acceptodds
Under review as a conference paper at ICLR 2027

Prompt Prototype Learning (PPL): Rethinking Text Supervision for AI-Generated Image Detection

Abstract

Contrastive vision-language models (VLMs) have shown promise for generalizable AI-generated image (AIGI) detection, yet existing approaches rely on fixed single-token class labels or per-sample captions, a brittle form of supervision that fails to capture diverse cues to generalize across unseen generators. We propose Prompt Prototype Learning (PPL), a novel framework that replaces fixed class tokens with ensemble-averaged prototype embeddings. For each class, we construct multiple prompts capturing complementary aspects of visual realism, including texture, boundary consistency, and high-level semantics. These prompts are encoded and aggregated into a normalized class prototype, providing richer and more stable contrastive supervision. Training images are aligned with these prototypes, improving generalization to variations in generation artifacts. We apply lightweight LoRA to the vision encoder and train exclusively on ProGAN with four object categories. Despite this limited training distribution, PPL generalizes strongly to unseen generators in the UniversalFakeDetect (97.3% mACC and 99.8% mAP) benchmark, as well as across four additional datasets covering recent generators and real-world transmitted images (94.1% mACC and 97.7% mAP), spanning GANs, diffusion models, and commercial tools, surpassing all prior methods. Ablation studies confirm that PPL is the primary driver of performance gains, with texture, boundary, and geometric prompts each contributing complementary signals.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative AI, AI-Generated Image, Prompt Prototype Learning, Vision-Text Supervision, Generalizaiton

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

This paper addresses cross-generator generalization in AI-generated image detection and proposes Prompt Prototype Learning (PPL). The method uses a frozen text encoder to encode manually designed forensic prompts covering texture, boundary, and geometric cues. Class prototypes for real and generated images are constructed by normalizing individual prompt embeddings, averaging them, and normalizing the resulting vectors again. During training, a prototype-alignment loss and a binary classification loss jointly guide LoRA adaptation of the vision encoder. At inference, only the adapted vision encoder and a linear classification head are used. The paper reports strong generalization across multiple cross-generator benchmarks and a dataset containing images transmitted through social media, while using only generated images from four ProGAN object categories and their corresponding real images for detector training.

Critical strengths

1.Simple and task-relevant text supervision. The method aggregates prompts covering multiple forensic cues into class prototypes, providing richer textual supervision than individual class tokens. Normalizing each prompt embedding before averaging gives each prompt equal weight in prototype construction and reduces reliance on any single prompt. 2.Strong cross-generator generalization. The paper reports strong performance across multiple benchmarks containing unseen generators, as well as a dataset reflecting social-media transmission, while using only generated images from four ProGAN object categories and their corresponding real images for detector training. 3.Broad experimental coverage. The paper includes ablations on prompt content, prompt count, prototype normalization, and different vision backbones. Comparisons with and without PPL under the same backbone support its effectiveness. The evaluation also covers robustness to compression, blur, noise, and their combinations.

Critical weaknesses

1.The contribution beyond C2P-CLIP requires a more controlled comparison. The proposed method shares the overall framework of C2P-CLIP, including text-based supervision, LoRA adaptation of the vision encoder, joint contrastive and classification objectives, and vision-only inference. Its main methodological distinction is replacing per-image alignment with augmented captions by class-level alignment with prototypes constructed from multiple forensic prompts, alongside the adoption of a larger vision backbone. Although the paper includes ablations across backbones and prompt constructions, it lacks a direct comparison with C2P-CLIP under matched backbone and training configurations. This limits the assessment of the proposed supervision strategy’s incremental benefit over the existing framework. 2.The role of fine-grained forensic semantics lacks direct validation. The paper partly attributes its improvements to the texture, boundary, and geometric forensic knowledge encoded in the text prototypes. Ablations involving single class tokens, random captions, and different attribute groups support the effectiveness of the prompt design, but do not directly establish that these descriptions correspond to the visual attributes actually used by the detector. The current experiments do not clearly disentangle the effects of specific attribute semantics, real/fake class keywords, and the prototype constraint itself. Consequently, the empirical evidence supports the method’s effectiveness more strongly than its proposed semantic explanation.

Questions for authors

1.Could the authors directly compare PPL with C2P-CLIP using the same vision backbone and otherwise matched training settings, to isolate the contribution of class-level prototype supervision relative to per-image caption alignment? 2.How would performance change if the real/fake class keywords were retained but the forensic attribute descriptions were removed, or if prototypes were constructed using class identifiers unrelated to authenticity, such as “Trump/Biden”? With the training configuration, prompt count, and prototype aggregation procedure held constant, could these controls clarify whether the specific forensic descriptions provide benefits beyond generic class-discriminative signals?

3 comments
AuthorsOpenReview ↗

Clarifications and rebuttal: matched C2P-CLIP comparison and role of forensic descriptions

We thank the reviewer for the precise and constructive suggestions. Both requested controls were run on all five benchmarks with PE-Core-G/14 and identical training.

W1/Q1: Matched comparison with C2P-CLIP.

We trained C2P-CLIP and PPL with the same backbone, data, and settings. Values are the average ACC / AP over UFD, AIGC, AIGB, AIGH, and TruFak:

Table R1. Matched C2P-CLIP vs. PPL. Avg ACC / AP (%).

BackboneC2P-CLIPPPLGain
CLIP-L/1472.1 / 87.480.9 / 89.6+8.8 / +2.2
PE-Core-G/1491.4 / 96.794.8 / 98.1+3.4 / +1.4

With PE-Core, PPL leads on every benchmark, e.g., by +7.8 ACC on TrueFake. Since both methods share the LoRA adaptation, the joint objective, and vision-only inference, this isolates class-level prototype supervision against per-image caption alignment.

W2/Q2: Role of the forensic descriptions.

K=6, aggregation, and training are fixed throughout:

Prototype construction (PE-Core-G/14)Avg ACCAvg AP
C2P-CLIP, Camera/Deepfake91.496.7
C2P-CLIP, Biden/Trump91.197.0
PPL, generic templates (real/fake keywords, no descriptions)90.896.9
PPL, descriptions kept, authenticity words → Camera/Deepfake92.197.2
PPL, descriptions kept, authenticity words → Biden/Trump91.797.3
PPL, full forensic prompts94.898.1

Keywords without descriptions:

Six generic templates (e.g., "a photo of a real/fake image") reach 90.8 ACC, no better than C2P-CLIP.

Descriptions without authenticity semantics:

Replacing every authenticity-related word (real, camera, genuine, authentic, natural, and their fake counterparts) with neutral identifiers yields 92.1 (Camera/Deepfake) and 91.7 (Biden/Trump). The two neutral variants differ by only 0.4 ACC, so the drop reflects the removed semantics rather than the class name. C2P-CLIP is nearly unaffected by the same keyword swap (91.4 → 91.1), indicating that its keywords act only as class identifiers. This gives a clear decomposition. Generic templates give no gain over C2P-CLIP. Forensic attribute descriptions without authenticity wording add about +1 ACC. Tying those attributes explicitly to authenticity adds a further +2.7 to +3.1 ACC. Neither component alone explains PPL; the gain arises from their combination. We agree that these controls show the forensic semantics matter without proving that the detector uses the specific visual attributes described. Grad-CAM (Fig. 5) offers partial support: with PPL, attention shifts from isolated high-frequency points to coherent structural regions. We will revise the text to present the semantic explanation as consistent with, rather than established by, these results.

Dear Reviewer,

Thank you for reviewing our rebuttal and for raising your score. We’d like to briefly address two points that may remain open.

Backbone vs. method contribution.

We acknowledge that Table R1 shows the backbone contributes substantially. In the revision, we will report C2P-CLIP and PPL under matched backbones in the main results table, so that the improvement attributable to prototype supervision (+3.4 ACC / +1.4 AP on PE-Core, +8.8 / +2.2 on CLIP-L) is stated explicitly and separated from backbone effects.

Forensic-semantics interpretation.

We will revise the text to present this explanation as consistent with, rather than established by, the controls, as discussed in our rebuttal.

If any other concerns remain, we would be glad to address them within the discussion period.

Best regards,

The Authors

Dear Reviewer vPfV,

Thank you for your constructive feedback and for engaging with our rebuttal. Under matched backbones and training, PPL outperforms C2P-CLIP (+3.4 ACC on PE-Core-G/14, +8.8 on CLIP-L/14), and the controls you suggested show that generic templates (90.8) and authenticity-neutral identifiers (91.7) both fall below the full forensic prompts (94.8).

We believe these results address all the concerns raised in your review. If any remain, we would be glad to clarify them before tomorrow's deadline; otherwise, we would be grateful if you would consider them in your final assessment.

Best regards, The Authors

Reviewer_u9MyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Prompt Prototype Learning (PPL) to address the limitations of existing single class-token or image-specific caption-based supervision approaches in AI-generated image detection. The proposed method constructs prompts describing various visual characteristics related to image authenticity, such as Texture, Boundary, and Geometry. The text embeddings of these prompts are normalized and averaged to generate Real/Fake class prototypes. The method then combines Binary Cross-Entropy (BCE) Loss with a prototype-based Contrastive Loss to guide the representation learning of the Vision Encoder. The model is trained exclusively on ProGAN data and evaluated across five benchmarks. It achieves 97.3% mACC and 99.8% mAP on UFD, as well as an average of 94.1% mACC and 97.7% mAP across four additional datasets. The proposed method also demonstrates greater robustness than existing approaches against image distortions such as JPEG compression, blur, and noise. Furthermore, extensive ablation studies are conducted to analyze the effects of prompt configurations, prototype normalization, and model settings.

Critical strengths
  • One of the main strengths of this paper is that it achieves strong detection performance across various generative models using only limited ProGAN training data. The proposed method achieves high average accuracy on multiple benchmarks covering not only traditional GANs but also Diffusion, Autoregressive, and Hybrid generative models. These results demonstrate the potential to detect images generated by unseen models, even when trained on a limited dataset.

  • Another strength is the use of specific visual artifacts in text supervision, rather than relying solely on simple Real/Fake class-level descriptions. The authors construct prompts that describe various visual characteristics, such as Texture, Boundary, and Geometry, and integrate them into prototypes to guide image representation learning. This approach is meaningful as it incorporates prior knowledge of synthetic image characteristics into the learning process.

Critical weaknesses
  • The paper claims improved generalization to unseen generative models as its main contribution, but it remains unclear whether these improvements can be attributed to the proposed method itself. While most existing methods use CLIP ViT-L/14, PPL employs PE-Core-G/14, a substantially larger backbone with approximately 1.88B parameters. In particular, IAPL achieves an average accuracy of 95.61% using CLIP ViT-L/14, whereas PPL achieves 97.3% with a much larger backbone. However, when PPL is applied to CLIP ViT-L/14, its accuracy drops to 94.4%, which is lower than that of IAPL.

  • The authors provide ablation results showing that PPL improves performance when applied to the same backbone. However, these results alone are insufficient to demonstrate that PPL offers superior generalization compared to existing methods. Without a direct comparison under the same backbone and experimental conditions, it is difficult to determine whether the reported performance gains come from the proposed method or the use of a larger pretrained model.

  • Furthermore, the rationale behind the prompt selection process is not sufficiently explained. The paper aims to improve generalization by integrating diverse artifact-related information, yet the actual prompt selection process primarily relies on the detection accuracy of individual prompts. In particular, the fact that using all 12 prompts leads to lower performance raises questions about how integrating diverse artifact-related information contributes to improved generalization.

Questions for authors
  • The paper demonstrates strong generalization performance across various image generation models, but its detection accuracy is relatively lower on certain high-quality generative models. Given the rapid advancements in image generation quality, can the proposed method maintain its detection performance on recently released state-of-the-art image generation models?
3 comments
AuthorsOpenReview ↗

Rebuttal and clarifications on IAPL, backbone, and prompt selection

We thank the reviewer for the careful reading. All new experiments use the identical ProGAN training set and settings, and we report average ACC / AP (%) over the five benchmarks (UFD, AIGC, AIGB, AIGH, TrueFake).

W1: IAPL vs. PPL with CLIP-L/14.

The 95.6% figure is IAPL's mean ACC on UFD only. There, PPL with CLIP-L/14 (94.4%) is indeed lower, and IAPL is also stronger on AIGC; we will state this explicitly. Our main claim, however, concerns generalization to newer and harder distributions, where PPL leads under the same backbone:

Table R1. IAPL vs. PPL, CLIP-L/14.

CLIP-L/14UFDAIGCAIGBAIGHTruFakAvg ACC / AP
IAPL95.691.570.265.255.075.5 / 84.0
PPL94.484.672.885.167.480.9 / 89.6

The gains are largest on AIGH (+19.9) and TrueFake (+12.4).

W2: Same-backbone comparison.

We re-trained C2P-CLIP, IAPL, and FatF on PE-Core-G/14. Values are the average ACC / AP over the five benchmarks:

Table R2. Baselines re-trained per backbone. Avg ACC / AP (%).

MethodCLIP-L/14PE-Core-G/14
C2P-CLIP72.1 / 87.491.4 / 96.7
IAPL75.5 / 84.087.5 / 94.1
FatF69.6 / 81.681.4 / 93.6
PPL80.9 / 89.694.8 / 98.1

PE-Core lifts every method (+11.8 to +19.3 ACC), confirming that the backbone matters. However, it is important to note that PPL remains best under both CLIP and PE-Core, and with PE-Core it leads on every benchmark (+7.3 ACC over IAPL). Tables 3–4 further show that PPL improves every backbone it is applied to (+1.5 to +3.5 mACC on UFD), and that PE-Core-G/14 without PPL (93.8%) is below PPL with CLIP-L/14 (94.4%).

W3: Prompt selection and the 12-prompt result.

Selection and diversity are complementary. Axis coverage helps: every pairwise category combination (96.2–96.8% mACC) beats every single category (95.1–95.5%) (Table 14). Selection does not remove diversity either, because the six selected prompts include all axes (texture 2, 3; boundary 6, 7; geometry 9, 12). What it removes is the weaker prompt within each axis. Using all 12 adds weaker or redundant descriptions, which pull the prototype toward less discriminative directions (96.1%). The best prototype therefore combines coverage across axes with quality within each axis. Two further controls show that diverse forensic content, not prompt count, drives the gains (Table 16). (i) Replacing PPL’s prompts with six generic templates (“a photo of a real/fake image”) lowers the five-benchmark average from 94.8% to 90.8%, so adding more prompts without forensic content does not help. (ii) Condensing all forensic cues into a single comprehensive prompt reaches 92.5–94.6% on UFD, below PPL’s 97.3%, so the content helps most when it is spread across diverse prompts.

Q1: Recent generators.

The additional benchmarks already include many recent models: FLUX.1-dev, SD3, Imagen 3, DALL·E 3, and Midjourney v6.1 (AIGB); FLUX, SD3.5-L, Infinity, Janus-Pro, and VAR (AIGH); and social-media-transmitted images (TrueFake) at Table 7. PPL achieves the best average ACC and AP on all four, including 97.8% on AIGH and 99.2–99.5% on the Janus family. We agree that some commercial and community sources remain harder (ACC: Imagen3 75.1%, Midjourney 84.9%, CommunityAI 78.6%). AP stays considerably higher on these sources, indicating mainly a shift of the fixed threshold rather than a loss of ranking signal. Multi-source training or threshold calibration may help, as noted in our limitations.

I appreciate author's responses and experiments to resolve my concerns. My main concern was the lack of comparison with other methods using the same backbone. I read Table R1 and R2 which report results obtained using the same training set and settings, and Table R1 shows that PPL is worse than IAPL on UFD, AIGC but generalizes well to harder distributions. Table R2 shows that PPL generalizes well on both backbones when compared with other methods using same backbone. After reading author's responses, I think my previous concerns are almost resolved and I have updated my score to 3.

We sincerely thank the reviewer for carefully reading our response and for raising the score. We are glad that the same-backbone comparisons resolved the main concern.

In the revised version, we will:

  1. add Tables R1 and R2 to the paper so that all methods are compared under the same backbone and training settings;
  2. state explicitly that IAPL outperforms PPL on UFD and AIGC with CLIP-L/14, while PPL generalizes better to the newer and harder benchmarks (AIGB, AIGH, TrueFake);
  3. clarify how prompts are selected and why using all 12 prompts performs worse, including the generic-template and single-prompt controls;

Thank you again for the constructive feedback, which has clearly strengthened the paper.

Reviewer_5k4nOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Prompt Prototype Learning (PPL) for AI-generated image detection. The goal is to improve generalization to unseen generators. The main idea is to replace single text labels with prototypes built from multiple prompts. The prompts are based on texture, boundaries, and geometric properties. The method uses a pretrained vision encoder with LoRA and contrastive learning. It is trained on ProGAN images from four categories. The validation covers five benchmarks with GANs, diffusion models, and commercial generators. The results show high performance. PPL achieves 97.3% mean accuracy on UniversalFakeDetect and 94.1% on four other datasets. It outperforms the compared methods. Ablation studies examine the effect of prompts and model components, proving and justifying the proposal.

Critical strengths
  1. The use of multiple text prompts is logic, well justified and easy to implement. It improves performance without extra inference cost. Ablation studies support its contribution. This makes the method practical and of interest to the community although it is not clear that the method will be available through github or any other repository.

  2. The method is evaluated on five benchmarks with many unseen generators. It outperforms the compared methods. The gains are large on some datasets. This supports its generalization claims.

Critical weaknesses
  1. The main idea is based on existing CLIP prompt ensembling. The prompts are manually designed and averaged. The technical contribution is rather incremental. The difference from existing prompt-based methods, sometimes justified as using several vs only 1 prompt, is weak i.e. not sufficiently strong. In fact Fig.2 shows several prompts but it is not clear that one is focused on texture, another on color, etc. It is not clear how the prompts themselves are generated.
  2. The method uses a very large pretrained vision encoder. The gains may partly come from this backbone. More controlled comparisons with simpler prompt ensembles are needed. The results are strong, but the contribution of the proposed supervision is not fully isolated. Fig 7b shows that with 6 prompts the ebst results are obtainned but until which extent this is generalizable for any dataset? Also the difference with other number of prompts even with 2 is nto so big.
Questions for authors
  1. What is the main technical difference between PPL and standard CLIP prompt ensembling? Can you show that averaging prompts during training provides benefits beyond existing approaches?
  2. How much of the improvement comes from the large PE-Core-G/14 backbone rather than PPL? Can you compare against simpler prompt-based methods using exactly the same backbone and training settings?
  3. How were the prompts designed and selected? How sensitive are the results to different prompt sets? Were any test datasets used during prompt selection?
  4. Can you clarify how training and test data overlap was excluded? How were classification thresholds selected across the five benchmarks? These details are important to me to assess the reported generalization of the proposed method.
1 comment
AuthorsOpenReview ↗

Clarifications and new controls (generic templates, re-trained baselines on PE-Core)

We thank the reviewer for the constructive feedback. All new experiments use the same ProGAN training set and settings, and we report average ACC / AP (%) over the five benchmarks (UFD, AIGC, AIGB, AIGH, TrueFake).

W1/Q1: PPL vs. standard CLIP prompt ensembling.

CLIP prompt ensembling averages generic templates to build a zero-shot classifier at inference. PPL differs in three ways (Sec. 2.4). (i) The ensemble is a fixed training target that shapes the LoRA-adapted vision features and is discarded afterwards, so inference uses only the vision encoder and a linear head. (ii) The prompts describe forensic evidence (sensor noise, optical blur, lens distortion, GAN/diffusion artifacts; Figs. 3 and 9) rather than rephrasing a class name. (iii) The prototypes are fixed and class-level, unlike the per-image captions of C2P-CLIP or the test-time-adapted prompts of IAPL, so PPL needs no captioner or per-sample optimization. To test this directly, we replaced our prompts with six generic templates (e.g., "a photo of a real/fake image"), keeping K, aggregation, and training identical:

Table R1. Generic vs. forensic prompts (K = 6, same training). PE-Core-G/14.

PE-Core-G/14Avg ACCAvg AP
C2P-CLIP91.496.7
PPL, generic templates (K=6)90.896.9
PPL, forensic prompts (K=6)94.898.1

Averaging generic prompts gives no gain over C2P-CLIP. The +4.0 ACC comes from the forensic content. Consistent with this, a single class token (94.4% mACC on UFD, Fig. 7(c)) and single comprehensive LLM-written prompts (92.5–94.6%, Table 16) both fall below PPL (97.3%). Fig. 2 shows illustrative prompts only; the full categories are in Figs. 3 and 9, and we will label them in Fig. 2.

W2/Q2: Contribution of the backbone.

We agree the backbone contributes to generalization, and we isolate it in two ways. (i) Tables 3–4: PPL improves every backbone on UFD (CLIP-L/14 92.9→94.4, DINOv2-L/14 90.5→93.1, PE-Core-L/14 93.1→96.3, PE-Core-G/14 93.8→97.3). PE-Core-G/14 without PPL is below PPL with CLIP-L/14. (ii) We re-trained the closest baselines on PE-Core-G/14:

Table R2. Baselines re-trained per backbone (Avg. ACC/AP).

MethodCLIP-L/14PE-Core-G/14
C2P-CLIP72.1 / 87.491.4 / 96.7
IAPL75.5 / 84.087.5 / 94.1
FatF69.6 / 81.681.4 / 93.6
PPL80.9 / 89.694.8 / 98.1

PE-Core lifts all methods (+11.8 to +19.3 ACC). Even so, PPL is best under both CLIP and PE-Core. Its gains over the baselines range from +3.4 to +13.4 ACC. In Fig. 7b, mACC varies only between 96.1% and 97.3% for K=2–12, so PPL is not sensitive to K. We fixed K=6 once and applied it unchanged to the four other benchmarks.

Q3: Prompt design and selection.

The prompts were written manually from known forensic properties (Sec. 3.3). Real prompts span texture, boundary optics, and geometry. Fake prompts span general, GAN, and diffusion artifacts, describing artifact families rather than specific models. Each prompt was scored individually, and the top K were averaged; the selected six include all axes. We acknowledge that these ablations were run on UFD. AIGC, AIGB, AIGH, and TrueFake were never used for ablation, and PPL's largest gains appear there (+11.9 ACC over the second-best method, +26.1 on TrueFake). We will state this protocol explicitly. Within the library, sensitivity is low: single prompts span 94.1–95.3% and structured sets 95.1–97.3%. Content matters, as intended: neutralizing the authenticity wording costs 2.7–3.1 ACC, and generic templates cost 4.0 ACC.

Q4: Overlap and thresholds.

Training uses only real and ProGAN images from four LSUN categories (Wang et al., 2020). ProGAN appears at test time only in the disjoint UFD ProGAN split, and no other evaluated generator is seen in training. All benchmarks are evaluated zero-shot with the same weights and a single fixed threshold (0.5 on the sigmoid output), without calibration or test-time adaptation. AP is threshold-free and supports the same conclusions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of rebuttal: matched-backbone comparisons, forensic-prompt controls, and evaluation protocol addressing all reviewer concerns.

Dear Area Chair,

We thank you and the reviewers for their time and constructive feedback. As we have not yet received responses from the reviewers, we provide a concise summary of the main concerns and how we addressed them.

Overview. Reviewer 5k4n (weak accept) found the method well justified and the evaluation broad. Reviewers u9My and vPfV acknowledged the strong cross-generator generalization. The shared central question was how much of the gain comes from PPL versus the larger PE-Core-G/14 backbone. Reviewer vPfV raised their score after our rebuttal.

1. Backbone vs. method contribution (all reviewers).

We re-trained C2P-CLIP, IAPL, and FatF on PE-Core-G/14 with identical data and training settings (average ACC / AP over five benchmarks: UFD, AIGC, AIGB, AIGH, and TrueFake):

MethodCLIP-L/14PE-Core-G/14
C2P-CLIP72.1 / 87.491.4 / 96.7
IAPL75.5 / 84.087.5 / 94.1
FatF69.6 / 81.681.4 / 93.6
PPL80.9 / 89.694.8 / 98.1

PE-Core improves every method (+11.8 to +19.3 ACC), confirming that the backbone matters. However, PPL remains the best method under both backbones. Against C2P-CLIP, the closest framework, the matched gain is +3.4 ACC on PE-Core-G/14 and +8.8 ACC on CLIP-L/14. The matched-backbone results will be included in the main results table.

2. Difference from prompt ensembling and the role of forensic content (5k4n, vPfV).

With the number of prompts (K=6), aggregation, and training held fixed (PE-Core-G/14, average ACC):

  • Generic templates ("a photo of a real/fake image"): 90.8, no better than C2P-CLIP (91.4).
  • Forensic descriptions with authenticity words replaced by unrelated identifiers (Biden/Trump): 91.7.
  • Full forensic prompts: 94.8.

The gain therefore requires both the forensic content and its link to authenticity; it does not come from prompt count or ensembling alone. Following Reviewer vPfV, we will present the semantic interpretation as consistent with, rather than established by, these controls.

3. Prompt selection protocol (5k4n, u9My).

Prompt and K ablations were run on UFD only. AIGC, AIGB, AIGH, and TrueFake were never used for any design decision, and K=6 was fixed once and applied unchanged. The selected prompts cover all three axes (texture, boundary, geometry), and sensitivity is low (96.1–97.3% mACC for K=2–12). We will state this protocol explicitly.

4. Evaluation protocol (5k4n).

Training uses only ProGAN and real images from four LSUN categories. All benchmarks are evaluated zero-shot with the same weights and a single fixed threshold (0.5), without calibration or test-time adaptation. AP, which is threshold-free, supports the same conclusions.

We believe our responses and the additional controlled experiments address all concerns raised by the reviewers, including the attribution of performance gains, prompt ensembling, the prompt selection protocol, and the evaluation setup. We appreciate your consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.