acceptodds
Under review as a conference paper at ICLR 2027

One Step Generation via Condition Shifting

13 views

Abstract

One-step text-to-image generation must preserve image quality and text alignment without repeated sampling updates. Trajectory-based objectives learn the mapping from noise to data, while distribution matching provides feedback on the samples that the generator actually produces. This feedback often requires an auxiliary network to estimate the generated distribution, increasing training cost as the backbone grows. Self-adversarial methods reduce this cost by learning real and fake trajectories within one model, using separate inputs for each task. We propose **APEX**, which learns the fake trajectory through condition shifting while retaining the original positive-time domain. Under the original condition, the model learns the real trajectory; under a deterministically shifted condition, it learns from its own outputs for the same text prompt. Flow matching on generated samples trains the shifted-condition prediction to follow the current generated distribution. Evaluating both velocities at the same noisy sample and time gives an internal distribution-matching correction, without a separate discriminator or fake-score network. Under score–velocity duality, this difference estimates the score discrepancy between the reference and generated distributions. We use it to construct a detached target for the one-step prediction and combine the resulting correction with the RCGM base objective. Both tasks share the existing network parameters, and inference uses only the original condition in a single forward pass. Across 0.6B–20B backbones, APEX supports full-parameter and LoRA tuning. SANA 0.6B achieves 0.84 GenEval with 0.20-second latency at one network function evaluation (NFE), and Qwen-Image 20B with LoRA reaches 0.89 GenEval at 1 NFE. Under matched training on Z-Image 6B at 2 NFEs, APEX improves over TwinFlow on all five reported metrics, including Chinese and English text rendering by 10.03 and 14.09 points. Condition ablations favor affine shifting over the other tested constructions on GenEval, DPG-Bench, and WISE.

Primary area:
generative models
Keywords:
Few-Step

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VNahOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an extension of step-distillation techniques based on the TwinFlow framework. TwinFlow constructs a fake-score estimator by extending the timestep domain to negative values, using the sign of the timestep to distinguish the real and fake trajectory estimation tasks. In contrast, APEX retains the original timestep domain and uses a shifted text embedding as the conditioning signal to distinguish the fake-trajectory task.

Critical strengths
  1. Clear presentation and implementation. The method builds upon prior works, particularly TwinFlow and RCGM, making the overall methodology and implementation relatively easy to follow.

  2. Promising empirical results. APEX achieves competitive or improved FID, CLIP, and GenEval scores over prior methods in Tables 1 and 2, demonstrating its potential for efficient text-to-image generation.

Critical weaknesses
  1. Unclear motivation and insufficient justification for the proposed conditioning mechanism. The main contribution appears to be replacing TwinFlow's timestep-based separation with condition-space separation. However, the paper does not sufficiently explain the limitations of TwinFlow that motivate this change or establish why the proposed alternative is preferable. More importantly, the rationale for the proposed conditioning mechanism is unclear. In TwinFlow, extending the timestep domain to negative values explicitly introduces a new input region to distinguish fake-trajectory estimation from real-trajectory estimation. In contrast, shifting the text embedding does not inherently guarantee that the model will learn a distinct fake-score estimator. Since the text embedding space is implicitly learned, the shifted embeddings may remain within the same semantic representation space as the original embeddings, and the paper does not explain how the shift reliably distinguishes the two estimation tasks. Consequently, the motivation and methodological justification for condition shifting remain insufficiently established.

  2. The evaluation lacks fair and comprehensive comparisons.

  • Latency and throughput comparisons are unclear. For methods using the same backbone, such as Qwen-Image, I would expect inference latency and throughput to be comparable under consistent evaluation settings. However, the reported numbers for APEX differ from those of Qwen-Image-Lightning, RCGM, and TwinFlow. Similar concerns apply to methods based on SANA. The paper should clarify whether these differences arise from different hardware, inference configurations, resolutions, or implementation optimizations, and provide a more controlled comparison.
  • The evaluation protocol and presentation of results could be improved. DPG-Bench is also an important benchmark for text-to-image generation and should be reported alongside other evaluation metrics rather than treated as a complementary evaluation (e.g. SANA paper and Qwen-Image report). Table 3 could be integrated with Table 1 to facilitate direct comparison. The main comparisons should focus on few-step generation methods, while comparisons with full-model systems such as GPT Image 1 [High] could be included as a reference for the performance gap to more capable systems.
  • The most informative comparison is insufficiently presented. Table 4 could provide particularly valuable evidence by comparing different step-distillation methods trained on the same dataset under controlled conditions. However, the paper reports only the NFE = 2 setting, despite being titled One-Step Generation. Including NFE = 1 results and providing a more comprehensive comparison with prior methods would make the table substantially more informative and better support the paper's central claim.
  • The ablation study on alternative condition constructions lacks sufficient context. The ablation study in Table 5 investigates alternative condition constructions, but the rationale behind the conditioning framework itself is not adequately established earlier in the paper. Without a clear explanation of why condition shifting should distinguish real and fake trajectory estimation, it is difficult to interpret these ablations or understand the principles underlying the design choices.
Questions for authors
  1. What specific limitations of TwinFlow motivate replacing timestep-based separation with condition-space separation?

  2. How does shifting the text embedding enable the model to distinguish real-trajectory estimation from fake-trajectory estimation?

1 comment

Dear Reviewer,

Thank you for your time and effort in reviewing the submissions. We kindly ask you to check the authors' responses and engage in further discussion if necessary. Please let us know promptly whether your concerns have been addressed or if any issues remain unresolved.

Your timely feedback would be greatly appreciated. Thank you for your cooperation!

Reviewer_nXnUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes APEX, a one-step text-to-image training framework that learns the fake trajectory via an affine shift of the text condition instead of TwinFlow’s negative time. Under the original condition the model learns the real trajectory; under a shifted condition it learns from its own outputs for the same prompt while staying in the positive-time domain. Evaluating real and shifted-condition velocities at the same noised sample and time estimates the score discrepancy and supplies a detached correction target for the one-step prediction, which is combined with the RCGM base objective. Both branches share the network parameters, no auxiliary discriminator or fake-score network is needed, and inference uses only the original condition in a single forward pass. authors report results from 0.6B to 20B parameters with full-parameter and LoRA tuning, matched retrains against TwinFlow on Z-Image 6B and Qwen-Image 20B, and ablations over condition constructions, shift parameters, loss weights, and data/update budgets. The headline numbers are 0.84 GenEval at 1 NFE for SANA 0.6B (0.20 s latency) and 0.89 for Qwen-Image 20B with LoRA.

Critical strengths
  • comprehensive controlled comparision and experiments. the paper’s matched retraining against TwinFlow on Z-Image 6B is exactly the controlled comparison this literature usually lacks most of the time. For example, it holds the backbone, data, updates, batch size, evaluator, and NFE fixed, and still shows improvement across different benchmarks, e.g. +5.85 GenEval, +3.68 DPG-Bench, +10.03 LongText-CN, +14.09 LongText-EN, and +1.16 OneIG. this is the evidence that makes the empirical claim credible.

  • ** the proposed method works across scales **. authors demonstrate that the proposed method works from 0.6B to 20B with both full-parameter and LoRA tuning, reaches 0.84/0.89 GenEval at 1 NFE with 0.20/0.39 s latency, and needs only one forward pass at inference. This shows the result is not a small-model artifact and makes the efficiency claim practically relevant, which matters for a one-step generation paper.

Critical weaknesses
  • λscoreλ_{score} term is not isolated. the paper never isolates its own correction term. the value of λ_score is not reported, and there is no λscoreλ_{score} = 0 run. Table 7 ablates a different endpoint formulation and the paper itself says those weights do not determine λscoreλ_{score} in Eq. (9). Because the detached score correction is the paper’s main mechanism, the matched gains could come from from the extra fake-flow task rather than from the correction. it is the main score-limiting gap, but it is fixable.

  • The strongest matched evidence is at 2 NFE, while the title and headline numbers promise one step. the gain +5.85 GenEval is a 2-NFE result, and the 1-NFE matched Qwen-Image gain is +0.02 GenEval with a lower DPG-Bench score. In addition, there are also no seeds, confidence intervals, or significance tests anywhere. without repeated runs, the small Qwen-Image margins and one-step framing are not statistically established.

Questions for authors
  1. What are the exact (a, b) and λ_score used for the headline models (Tables 1-3, 9), and what happens under the Z-Image matched protocol with λscoreλ_{score} = 0 ?
  2. Can authors report the Z-Image matched comparison at 1 NFE, and 2–3 seeds for the Qwen-Image matched comparison (Table 9) and Table 5, with std or paired intervals?
  3. Is the gain over TwinFlow attributed to condition shift, correction, or RCGM base?
Reviewer_1T8vOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces APEX, a self‑adversarial one‑step (1 NFE) text‑to‑image generation framework. It eliminates auxiliary discriminator or fake‑score networks by applying deterministic affine condition‑shifting to text embeddings to distinguish real‑data and fake generated‑sample trajectories within the original positive‑time domain. Drawing on score‑velocity duality, it derives an internal correction signal from real‑fake velocity discrepancy. It merges this corrective loss with the RCGM base objective, shares all network parameters, supports both full‑parameter and LoRA fine‑tuning and requires only one forward pass during inference. Its key contributions are the condition‑shifting self‑adversarial training paradigm, the detached velocity‑difference correction for one‑step prediction and thorough large‑scale evaluations across 0.6B‑20B‑parameter backbones. Validations show SANA 0.6B reaches 0.84 GenEval with 0.20‑second latency and LoRA‑tuned Qwen‑Image‑20B achieves 0.89 GenEval at 1 NFE. Controlled matched retraining against TwinFlow delivers overall metric improvements and prominent Chinese and English text‑rendering gains on Z‑Image‑6B with mixed results on Qwen‑Image‑20B. Ablations verify affine shifting as the optimal conditioning setup with evaluations covering GenEval, DPG‑Bench, WISE and other benchmarks.

Critical strengths
  1. This paper presents condition shifting self‑adversarial learning, a mechanism to produce internal distribution matching corrections entirely within a single model, removing the need for separate discriminator or fake‑score networks required by prior distribution‑matching distillation methods such as DMD. Unlike TwinFlow which relies on negative‑time inputs to separate real and fake trajectories, APEX keeps all training within the native positive‑time domain by transforming text embeddings. This matters for assessment because it solves a key practical pain point of fast image‑generation distillation: heavy extra training compute and memory overhead from auxiliary networks. The design is general and works across diverse backbones (0.6B‑20B) for both full‑parameter and LoRA fine‑tuning. It demonstrates a feasible path to high‑quality 1 NFE generation that is easy to adopt for existing flow‑matching models, raising the real‑world utility of the approach and representing a meaningful methodological advance for few‑step generative modelling.
  2. The paper performs controlled matched retraining against the strong baseline TwinFlow, fixing backbone, training data, update budget and evaluation protocol to enable fair comparison rather than only reporting disjoint system‑level numbers. On Z‑Image 6B APEX outperforms TwinFlow across all five metrics with very large improvements in Chinese and English text rendering. Across system evaluations it achieves state‑of‑the‑art GenEval scores for 1 NFE generation. This is consequential to assessment because one‑step generative methods frequently suffer from severe quality degradation at 1‑NFE. The matched experiments give credible evidence that the proposed condition‑shifting mechanism itself drives performance gains rather than differences in training setup. Combined with broad benchmarking over compositional alignment, prompt‑following and text‑rendering tasks, it provides solid empirical validation that supports confidence in the method’s effectiveness.
Critical weaknesses
  1. The score‑velocity duality identity which underpins APEX’s internal correction requires that FθF_\theta under original condition cc and shifted‑condition cshiftc_{\text{shift}} respectively converge to the true conditional‑mean velocity fields of pr(⋅∣c)p_r(\cdot|c) and pθ(⋅∣c)p_\theta(\cdot|c). However affine condition shifting cshift=ac+b1c_{\text{shift}}=ac+b\mathbf 1 only applies a linear transformation to text embeddings. Formal proof is required for the claim that the population optimum of flow‑matching loss trained on cshiftc_{\text{shift}} recovers exactly the velocity field of pθ(⋅∣c)p_\theta(\cdot|c), the generated distribution for the original prompt.
  2. Under controlled matched retraining, APEX surpasses TwinFlow on all metrics for Z-Image 6B with substantial gains in text rendering. But under an identical matched protocol for Qwen‑Image 20B, APEX improves GenEval and WISE but underperforms TwinFlow on DPG‑Bench for fine‑grained prompt following. The paper provides little investigation into the root causes of these backbone‑dependent divergent outcomes. This weakness weighs heavily on overall assessment because the core claim rests on condition‑shifting delivering reliable internal distribution‑matching correction. Mixed empirical results cast doubt on the generality of the proposed mechanism.
Questions for authors
  1. The correction builds on the premise that flow‑matching under cshiftc_{\text{shift}} yields a velocity field representative of pθ(⋅∣c)p_\theta(\cdot|c). The manuscript would benefit from included theoretical analysis or controlled empirical evidence demonstrating how closely the velocity field learned under cshiftc_{\text{shift}} approximates the true velocity of pθ(⋅∣c)p_\theta(\cdot|c). A large mismatch in this approximation could explain the mixed DPG‑Bench outcomes observed in matched Qwen‑Image 20B retraining.
  2. Under score‑velocity duality the velocity difference approximates score discrepancy when both velocity branches converge to their respective population‑optimum fields. What empirical sanity checks were performed to verify that both the original condition and shifted condition velocity fields are sufficiently well‑converged in their practical training setup?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.