFG augmentation as the primary acceleration driver, and DMD-R integrates reinforcement learning; rCM and TMD target video generation.Source: A RELATED WORK · p. 18

B.1 DETAILED IMPLEMENTATION ON MULTIMODAL GENERATIVE MODELS

Multimodal-model experiments report 1-NFE GenEval results for Qwen-Image, alongside training-data and configuration details for the reported runs.

B.1 DETAILED IMPLEMENTATION ON MULTIMODAL GENERATIVE MODELS

  • OpenUni full-parameter fine-tuning used 512×512 resolution, batch size 128, and 600,000 steps; Qwen-Image LoRA used rank and alpha 64, while full-parameter training added a time embedder.The Qwen-Image-Edit LoRA setup used rank 64 and alpha 32; the LoRA setups comprised approximately 420M trainable parameters.Source: §B.1 · p. 19
  • 0.86 GenEval at 1-NFE is reported for Qwen-Image-TWINFLOW, rising to 0.90 with LLM-rewritten prompts versus 0.91 for Qwen-Image-RL.The authors note gains in Colors and Attribute Binding with rewritten prompts.Source: §B.1 · p. 19, Table 5 · p. 19

B.2 DETAILED IMPLEMENTATION ON DEDICATED TEXT-TO-IMAGE MODELS

Dedicated text-to-image model experiments report GenEval subtask results for TWINFLOW at 1-NFE, with training data and full-parameter tuning details.

B.2 DETAILED IMPLEMENTATION ON DEDICATED TEXT-TO-IMAGE MODELS

  • Training used BLIP-3o-60K, Echo-4o without its multi-reference split, and ShareGPT-4o-Image, totaling approximately 200,000 text-to-image samples.Source: §B.2 · p. 19
  • SANA-0.6B and SANA-1.6B experiments used full-parameter tuning for 30,000 steps, with batch sizes of 128 and 64, respectively.Source: §B.2 · p. 19, Table 6 · p. 20
  • At 1-NFE, TWINFLOW-0.6B/1.6B scores 0.84/0.79 on GenEval Position and 0.70/0.68 on Attribute Binding, outperforming listed multi-step models except FLUX-Schnell, excluding overall score.Source: §B.2 · p. 20

B.3 EXPLORATION ON IMAGE EDITING

A preliminary image-editing exploration fine-tuned Qwen-Image-Edit and reported 2-NFE results, while identifying further scaling as a possible route to 1-NFE performance.

B.3 EXPLORATION ON IMAGE EDITING

  • 3.47 on ImgEdit at 2 NFEs is reported for Qwen-Image-Edit-TWINFLOW, surpassing all multi-step models except Qwen-Image-Edit itself.The model edits in 2 to 4 NFEs.Source: §B.3 · p. 20
  • The preliminary exploration used approximately 15,000 square images, LoRA fine-tuning, and fixed 512×512 training resolution; test resolution matched the input image size.Source: §B.3 · p. 20
  • The authors state that further scaling could yield strong single-step image-editing performance, despite the preliminary dataset use and training strategy.Source: §B.3 · p. 20

C THEORETICAL ANALYSIS

This section derives a relation between the score and velocity using flow matching, then rearranges it to express the score in terms of the model output.

C.1 TRANSFORMATION FROM SCORE TO VELOCITY

  • The section derives a score–velocity relation by substituting the model output into the flow-matching expression.It begins with f_x(F_θ(x_t, t), x_t, t) = α(t) · F_θ(x_t, t) − ˆα(t) · x_t.Source: §C.1 · p. 21
  • The cited theorem expresses the velocity as x_t plus α²(t) times the score ∇_{x_t} log p_t(x_t).The section defines s(x_t) = ∇_{x_t} log p_t(x_t).Source: §C.1 · p. 21
  • Rearranging the score term yields t²s(x_t) = −t x_t + (t² − t)F_θ(x_t, t).Source: §C.1 · p. 21

D VISUALIZATION RESULTS

The visual comparison reports that Qwen-Image-Lightning produces repetitive outputs, while Qwen-Image-TWINFLOW retains variety and high-quality generation. The section also describes LPIPS-based diversity quantification and shows example generations from both models.

D.1 COMPARISON OF QWEN-IMAGE-TWINFLOW AND QWEN-IMAGE-LIGHTNING

  • Qwen-Image-Lightning shows obvious diversity degradation, while Qwen-Image-TWINFLOW retains varied, high-quality outputs.The comparison reports that Lightning’s images are often similar despite different latent noise; Table 9 quantifies diversity using LPIPS.Source: §D.1 · p. 21, Table 9 · p. 22, Figure 5 · p. 22
  • Diversity is measured as the mean pairwise LPIPS distance among four generated samples per prompt across the evaluated samples.Source: §D.1 · p. 21, Table 9 · p. 22
  • Figure 5 compares 1-NFE outputs prompted from DPG-Bench, while Figure 6 visualizes Qwen-Image-TWINFLOW generations at training steps from 200 to 6400.The images are described as beginning to take shape at 200–400 steps and refining visual details from 800–6400 steps.Source: Figure 5 · p. 22, Figure 6 · p. 22

D.2 VISUALIZATION RESULTS ACROSS TRAINING STEPS

The visualizations show how 1-NFE generations change during training, document selected prompts, and provide additional high-resolution samples and fake-trajectory images.

D.2 VISUALIZATION RESULTS ACROSS TRAINING STEPS

  • At 200–400 steps, 1-NFE performance converges rapidly; from 800–6400 steps, training refines visual details and image fidelity.Source: §D.2 · pp. 22–23

D.3 SELECTED PROMPTS USED FOR VISUALIZATION

  • The selected prompts for Figures 1 and 3 are documented to support reproducibility, including a Hong Kong street scene and a Baroque-style still life.Source: §D.3 · p. 24

D.4 HIGH RESOLUTION VISUALIZATION

  • Further Qwen-Image-TWINFLOW visualizations use randomly chosen prompts and are presented without curation or cherry-picking.Source: §D.4 · p. 24

D.5 FAKE TRAJECTORY VISUALIZATION

Related papers

Loading 2512.05150v2…