acceptodds
Under review as a conference paper at ICLR 2027

EFG: Encoder-Free Forensic Guard for Synthetic Image Detection and Explanation

Abstract

AI-generated image detection has recently benefited from vision-language models (VLMs), which provide natural-language explanations to justify their judgments. However, existing VLM-based detectors heavily rely on visual encoders, which compress images into high-level semantic features, inevitably discarding low-level synthetic artifacts critical for forensic discrimination. Consequently, this compression creates a fundamental bottleneck, with detection-relevant signals largely lost before the model even begins reasoning. In this work, we propose EFG (Encoder-Free Forensic Guard), a specialized VLM that captures comprehensive forensic signals without visual encoders. Specifically, EFG converts the input image into token sequences via a lightweight patch embedding layer and directly feeds these raw pixel-level tokens into the large language model. This design preserves complete visual signals without encoder-induced information loss, while adaptively fusing low-level forensic traces with high-level semantics across layers. Beyond detection, EFG can produce natural-language explanations to justify its forensic judgments. Extensive experiments show that EFG achieves state-of-the-art performance across three AI-generated image detection benchmarks. Our work highlights the potential of encoder-free, pixel-space learning for VLM-based forensic systems. Code will be open-sourced upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthetic Image Detection, Deepfake Detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YvLAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EFG (Encoder-Free Forensic Guard), an encoder-free vision-language framework for AI-generated image detection and explanation. The authors argue that conventional VLM-based forensic detectors suffer from a visual encoder bottleneck, where semantic-oriented encoders discard low-level forensic artifacts required for reliable detection. To address this issue, EFG removes the external visual encoder and directly converts raw image patches into visual tokens through a lightweight patch embedding module, allowing the language model to reason over both low-level forensic traces and high-level semantics. The model is fine-tuned with LoRA and evaluated on multiple synthetic image detection benchmarks with both detection accuracy and explanation quality. The experimental results show strong performance compared with existing VLM-based forensic detectors and conventional detection approaches.

Critical strengths
  1. Important and timely problem. The paper addresses an important limitation of current VLM-based synthetic image detectors. Improving both detection accuracy and explanation ability is valuable for trustworthy AI-generated content analysis.
  2. Interesting encoder-free formulation. The key idea of removing the visual encoder and directly feeding pixel-level tokens into a native VLM is interesting. The paper provides a different perspective from previous approaches that attempt to adapt or fine-tune existing visual encoders.
  3. Strong empirical evaluation. The paper evaluates EFG on multiple benchmarks, including FakeClue, LOKI, Holmes-Set, and DMImage. The results demonstrate improvements over both conventional detectors and specialized forensic VLMs.
  4. Practical explainability advantage. Compared with traditional binary classifiers, EFG can provide natural-language explanations describing potential artifacts and inconsistencies, which improves interpretability.
Critical weaknesses
  1. The contribution of the encoder-free design is not sufficiently isolated. The central claim of the paper is that visual encoders introduce a fundamental bottleneck for forensic detection. However, the current experiments do not fully isolate whether the improvement comes from removing the encoder itself or from other differences, such as the choice of native VLM backbone, tokenization strategy, or training procedure. A more controlled comparison using the same backbone with encoder-based and encoder-free variants would provide stronger evidence.
  2. Limited methodological novelty beyond applying native VLMs to forensics. Encoder-free or native vision-language architectures have been explored in previous works. The main novelty of this paper comes from applying this paradigm to synthetic image detection rather than introducing a fundamentally new multimodal learning mechanism. The authors should better clarify the forensic-specific contribution compared with existing native VLM architectures.
  3. Insufficient analysis of how forensic signals are preserved. The paper claims that removing the visual encoder allows preservation of low-level forensic cues. However, there is limited analysis showing which visual tokens or internal representations capture these cues. More detailed visualization or attribution studies would make the proposed mechanism more convincing.
  4. Explanation evaluation remains incomplete. Although EFG generates better explanations according to automatic metrics, metrics such as ROUGE-L and CSS do not fully measure whether explanations are actually grounded in correct forensic evidence. Human evaluation or artifact-level verification would strengthen the explanation claims.
  5. Lack more visualization examples. The paper lacks sufficient visualization examples to qualitatively demonstrate the effectiveness of the proposed method. More visual comparisons, failure cases, and intermediate results would be helpful to provide stronger evidence and make the claims more convincing.
Questions for authors
  1. Can the authors provide a controlled comparison where the only difference is whether the visual encoder is used?
  2. Which components of the native visual tokens contribute most to forensic detection?
  3. How reliable are the generated explanations compared with human forensic judgments?
  4. Could you give more visualization examples to illustrate?
  5. Could the authors provide robustness evaluations under additional degradations, such as Gaussian noise or resizing?
3 comments
AuthorsOpenReview ↗

Response to Reviewer YvLA (1/2)

Dear Reviewer YvLA,

We thank you for the insightful and valuable comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns.

Weakness 1: Encoder-free contribution not isolated

Thank you for this precise comment. To further isolate the encoder variable, we add the controlled comparison: the same NEO-2.2B backbone, training data, LoRA configuration, schedule and resolution — only the visual front-end differs. Specifically, we replace EFG's patch embedding with (i) a frozen CLIP-ViT-L/14, (ii) a CLIP-ViT-L/14 fine-tuned on FakeClue, and (iii) a frozen SigLIP-400M, each connected to NEO through a linear projection.

Table 1: Same-backbone controlled comparison on FakeClue.

Visual front-endAccF1ROUGE-LCSS
CLIP-ViT-L/14 (frozen)89.688.943.764.2
CLIP-ViT-L/14 (fine-tuned)93.492.749.870.6
SigLIP-400M (frozen)91.290.546.467.1
EFG (patch embedding)97.196.959.179.0

All encoder-based variants underperform EFG on all four metrics. Fine-tuning the encoder partially recovers accuracy, consistent with Figure 1, but does not close the gap. Notably, the drop is much larger for ROUGE-L and CSS than for accuracy: adding a ViT introduces an additional feature level that is not aligned with the image-level features, but we have kept all other experimental conditions as consistent as possible.

Weakness 2: Limited methodological novelty

Thank you for this comment. The novelty of this work is a forensic finding: the visual encoder is a bottleneck for synthetic-image detection. Encoder-free design is the instrument to verify this, and it surpasses encoder-based methods in performance. We substantiate this with three observations. (1) We confirm that the encoder affects forgery features: tuning it improves Qwen2-VL-2B from 89.8 to 90.7 and LLaVA-1.6-7B from 90.1 to 91.2 on Holmes-Set (Figure 1). (2) Prior forensic works act on the encoder side — AIGI-Holmes fine-tunes it and Deep-VRM injects its features into deeper layers — and both improve detection. (3) We take this one step further: with the backbone, data and training fixed, removing the encoder improves both detection and explanation quality (Table 1), and the gain carries over to the unseen DMImage (93.3 vs. 91.7 for FakeVLM). This yields, to our knowledge, the first encoder-free forensic VLM.

Weakness 3: How forensic signals are preserved

Thank you for this important request. The forensic cues in EFG are not confined to a small, identifiable subset of tokens; they are distributed across the visual tokens produced by the patch embedding. We support this with three observations.

(1) In the same-backbone controlled comparison (Table 1), fixing the backbone, data, LoRA configuration, schedule and resolution and replacing EFG's patch embedding with a frozen/tuned CLIP-ViT-L/14 or a frozen SigLIP-400M lowers accuracy to 89.6–93.4 and, more sharply, ROUGE-L/CSS to 43.7–49.8 / 64.2–70.6, against 59.1 / 79.0 for EFG: the encoder removes information on which the explanation depends.

(2) Tuning the encoder improves two encoder-based VLMs (Qwen2-VL-2B 89.8 → 90.7, LLaVA-1.6-7B 90.1 → 91.2; Figure 1), confirming that the semantic alignment of the encoder is the limiting factor.

(3) Under low-level degradations (Table 4 of the paper), EFG stays accurate where conventional detectors collapse, and its explanations describe concrete low-level artifacts (texture, saturation, smoothness) that annotators verify match the image (content description 4.52 vs. 3.87 / 4.31).

Together, these results indicate that low-level forensic signals remain accessible at the patch level and are used by the model, rather than being compressed into high-level semantics.

AuthorsOpenReview ↗

Response to Reviewer YvLA (2/2)

Weakness 4: Explanation evaluation incomplete

Thank you for this comment. As ROUGE-L and CSS measure text similarity rather than grounding, we add a human evaluation.

Five annotators independently evaluate the same 50 images (balanced between real and fake and covering multiple generators) together with the answers of three models (EFG, FakeVLM and AIGI-Holmes). Two key aspects are assessed: (1) content description — whether the description correctly reflects the content of the image (1–5); (2) credibility — whether the annotator finds the judgment and its justification trustworthy (1–5). Annotators also select the most satisfactory answer among the three models.

Table 2: Human evaluation of model answers.

CriterionEFGFakeVLMAIGI-Holmes
Content description (1–5)4.523.874.31
Credibility (1–5)3.763.183.43
Win rate (%)40.627.432.0

EFG obtains the highest content-description and credibility scores and is preferred in 40.6% of the cases (vs. 27.4% and 32.0%). Since the image set is balanced between real and fake, these results cover both classes. We therefore revise the claim from "better explanations" to "more accurate and better-grounded explanations".

Weakness 5: More visualization examples

Thank you for this reminder. We will add more comparison cases and failure cases to the Appendix, together with a dedicated failure-case figure.

Question 1: Controlled comparison

Please see Weakness 1 (Table 1). The backbone, data, LoRA configuration, schedule and resolution are identical; only the visual front-end differs.

Question 2: Contribution of native visual tokens

We compare the image regions to which the fake token responds: no consistent, dominant subset of native visual tokens is found, and the response is distributed over the whole image. This indicates that all regions contribute to the fake prediction, which is consistent with the fact that the entire synthetic image is generated rather than locally manipulated.

Question 3: Reliability versus human judgments

Please see Weakness 4 (Table 2): a five-annotator evaluation of content description, credibility and preference, where EFG is rated highest on all three.

Question 4: More visualization examples

Please see Weakness 5.

Question 5: Robustness to Gaussian noise and resizing

We extend Table 4 of the paper with resizing (0.5–2.0×); the Gaussian blur results are reproduced from Table 4 of our paper.

Table 3: Additional degradation robustness (FakeClue accuracy).

MethodGaussian blur σ=1Resize (0.5–2.0×)
CNNSpot47.651.2
UnivFD53.360.1
NPR59.385.6
AIDE68.781.3
FakeVLM90.695.1
AIGI-Holmes91.394.8
Ivy-xDetector83.590.3
EFG92.095.2

EFG remains the most accurate under both degradations, and its drop under resizing is smaller than under Gaussian blur.

Dear Reviewer YvLA,

We are deeply grateful for the time and effort you have invested in reviewing our paper. To conserve your valuable time, here we summarize the key points of our rebuttal to facilitate a quick read.

  • Same-backbone comparison: we add a controlled comparison with frozen ViTs and fine-tuned ViT; EFG outperforms all of them on all four FakeClue metrics. (Please refer to Weakness 1 for details)

  • Contribution reframed: the novelty is the forensic finding, not the architecture. (Please refer to Weakness 2 for details)

  • Mechanism: encoder-based variants lose accuracy and, more sharply, explanation quality. (Please refer to Weakness 3 for details)

  • Explanation evaluation: we add a five-annotator human evaluation; EFG achieves the best content-description, credibility and preference scores. (Please refer to Weakness 4 for details)

  • Robustness: blur and resizing results added. (Please refer to Question 5 for details)

We humbly hope that our responses have adequately addressed your concerns. If you have further concerns, please let us know and we will continue actively responding to your comments and improving our submission.

Thank you for your invaluable time and consideration.

Best regards,

Authors of Paper 33562

Reviewer_Ui9POpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the inherent limitations of conventional Vision-Language Models (VLMs) in AI-generated image detection and explanation. The authors identify the pre-trained visual encoder as a critical bottleneck that compresses images into high-level semantic features, inevitably discarding fine-grained low-level forensic artifacts vital for forgery discrimination. To overcome this limitation, the paper proposes EFG, a specialized native VLM that bypasses the external visual encoder entirely.

Critical strengths

The paper accurately targets and empirically validates a fundamental bottleneck in existing VLM-based forensics: semantic-centric visual encoders discard critical low-level forgery cues, which fundamentally limits both detection accuracy and explanation faithfulness.

Removing the pre-trained visual encoder and adopting a native, encoder-free VLM framework (leveraging a lightweight patch embedding layer) is a refreshing and effective paradigm shift for forensic tasks. It bridges the gap between raw pixel statistics and linguistic reasoning without multi-stage training misalignment.

Critical weaknesses
  1. Removing the vision encoder directly is an interesting way to preserve multi-level information. However, the authors do not provide direct evidence showing that LoRA-trained large language models can extract multi-level information even without the vision encoder. It is possible that the model only maps features to high-level semantics in the end.
  2. How do the authors verify that these explainable answers are correct and correspond to the input image? For example, what exactly are chaotic texture, unnaturally vibrant colors, and overly smooth regions? Could both real and fake images have these properties?
  3. There is a lack of extensive quantitative results and ablation studies. For different questions, does the fine-tuned model retain the original capabilities of the VLM? We also need error-case analysis: for instance, cases where the model outputs neither "real" nor "fake", or returns other neutral descriptions.
Questions for authors

Processing raw pixel tokens directly without a heavy visual encoder avoids compression loss, but it significantly increases the sequence length of visual tokens fed into the LLM (especially under dynamic high resolutions up to 1024×10241024 \times 1024). Could the authors discuss the trade-off in terms of training GPU memory consumption, inference latency, and throughput compared to conventional encoder-based VLMs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ui9P

Dear Reviewer Ui9P,

We thank you for the insightful and valuable comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns.

Weakness 1: No direct evidence of multi-level extraction

Thank you for this fundamental question. We do not claim that every level of information is provably preserved inside the LLM, but the paper already provides behavioral evidence that the encoder-free design avoids the semantic bottleneck.

(1) The encoder limits detection, and removing it improves performance. Tuning the encoder improves two encoder-based VLMs on Holmes-Set (Figure 1): Qwen2-VL-2B 89.8 → 90.7, LLaVA-1.6-7B 90.1 → 91.2. We further add controlled NEO-2.2B variants with frozen CLIP-ViT-L/14, fine-tuned CLIP-ViT-L/14 and frozen SigLIP-400M, which differ only in the visual front-end (see Reviewer YvLA); EFG is best on all four FakeClue metrics.

(2) Under identical training (Figure 4), EFG reaches 97.1 / 83.3 / 92.5 on FakeClue / LOKI / Holmes-Set, above Qwen2-VL-2B (94.3 / 79.6 / 89.8), InternVL3-2B (88.8 / 81.1 / 86.4) and SmolVLM2-2.2B (95.7 / 81.9 / 91.6); if all inputs were reduced to high-level semantics, these encoder-based models would not be systematically worse.

(3) Under low-level degradation (Table 4 of the paper, JPEG QF=70 / blur σ=2), EFG scores 94.7 / 90.9, versus 91.1 / 89.0 (FakeVLM), 93.4 / 89.8 (AIGI-Holmes) and 66.5 / 65.3 (AIDE), indicating that it uses more than high-level semantics.

Taken together, these results indicate that the encoder-free design keeps low-level information accessible and that the model actually uses it, rather than compressing all inputs into high-level semantics.

Weakness 2: Verification of explanation correctness

The reference annotations are ground truth produced by an expert VLM directly on the images, so comparing model outputs against them is meaningful; the terms used in the explanations (e.g., chaotic texture, unnaturally vibrant colors, overly smooth regions) correspond to the artifact categories in these annotations.

We further conduct a human evaluation to verify the explanations against the image. Five annotators independently evaluate the same 50 images (balanced between real and fake and covering multiple generators) together with the answers of three models (EFG, FakeVLM and AIGI-Holmes). Two key aspects are assessed: (1) content description — whether the description correctly reflects the content of the image (1–5); (2) credibility — whether the annotator finds the judgment and its justification trustworthy (1–5). Annotators also select the most satisfactory answer among the three models, reported as a win rate.

Table 1: Human evaluation of model answers.

CriterionEFGFakeVLMAIGI-Holmes
Content description (1–5)4.523.874.31
Credibility (1–5)3.763.183.43
Win rate (%)40.627.432.0

EFG obtains the highest content-description and credibility scores and is preferred in 40.6% of the cases (vs. 27.4% and 32.0%).

Weakness 3: Limited quantitative results and ablations

Thank you for this constructive comment. We would like to address the points below.

1. Expanded ablations. We further add a visual front-end ablation and robustness evaluations, reported in our response to Reviewer YvLA (Tables 1 and 3); EFG outperforms all encoder-based variants on all four FakeClue metrics and remains the most accurate under Gaussian blur and resizing.

2. Retention of general capabilities. We do not claim that EFG retains the general capabilities of the base VLM. EFG is fine-tuned exclusively as a synthetic-image detector, and this specialization inevitably sacrifices part of the model's general ability; the same holds for other dedicated forensic models, and prior work does not report capability retention as an evaluation criterion. We will state this trade-off explicitly in the revision.

3. Error-case analysis. Because the training data uses a standardized system prompt and a fixed answer format, EFG output either "real" or "fake" in 100% of the evaluation cases; there were no neutral, non-committal or malformed responses. Representative failure cases are added to the Appendix.

Question 1: Efficiency trade-off

Thank you for raising this practical point. The concern does not materialize in terms of token count: in conventional encoder-based VLMs, one visual token also corresponds to roughly a 32 × 32 image patch, which is the same granularity as EFG. For models of the same scale, the inference cost is therefore comparable. In practice, EFG-2.2B is faster than AIGI-Holmes-7B, mainly because it has fewer parameters.

The advantage of the encoder-free design is not efficiency but trainability: the parameters through which the image enters the LLM are updated during fine-tuning, so the model can learn to focus specifically on artifact features, whereas a frozen encoder cannot.

Dear Reviewer Ui9P,

We are deeply grateful for the time and effort you have invested in reviewing our paper. To conserve your valuable time, here we summarize the key points of our rebuttal to facilitate a quick read.

  • Multi-level extraction: We add behavioral and controlled evidence, and EFG outperforms encoder-based variants on all four FakeClue metrics. (Please refer to Weakness 1 for details)

  • Explanation verification: We conduct a five-annotator human evaluation, and EFG is best on content description, credibility and preference. (Please refer to Weakness 2 for details)

  • Capability and errors: EFG is a specialized detector, and it outputs a valid real/fake verdict in 100% of evaluation cases. (Please refer to Weakness 3 for details)

  • Efficiency: EFG's inference cost is comparable at equal model scale. (Please refer to Question 1 for details)

We humbly hope that our responses have adequately addressed your concerns. Thank you for your invaluable time and consideration.

Best regards,

Authors of Paper 33562

AuthorsOpenReview ↗

Additional Results for Question 1: Efficiency trade-off

Thank you for raising this practical question. The native vision input does not increase the sequence length or the inference cost: the patch-embedding layer folds the image into 32 × 32 patches, so a 1024 × 1024 input produces 1,024 visual tokens (256²: 64; 512²: 256; 768²: 576). This is comparable to the number of visual tokens produced by encoder-based VLMs at the same resolution (e.g., a 28 × 28-merge encoder yields ≈1.3k tokens at 1024²). The encoder-free design therefore does not lengthen the LLM sequence; what it removes is the ViT computation that precedes the LLM.

We measured EFG on a single H100 (bf16, batch size 1, greedy decoding, input resolution 256²–1024²), matching the evaluation protocol used in the paper:

InputVisual tokensVision front-endPrefill + vision6-token verdict512-token explanationPeak memory
256²640.71 ms1.4 ms286 ms–5.49 GiB
512²2560.70 ms4.4 ms301 ms–5.56 GiB
768²5760.83 ms3.0 ms297 ms25.0 s5.70 GiB
1024²1,0240.87 ms9.3 ms298 ms25.0 s5.93 GiB

(1) The measurements show that prefill is cheap and decoding dominates. The vision front-end costs less than 1 ms even at 1024², and the 16× increase in visual tokens from 256² to 1024² adds only ≈8 ms to prefill, because prefill is fully parallel. The dominant cost is the LLM: autoregressive decoding requires a full pass through the LLM for every generated token (≈48 ms/token), independently of image resolution. Consequently, the 6-token verdict runs at ≈3.3 images/s single-stream, while a 512-token explanation takes ≈25 s. Since the visual tokens are consumed only during prefill, image resolution has little effect on end-to-end latency; output length and LLM size are the controlling factors.

(2) For training, with LoRA r = 128, per-device batch 16, gradient accumulation 2, 1024² input, and gradient checkpointing, EFG peaks at 41.9 GiB allocated (52.8 GiB reserved) per GPU without ZeRO, and runs at 4.4 s per optimizer step (32 images) on one H100. The 4 × A100-40 GB setting in the paper uses ZeRO-3 sharding, which brings per-GPU memory below 40 GB.

(3) This decode-dominated profile also determines how EFG compares with larger forensic VLMs. Because runtime scales with the size of the LLM rather than with the image encoder, EFG's 2B scale gives it a substantially lower per-token cost and memory footprint than 7B baselines such as FakeVLM, while achieving higher accuracy (Table 1). At the same 2B scale (Section 4.4.1), EFG's visual token count is comparable to that of encoder-based VLMs, and the encoder computation is removed entirely.

We hope that the additional efficiency measurements reported in Question 1 adequately address your concerns. Thank you for your invaluable time and consideration.

Reviewer_e15fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces EFG, a vision-language approach to AI-generated image detection and explanation. Motivated by the hypothesis that semantic-oriented visual encoders suppress low-level forensic cues, EFG adopts the encoder-free NEO-2.2B backbone, converts images into visual tokens through lightweight convolutional patch embeddings, and applies LoRA-based supervised fine-tuning to generate authenticity judgments and explanations. Experiments cover FakeClue, LOKI, Holmes-Set, and DMImage, together with JPEG compression and Gaussian blur evaluations.

Applying native vision-language modeling to image forensics is a worthwhile contribution. However, the current experiments do not sufficiently establish competitiveness against recent methods, isolate the benefit of removing the visual encoder, or substantiate the claimed faithfulness of explanations.

Critical strengths
  • Adapting NEO to jointly produce authenticity judgments and explanations has merit, even though the backbone and optimization techniques are established.
  • EFG reports higher accuracy than the compared forensic VLMs on FakeClue, LOKI, and DMImage. It also obtains the highest reported accuracy under all four degradation conditions.
Critical weaknesses
  • Incomplete coverage of recent baselines. Detection comparisons omit DDA, OmniAID, and CO-SPY; DDA is even discussed in related work. Recent explainable forensic approaches such as BusterX++ and applicable Veritas-family models are also missing. General-purpose comparisons rely largely on earlier models such as Qwen2-VL and InternVL2. Qwen3-VL-Plus appears only in qualitative examples, while fine-tuned InternVL3 appears only in the architectural comparison. These omissions weaken both the SOTA claim and broad conclusions about general-purpose VLMs’ forensic limitations. Impact: High.
  • Explanation metrics do not establish faithfulness. ROUGE-L and CSS measure similarity to reference text, not whether the described artifacts exist or actually drive the prediction. Moreover, EFG’s FakeClue CSS is lower than FakeVLM’s (79.0 vs. 80.1). Human assessment of visual correctness and hallucination, or controlled evidence interventions, is needed to support stronger explanation claims. Impact: High.
  • Some conclusions exceed the evidence. Removing a separate encoder does not guarantee lossless information preservation. Generalization is also uneven: EFG trails AIDE by 4.5 points on Holmes-Set and improves over AIGI-Holmes by only 0.2 points. Small gains are reported without uncertainty estimates. Impact: Moderate.
Questions for authors
  • Can you add recent detection and explainable-forensics baselines, as well as systematic MLLM evaluations, under clearly documented protocols?
  • Are the small gains stable across runs or supported by paired uncertainty estimates?
  • Can you validate explanation quality and faithfulness beyond reference-text similarity?
3 comments
AuthorsOpenReview ↗

Response to Reviewer e15f (1/2)

Dear reviewer e15f,

We thank you for the insightful and valuable comments. We respond to each comment as follows and sincerely hope that our rebuttal could properly address your concerns.

Weakness 1: Incomplete coverage of recent baselines

We thank you for this suggestion and have additionally evaluated three categories of models.

(1) For recent detection models, we add DDA [1], OmniAID [2] and CO-SPY [3], all fine-tuned on FakeClue and evaluated on the same test splits; for DDA, we follow its original design and use the Stable Diffusion VAE for reconstruction alignment.

(2) For MLLM-based forensic detectors, due to limited computational resources, we directly evaluate BusterX++ [4] and Veritas [5] on FakeClue and LOKI.

(3) For general-purpose VLMs, our evaluation serves only to show that these models cannot be directly used for synthetic image detection without task-specific fine-tuning; we additionally add InternVL3 to this category. As shown in Table 1, EFG still outperforms all of them in detection, and also produces better explanations than the MLLM-based models.

Table 1: Expanded baseline comparison.

MethodFakeClue AccLOKI AccROUGE-L
DDA84.375.3–
OmniAID91.479.9–
CO-SPY92.380.6–
BusterX++80.568.344.9
Veritas72.264.842.2
InternVL368.657.215.8
EFG97.183.359.1

In Table 1, DDA, OmniAID and CO-SPY perform detection only and do not provide explanations.

Weakness 2: Faithfulness of explanation metrics

Thank you for raising this important point. We address your concerns in detail below.

(1) Credibility of the generated text. To thoroughly investigate the credibility of the models' generated text, we conduct an additional human evaluation. Five annotators independently evaluate the same 50 images (balanced real/fake and covering multiple generators) together with the answers of three models (EFG, FakeVLM and AIGI-Holmes). Each answer is assessed on three measures: (1) content description — whether the description correctly reflects the content of the image (1–5); (2) credibility — whether the annotator finds the judgment and its justification trustworthy (1–5); (3) preference — the annotator selects the most satisfactory answer among the three models, reported as a win rate (%). This yields 250 judgments for each criterion.

Table 2: Human evaluation of model answers.

CriterionEFGFakeVLMAIGI-Holmes
Content description (1–5)4.523.874.31
Credibility (1–5)3.763.183.43
Win rate (%)40.627.432.0

As shown in Table 2, EFG obtains the highest content-description and credibility scores and is preferred in 40.6% of the cases, indicating that its explanations are grounded in the image content rather than being merely lexically similar to the reference annotations.

(2) The CSS gap. FakeVLM is built on a 7B VLM, whereas our EFG uses a 2B backbone; since the FakeClue explanations are generated by an LLM, larger models naturally obtain higher similarity scores on this metric. On LOKI, whose annotations are human-written, EFG surpasses FakeVLM in accuracy (83.3 vs. 82.6), ROUGE-L (21.6 vs. 18.4) and CSS (56.7 vs. 55.5), indicating that our explanations are more faithful at the semantic level rather than to a specific annotation format.

AuthorsOpenReview ↗

Response to Reviewer e15f (2/2)

Weakness 3: Conclusions exceed the evidence

We sincerely thank you for this fair criticism. We would like to clarify the three points below.

1. Clarifying "lossless". By "lossless" we mean that the LLM has direct access to the full pixel content of the image and decides internally which signals are useful and which can be ignored, whereas a ViT encoder compresses the image into a semantic space by design. The encoder-free design therefore retains a richer set of information. We agree that "lossless" is not entirely appropriate and will revise the wording to the more precise "avoids encoder-induced semantic compression".

2. The Holmes-Set gap is concentrated. As shown in Table 2 of the paper, EFG is better than AIGI-Holmes on 7 of 10 generator categories (Janus +1.4, J-1B +3.3, J-7B +6.8, Infinity +1.4, PixArt +1.3, SD3.5 +2.0, FLUX +5.6) and is best overall on 4 categories. The three categories where EFG loses (Show-o, L-GEN, VAR) are all autoregressive/token-based generators, which we already discuss in Section 4.2: their artifacts are absent from the training data, and since EFG is designed to exploit low-level artifact cues, this constitutes a distribution shift in precisely the signals it relies on. To verify this, we additionally include VAR images in the training set and re-evaluate the three AR categories.

Table 3: Accuracy after adding VAR to training.

MethodShow-oL-GENVARAvg
AIGI-Holmes98.198.497.998.1
EFG98.698.498.998.6

With VAR in training, EFG improves substantially on all three categories and surpasses AIGI-Holmes on Show-o (98.6 vs. 98.1) and VAR (98.9 vs. 97.9), confirming that the gap is caused by the missing AR artifacts rather than by the architecture. In the original Holmes-Set setting (Table 2 of the paper), AIDE cannot provide explanations, whereas EFG is the best among explanation-capable detectors (92.5 vs. 92.3).

3. Uncertainty results. All models use greedy decoding to reduce uncertainty, so the outputs are deterministic for a given checkpoint.

Question 1: Recent baselines

Please see Weakness 1. EFG remains the most accurate and produces better explanations than the MLLM-based models.

Question 2: Stability and uncertainty

Please see Weakness 3. All models use greedy decoding to reduce uncertainty.

Question 3: Explanation faithfulness

Please see Weakness 2. We have conducted a human evaluation with five annotators on 50 images.

[1] Chen et al. Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable. NeurIPS 2025.

[2] Guo et al. OmniAID: Decoupling Semantics and Artifacts for Universal AI-Generated Image Detection in the Wild. ICML 2026.

[3] Cheng et al. CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI. CVPR 2025.

[4] Wen et al. BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM. NeurIPS 2025.

[5] Tan et al. Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning. ICLR 2026.

Dear Reviewer e15f,

We are deeply grateful for the time and effort you have invested in reviewing our paper. To conserve your valuable time, here we summarize the key points of our rebuttal to facilitate a quick read.

  • Recent baselines added: DDA, OmniAID, CO-SPY, BusterX++, Veritas and InternVL3, all evaluated under documented protocols (Table 1). (Please refer to Weakness 1 for details)

  • Explanation faithfulness established: We conduct an additional human evaluation of content description, credibility and pairwise preference, and discuss the CSS gap transparently. (Please refer to Weakness 2 for details)

  • Claims clarified and better supported: we clarify the intended meaning of "lossless" and replace it with a more appropriate description, analyze the Holmes-Set gap per generator, and verify it by adding VAR to training. (Please refer to Weakness 3 for details)

We humbly hope that our responses have adequately addressed your concerns. Thank you for your invaluable time and consideration.

Best regards,

Authors of Paper 33562

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We thank you all for the time and effort invested in evaluating our submission. Below we provide a faithful and brief summary of our rebuttal.

We are grateful to the three reviewers (e15f, Ui9P, and YvLA) for their constructive feedback. All recognized the core strengths of EFG: (1) it targets a real bottleneck, as visual encoders discard critical low-level forgery cues (Ui9P); (2) the encoder-free formulation is an "interesting" and "refreshing and effective paradigm shift" (YvLA, Ui9P); and (3) its improvements across multiple benchmarks are strong (YvLA, e15f).

Alongside these strengths, the reviews raise a common set of concerns — causal attribution, explanation grounding, baseline coverage, and generalization, robustness and efficiency — which we address below with new experiments.

First, on whether the improvement truly comes from removing the encoder (raised by Ui9P and YvLA), we add a controlled comparison in which only the visual front-end differs — same NEO-2.2B backbone, data, LoRA configuration and resolution. EFG reaches 97.1 accuracy, against 89.6–93.4 for frozen or fine-tuned CLIP-ViT-L/14 and frozen SigLIP-400M (response to Reviewer YvLA, Table 1). The gap is largest in explanation quality (e.g., ROUGE-L 59.1 vs. 43.7–49.8), indicating that the encoder removes information on which the explanation depends.

Second, on explanation faithfulness (raised by all three), we assess the explanations against explicit human criteria: five annotators independently rate the same 50 balanced images across three models, and EFG is preferred in 40.6% of cases (vs. 27.4% and 32.0%), with the highest content-description (4.52 vs. 3.87 / 4.31) and credibility (3.76 vs. 3.18 / 3.43) scores (response to Reviewer YvLA, Table 2).

Third, on comparison with recent studies (raised by e15f), we add DDA, OmniAID, CO-SPY, BusterX++, Veritas and InternVL3 with each method's protocol stated; EFG remains best, reaching 97.1 and 83.3 on FakeClue and LOKI respectively (response to Reviewer e15f, Table 1).

Finally, we add three experiments that clarify several issues: (1) a per-generator analysis of the Holmes-Set gap (Reviewer e15f); (2) additional degradation experiments (Reviewer YvLA); and (3) measurements of output validity and native-token efficiency (Reviewer Ui9P).

We believe the added evidence materially strengthens the paper and clarifies the scope of its claims, and we are grateful that the reviews have improved the manuscript.

Thank you again for your efforts in handling our submission.

Sincerely,

Authors of Paper 33562

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.