acceptodds
Withdrawn from ICLR 2027

SynerGenDet: Internal Flow Forensics and Timestep Adaptive Alignment for Unified Image Generation and Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image generation and generated-image detection have both advanced rapidly, but they iterate independently for a long time, lacking a mechanism for mutual promotion and co-evolution. Although some recent work has attempted to use a unified architecture to achieve co-evolution of generation and detection, it has not fully leveraged the advantages of a unified architecture: detection remains a purely semantic judgment, while generation does not effectively exploit detection feedback. Therefore, we propose SynerGenDet: a unified framework that deeply couples generation and detection. On the detection side, we leverage the model’s native generative capability for authenticity discrimination in a single unified forward pass, without requiring an auxiliary generative model or iterative denoising. Specifically, we use its one-step flow matching error as a generative forensic cue that complements semantic evidence. On the generation side, we introduce Timestep adaptive Alignment, which guides image generation with detector feedback at the right coarse-to-fine granularity. Experiments on multiple benchmarks show that SynerGenDet achieves state-of-the-art performance.

Primary area:
generative models
Keywords:
AI-Generated Image Detection, Generation–Detection Co-evolution, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_84AbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Building upon BAGEL and UniGenDet, this paper trains a model to simultaneously perform AI-generated image detection and image generation. In Stage 1, the generator's single-step flow (velocity) error on the input image is fed into the detection head. In Stage 2, the detector is frozen, and generator features are aligned with detector features using time-varying weights. Detection performance is evaluated on FakeClue, DMImage, and ARForensics, while generation quality is assessed using GenEval and FID.

Critical strengths
  1. Using the model's own flow path as a detection cue is a simple yet sound idea. The comparison with UniGenDet was conducted under controlled conditions, with identical initialization, data, and computational budgets.
  2. On the ARForensics benchmark, the worst-case accuracy is significantly better than that of UniGenDet (97.5 vs. 89.4 on LlamaGen).
Critical weaknesses
  1. The gains in detection performance are modest, and the evaluation scope is limited. Compared to UniGenDet, the improvement is +1.0 on FakeClue and +0.1 on ARForensics; notably, UniGenDet outperformed the proposed method on 5 out of the 7 generators in the ARForensics suite. A simple sweep of the detection threshold (tpt_p) could yield similar fluctuations in accuracy. No random seeds were reported. The only evidence supporting the "flow-error" feature claim is the comparison with UniGenDet, yet it is unclear whether the UniGenDet figures in Tables 2–3 stem from retraining or the use of original weights (though the Table 2 caption specifies "original model weights"). Key benchmarks such as GenImage, UniversalFakeDetect, and Chameleon were omitted, as were tests involving JPEG compression and resizing—factors critical for methods relying on reconstruction error cues.

  2. Table 2 is difficult to interpret. By reverse-engineering the real/fake ratios from the "Real," "Fake," and "Overall" accuracy figures, I estimate that real images constitute approximately 78% of the UniGenDet set but only about 25% of the SynerGenDet set. Since SynerGenDet’s accuracy on real images is actually lower, the reported overall gain is heavily influenced by this class distribution.

  3. The benefits on the generation side were not rigorously controlled. Most of the FID improvement (22.5 → 18.0) occurred during Stage 1; however, Stage 1 involved fine-tuning on 80K LAION images, and the FID was subsequently evaluated using LAION prompts. GenEval scores remained virtually unchanged (an average of 0.877 across six categories versus BAGEL’s 0.875). I would like to see the results of simply fine-tuning BAGEL on the same dataset, as well as the outcome of aligning with an encoder that is not a detector.

Questions for authors
  1. What are the respective numbers of real and fake images in Table 2? Were all baselines evaluated using the same data split?
  2. What is the FID score when BAGEL is fine-tuned solely on the LAION dataset?
  3. How much does detection performance drop if the eˉ\bar{e} component is removed from the detection head of the same checkpoint? Do Tables 1–3 utilize the Stage 1 or Stage 2 checkpoint? (Note: Stage 2 involves continued training of the generation expert responsible for producing eˉ\bar{e}.)
Reviewer_NecCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SynerGenDet, which combines image generation and generated-image detection within a unified model. In the first stage, Internal Flow Probing (IFP) fuses one-step flow-matching errors with semantic features for authenticity classification. In the second stage, the detection expert is frozen, and Timestep-Adaptive Alignment (TAA) adjusts global and local feature supervision according to the noise level to improve generation quality.

Critical strengths

1.IFP provides complementary forensic evidence through the model’s internal one-step flow-matching error, without requiring an additional generative model or iterative reconstruction. Isolating the probe stream from the clean-image input also prevents direct leakage of the prediction target.

2.The first-stage comparison with UniGenDet controls initialization, data, and training settings, providing reasonably convincing evidence for IFP’s detection benefits.

3.On ARForensics, the worst-generator accuracy improves from 89.4% to 97.5%, indicating more balanced cross-generator performance. Generation experiments also show improved FID while retaining strong compositional text-to-image generation capabilities.

Critical weaknesses

1.Unified generation and detection, as well as detector-guided generative alignment, have precedents, while the average accuracy improvement on ARForensics is only 0.1 percentage points and TAA reduces FID by only 0.4 relative to fixed weighting.

2.The comparison between BAGEL and “BAGEL + IFP” also includes additional joint training. Without corresponding generation results from a model trained without IFP under the same data and budget, this FID improvement cannot be clearly attributed to IFP.

3.Flow-error features are fused only in the final classification head and are not directly accessible to the explanation pathway, so the faithfulness of the explanations to the final classification has not been established. CSS also decreases from UniGenDet’s 79.8 to 78.5, indicating that explanation quality does not improve consistently across metrics.

Questions for authors

1.Under the same data and training budget, what FID does a jointly trained model without IFP achieve? Does TAA using detector features offer a distinct advantage over using features from the original understanding expert or a general-purpose visual encoder?

2.Are the authenticity predictions from the classification head consistent with those expressed in the textual explanations? Since the explanation pathway cannot directly access flow errors, how can the explanations be verified to reflect IFP’s contribution to the final decision?

3.Are the reported detection results obtained from the model after Stage 1 or Stage 2? Does updating the generation expert in Stage 2 change the flow-error distribution or detection performance?

Reviewer_wbQVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SynerGenDet, a two-stage framework that unifies image generation and forgery detection. In the first stage, Internal Flow Probing (IFP) extracts the single-step Flow Matching prediction error and fuses semantic features to enhance authenticity detection. In the second stage, Timestep-Adaptive Alignment (TAA) dynamically adjusts local and global feature alignment according to the generation timestep. Experiments achieve competitive results on multiple detection and generation benchmarks.

Critical strengths
  • Reasonable use of internal information from generative models. IFP directly leverages the Flow Matching Error of the unified model to assist forgery detection, without requiring an additional generative model or iterative denoising, which has certain technical value.
  • Good cross-generator generalization. It achieves high detection accuracy on multiple benchmarks and significantly improves the worst-case performance on unseen generators.
Critical weaknesses
  1. Limited core novelty. The overall two-stage framework follows UniGenDet; IFP is closely related to the existing generation-error-based detection method LaRE^2; and TAA mainly introduces local/global alignment and timestep weighting on top of DIGA, resulting in relatively limited methodological increments.
  2. Insufficient experimental gains and attribution. The overall improvement over UniGenDet is relatively small; TAA improves FID by only 0.4 compared with fixed weighting; and there is a lack of sufficient controlled-variable experiments and statistical stability validation.
Questions for authors
  1. I suggest that the authors clarify the substantive methodological differences between SynerGenDet and UniGenDet, particularly regarding IFP versus existing generative-error-based detection methods such as LaRE² and DIRE, and TAA versus DIGA
  2. Could the authors provide ablations comparing Semantic Only, Flow Error Only, and their fusion to demonstrate that IFP contributes independent forensic information?
  3. Could the authors compare TAA with reversed weighting, nonlinear scheduling, and other strategies, and report FID across multiple random seeds to validate the necessity and robustness of timestep-adaptive alignment?

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

Related papers

Discussion (0)

Sign in to comment.