PaperlayerView exact revision on arXiv

Source-linked AI summary

Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation

Chenxi Zhao, Chen Zhu, Xiaokun Feng, Aiming Hao, Jiashu Zhu, Jiachen Lei, Jiahong Wu, Xiangxiang Chu, Jufeng Yang

arXiv:2604.18168v1cs.CV

TL;DR

MeanFlow research has largely focused on class-label conditioning, leaving efficient text-conditioned generation difficult because text requires stronger semantic understanding and conventional LLM-encoder integration performs unsatisfactorily. The paper analyzes the representation properties needed in few-step generation, adapts MeanFlow with BLIP3o-NEXT’s text encoder, and reports competitive one-step text-to-image synthesis with significant benchmark improvements. Its findings provide practical guidance for extending MeanFlow to text conditioning.

  • Problem

    MeanFlow has primarily been studied with fixed class labels, while flexible text inputs create a more demanding and insufficiently understood conditioning problem.

  • Method

    The authors analyze text representations under limited denoising steps, identify discriminability and disentanglement as important properties, and adapt MeanFlow using BLIP3o-NEXT’s LLM-based text encoder.

  • Results

    The resulting EMF system achieves competitive one-step text-to-image generation, significant improvements across multiple benchmarks, and markedly improved synthesis quality.

  • Takeaways & Limitations

    Text-conditioned MeanFlow generation benefits from selecting representations with strong semantic discriminability and disentanglement rather than relying on conventional encoder integration alone.

Abstract

from arXiv · show

Few-step generation has been a long-standing goal, with recent one-step generation methods exemplified by MeanFlow achieving remarkable results. Existing research on MeanFlow primarily focuses on class-to-image generation. However, an intuitive yet unexplored direction is to extend the condition from fixed class labels to flexible text inputs, enabling richer content creation. Compared to the limited class labels, text conditions pose greater challenges to the model's understanding capability, necessitating the effective integration of powerful text encoders into the MeanFlow framework. Surprisingly, although incorporating text conditions appears straightforward, we find that integrating powerful LLM-based text encoders using conventional training strategies results in unsatisfactory performance. To uncover the underlying cause, we conduct detailed analyses and reveal that, due to the extremely limited number of refinement steps in the MeanFlow generation, such as only one step, the text feature representations are required to possess sufficiently high discriminability. This also explains why discrete and easily distinguishable class features perform well within the MeanFlow framework. Guided by these insights, we leverage a powerful LLM-based text encoder validated to possess the required semantic properties and adapt the MeanFlow generation process to this framework, resulting in efficient text-conditioned synthesis for the first time. Furthermore, we validate our approach on the widely used diffusion model, demonstrating significant generation performance improvements. We hope this work provides a general and practical reference for future research on text-conditioned MeanFlow generation. The code is available at https://github.com/AMAP-ML/EMF.

1 College of Computer Science, Nankai University 2 AMAP, Alibaba Group

MeanFlow has mainly been studied for class-label-conditioned generation, while flexible text conditioning remains an open direction with greater semantic demands. The paper analyzes this challenge and develops an efficient text-conditioned MeanFlow approach using discriminative text representations.

  • Motivation: MeanFlow and related few-step methods aim to reduce the many denoising iterations typically required for high-quality image generation.MeanFlow models average velocity between time steps and supports efficient one-step generation.
  • Motivation: Prior MeanFlow research primarily evaluates class-label-conditioned ImageNet generation, leaving flexible text conditioning largely unexplored.Text conditions offer richer content but impose greater demands on semantic understanding than fixed class labels.
  • Challenge: Directly integrating LLM-based text encoders with conventional diffusion training fails to produce satisfactory one-step text-to-image results.The authors therefore investigate why powerful text encoders do not transfer straightforwardly into the MeanFlow framework.
  • Analysis: Limited denoising iterations expose differences in text representations, whose velocity fields can vary substantially in semantic fidelity.Some models achieve strong final performance only because multiple denoising steps correct low-quality underlying velocity fields.
  • Approach and outcome: Using BLIP3o-NEXT’s LLM-based text encoder, the authors adapt MeanFlow for text-to-image generation and report competitive one-step synthesis with improved quality.The method is presented as the first effective extension of MeanFlow from fixed class labels to flexible text inputs.
  • Analysis: The paper identifies semantic discriminability and disentanglement as key properties of text representations for few-step generation.These properties improve semantic fidelity when only limited denoising iterations are available.

2. Related Work

MeanFlow research has focused on accelerating class-conditioned generation, while this work extends its conditioning to flexible text inputs and analyzes the text representations needed for few-step synthesis.

  • MeanFlow: MeanFlow learns a flow map that directly predicts transitions between time points, avoiding costly ODE integration during inference.The transition from z_t to z_r is modeled directly through the predicted average velocity.
  • MeanFlow: The method derives a self-consistent training target by differentiating the transition along the trajectory and computing the total derivative with JVPs.Stop-gradient is applied to the target for stable optimization.
  • Text-conditioned generation: Prior MeanFlow studies primarily address class-label-conditioned ImageNet generation, motivating extension to richer text-conditioned image creation.Text inputs impose greater semantic-understanding demands than fixed class labels.
  • Text representations: Under one-step or few-step sampling, BLIP3o-NEXT preserves semantic integrity better than SANA-1.5, which loses semantic fidelity as steps are reduced.The comparison links stronger subject preservation with greater robustness to velocity-field integration errors.
  • Text representations: Image–text retrieval and subsequence-similarity analyses identify discriminability and disentanglement as important properties of high-quality text representations.BLIP3o-NEXT and Gemma show strong disentanglement, while vision–language-aligned encoders retrieve more semantically relevant images than SANA-1.5 and T5.
  • Text-conditioned generation: The proposed adaptation uses BLIP3o-NEXT representations to extend MeanFlow toward semantically faithful one-step text-to-image generation.The approach is presented as the first extension from fixed class labels to flexible text inputs within MeanFlow-based one-step generation.

4. Experiment

The experiments evaluate text-to-image generation with MeanFlow across benchmarks, training settings, and sampling-step budgets. The method achieves strong quality with few sampling steps while preserving robust evaluation performance.

  • The training recipe uses approximately 170,000 samples, a 1e-5 learning rate, batch size 128, and 150 epochs based on BLIP3o-NEXT.
  • The evaluation uses GenEval for text–image faithfulness, DPG-Bench for long-form instruction following and compositional robustness, and HPS-v2 for human perceptual preferences.
  • 0.90 GenEval is achieved with 4 sampling steps, nearly matching BLIP3o-NEXT’s 0.91 and outperforming nearly all pretrained models requiring more than 20 steps.
  • The method surpasses every distilled model while continuing training from a single set of pretrained weights rather than requiring one or more teacher models.
  • 4-step sampling reaches high generation quality within roughly 10k training steps and reaches 0.90 GenEval by 60k steps.
  • 2-step sampling reaches 0.85 GenEval at 70k steps, while 1-step sampling reaches 0.74 at 90k steps.

5. Discussion

The discussion examines how MeanFlow scales with sampling steps and training domains, and compares its results with pretrained and distilled systems. Performance improves with additional steps, while domain and encoder choices affect convergence.

  • Performance rises from 0.74 at 1 step to 0.90 at 4 steps, approaching BLIP3o-NEXT’s 0.91 baseline obtained with 30 sampling steps.
  • The method’s 4-step GenEval performance attains the best distilled-model result and rivals larger models, according to Table 3.
  • 81.94 DPG-Bench is reached at 8 steps, up from 81.20 with 4 steps, indicating continued gains beyond the shortest sampling schedules.
  • SANA-1.5 encoder fine-tuning raises GenEval from 0.81 to 0.85, but an additional MeanFlow stage remains ineffective.
  • The SFT variant of BLIP3o-NEXT converges stably, whereas SANA-1.5 remains training-unstable regardless of text-encoder fine-tuning.

6. Conclusion

The paper extends MeanFlow from class-label conditioning to flexible text conditioning for efficient text-to-image synthesis. It identifies representation properties needed for few-step semantic fidelity and validates the approach on BLIP3o-NEXT.

  • The work presents the first implementation of extending MeanFlow’s class-label-conditioned one-step generation to flexible text conditioning.
  • High-quality few-step text representations require strong semantic discriminability and semantic disentanglement to preserve semantic fidelity.
  • The method adopts BLIP3o-NEXT’s LLM-based text encoder and adapts MeanFlow on top of the BLIP3o-NEXT framework for efficient text-conditioned synthesis.
  • Empirical results report competitive one-step text-to-image generation with markedly improved synthesis quality and significant improvements across multiple evaluation benchmarks.

8. Velocity Field Learning Challenges: Class-Label vs. Text Conditions

Class-label and text conditioning produce different velocity-field learning dynamics. Discrete, separated class features support smooth trajectories, whereas complex text conditions create winding paths and require more corrective refinement.

  • Figure 7 contrasts the smooth class-label path with the tortuous text-conditioned path.
  • Class-label features form sparse clusters with large inter-class margins, supporting smooth denoising trajectories and stable average-velocity prediction.
  • Textual conditions produce more tortuous denoising trajectories because their higher complexity and coupled nature separate instantaneous from average velocities.
  • The divergence between instantaneous and average velocities can cause early-stage semantic drift and necessitate additional corrective steps under text conditioning.
  • Strong discriminability and disentanglement are essential for preserving semantic fidelity when the velocity field is learned with limited denoising steps.

9. Additional Experiment on text encoder

Analysis of SANA-1.5 and OpenUni shows that text encoders suited to conventional diffusion refinement may not provide the discriminability required for few-step MeanFlow generation.

  • After training, Gemma becomes less discriminative but more disentangled, supporting language-space refinement in 20-step generation.
  • Few-step MeanFlow generation instead requires strong image–text discriminability.

10. Inference Time Comparison.

The timing comparisons show that MeanFlow substantially reduces sampling time relative to 30-step Flow Matching, while end-to-end gains are limited by autoregressive text-embedding generation.

  • 1.24 s versus 0.22/0.12/0.08 s: MeanFlow requires far less sampling time than BLIP3o-NEXT with 30 steps at 4/2/1 steps.These measurements use H200 hardware and time diffusion sampling for the same prompt.
  • 9.87 s versus 11.3 s: end-to-end generation is only modestly faster for the 4-step version than BLIP3o-NEXT.Most remaining time is spent generating autoregressive text embeddings.

11. User Study and ImageReward Result

The evaluation measures instruction following with PickScore and a user study comparing five models on 50 prompts. Both experiments report better performance for the proposed method under four-step generation.

  • 50 prompts and 20 users formed the user-study evaluation of which model best matched each prompt.Users compared images from five models for every prompt.
  • Both PickScore and the user study show that our method performs better when all models use four-step generation.

12. Additional Quantitative and Qualitative Results

Additional evaluations examine text-conditioned generation under limited denoising steps using quantitative benchmarks and visual comparisons. Across these tests, the method preserves image quality, semantic fidelity, and prompt adherence at few-step settings.

  • DPG-Bench evaluation: The method consistently outperforms equally sized distilled few-step models on DPG-Bench under the same step setting.DPG-Bench emphasizes long prompts with intricate attribute bindings and spatial relationships.
  • Vertical comparison across sampling steps: At one, two, four, and eight steps, MeanFlow is compared against a Flow Matching baseline using the same sampling-step configurations.
  • Vertical comparison across sampling steps: MeanFlow retains salient object structures and fine-grained textures at extremely low step counts, whereas Flow Matching becomes blurred and loses fine details.The resulting images are described as visually coherent and semantically faithful at a fraction of the baseline’s inference time.
  • Horizontal few-step comparison: At four steps, the model preserves fine-grained details and follows textual instructions more faithfully than existing distilled models across challenging prompts.
  • Qualitative results: At one sampling step, MeanFlow produces relatively sharp images compared with blurred few-step Flow Matching outputs.
  • Qualitative results: At four steps, the approach achieves higher semantic fidelity and richer visual details than existing distilled models on complex text prompts.The comparisons include cases where competing models fail to render described content accurately.
Loading 2604.18168v1…