acceptodds
Withdrawn from ICLR 2027

ChromaGuard: Simple Chroma, Strong Generalization for Synthetic Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pretrained models have enabled remarkable performance in synthetic image detection by providing rich and transferable visual representations. However, their semantic information can become spuriously associated with forgery labels, causing detectors to rely on semantic shortcuts rather than synthesis artifacts and limiting generalization. To address this challenge, we propose ChromaGuard, a framework for generalizable synthetic image detection that employs chroma guidance to reduce semantic reliance. Through visualization analysis, we reveal that YCbCr chroma is very weakly related to image semantics. Building on this observation, ChromaGuard treats the RGB image and its chroma component as complementary views. The RGB branch preserves the complete image information to avoid missing synthesis artifacts, while the chroma branch guides the detector toward cues weakly associated with semantic content. A simple bidirectional Chroma-RGB interaction module connects the two branches, which alternately injects chroma guidance into RGB features and refines chroma features using the updated RGB representation. To further enhance sensitivity to synthesis artifacts within RBG branch, we introduce a DCT high-frequency adapter to mine frequency-domain artifacts left by generators. Evaluations of cross-generator generalization, semantic sensitivity, and robustness demonstrate its strong performance and reliability, supporting the effectiveness of Chroma guidance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Chroma, Semantic, Synthetic Image Detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_16VjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ChromaGuard, a framework for generalizable synthetic image detection that aims to reduce semantic shortcut learning in pretrained visual models. The authors observe that YCbCr chroma components exhibit weak semantic association compared with other image components, and use chroma as a guidance signal to encourage detectors to rely more on synthesis artifacts rather than image semantics. The proposed framework introduces a bidirectional Chroma-RGB interaction module that exchanges information between RGB features and chroma features, together with a DCT high-frequency adapter to enhance sensitivity to frequency-domain artifacts. Extensive experiments are conducted on cross-generator generalization, semantic sensitivity, and compression robustness, showing strong performance compared with existing synthetic image detectors.

Critical strengths
  1. The paper identifies an important generalization problem. The paper studies semantic shortcut learning in synthetic image detection, which is an important issue for deploying detectors across unseen generators. The motivation is clear and well aligned with current challenges in AI-generated content detection.
  2. The chroma-guided perspective is interesting. Using YCbCr chroma as a weakly semantic guidance signal is a simple but interesting observation. Compared with methods that suppress semantic information after feature extraction, the proposed approach attempts to influence representation learning from an earlier stage.
  3. The experimental evaluation is comprehensive. The paper evaluates the method under multiple settings, including cross-generator generalization, semantic sensitivity, and JPEG compression robustness. The results show consistent improvements over many existing baselines.
  4. The proposed framework is simple and practical. The method does not require complex training objectives or additional large models. The design based on RGB-chroma interaction and frequency adaptation is relatively lightweight and potentially easy to apply.
Critical weaknesses
  1. Limited methodological novelty and insufficient positioning with prior color-based approaches. Although the motivation is interesting, the proposed framework mainly combines existing ideas including pretrained CLIP features, cross-attention based feature interaction, and frequency-domain artifact modeling. The main novelty comes from selecting chroma as a guidance signal rather than introducing a fundamentally new detection paradigm. Moreover, the use of color-space information for synthetic image detection has been explored in previous studies, such as CHROMA: Detecting AI-Generated Images Through Inter-channel Color-Space Correlations. However, this line of work is not sufficiently discussed or compared with in the paper. The lack of a more comprehensive comparison with existing color-based detection approaches makes it difficult to clearly assess the unique contribution of ChromaGuard. The authors should better clarify the conceptual differences between ChromaGuard and previous semantic suppression approaches, as well as prior methods that exploit color-space correlations for synthetic image detection.
  2. The causal relationship between chroma guidance and improved generalization is insufficiently demonstrated. The paper shows that chroma has weak semantic association, but the connection between this property and improved synthetic image detection is not fully established. It remains unclear whether chroma improves performance because it reduces semantic reliance or simply because it provides additional useful visual information. More controlled experiments isolating these effects would strengthen the claims.
  3. Comparison with recent foundation-model-based detectors is limited. The evaluation mainly compares with existing synthetic image detectors, but stronger comparisons with recent multimodal or foundation-model-based forensic approaches would make the conclusions more convincing.
  4. Limited visualization and qualitative analysis. Although the paper provides quantitative evaluations and t-SNE visualizations, the qualitative analysis of ChromaGuard remains limited. Since the central motivation of the paper is that chroma guidance helps the model reduce semantic reliance and focus on synthesis artifacts, more visualization examples would be valuable to better understand this mechanism. For example, showing attention maps, feature responses, or comparisons between RGB-only and chroma-guided representations on challenging synthetic images would make the proposed approach more convincing and provide deeper insights into how chroma guidance improves generalization.
Questions for authors
  1. How much of the improvement comes from semantic suppression versus the additional chroma information itself?
  2. Would the chroma guidance remain effective if the backbone is not CLIP but another pretrained vision model?
  3. Can the authors provide stronger analysis explaining why chroma provides better guidance than other low-semantic image components?
  4. Could you give more visualization examples?
Reviewer_hgvtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ChromaGuard, a framework that aims to reduce semantic reliance and improve the generalization of AI-generated image detection. Based on a normalized mutual information (NMI) analysis, the authors argue that YCbCr chroma is weakly correlated with image semantics. Building on this, they construct a dual-branch network with an RGB branch (CLIP-based) and a chroma branch that interact through multi-stage bidirectional cross-attention, and introduce a DCT high-frequency adapter into the RGB branch to enhance the detection of generation artifacts. The experiments cover cross-generator generalization, semantic sensitivity, and robustness to JPEG compression.

Critical strengths

1. Good detection performance on the adopted benchmark. The method achieves an average ACC of 95.91% over the 16 generator subsets of AIGCDetectBenchmark, outperforming the compared baselines. The ablation study shows that both the Chroma-RGB interaction module and the DCT high-frequency adapter bring clear improvements, suggesting that the proposed combined framework has some practical value.

2. Addresses the important problem of semantic reliance in AI-generated image detection. The paper attempts to exploit the weak semantic correlation of chroma to guide the detector toward generation artifacts, and evaluates the method in terms of cross-generator generalization, semantic sensitivity, and compression robustness. This research problem is of practical significance.

Critical weaknesses

1. Limited methodological novelty and insufficient differentiation from prior work.

The method mainly consists of chroma feature extraction, bidirectional feature interaction, and a DCT high-frequency adapter, but the overall design overlaps substantially with existing work.

First, the bidirectional Chroma-RGB interaction module is highly similar to the feature injector and extractor in ViT-Adapter (ICLR 2023). More directly, DeepFake-Adapter (IJCV 2025) has already applied the same bidirectional interaction to forgery detection. It evenly splits a frozen ViT into 3 stages: at the beginning of each stage, cross-attention is performed with the ViT features as the query and the features of an auxiliary CNN branch as the key/value, and the result is added back to the ViT residually; at the end of each stage, the auxiliary branch is updated with its own features as the query and the ViT features as the key/value, and then passed to the next stage. Eqs. (1)–(2) of this paper are identical in form to Eqs. (5)–(6) of DeepFake-Adapter, and the 3-stage split is also the same; as far as the interaction module is concerned, the main difference is only that the input of the auxiliary branch is changed from the RGB image to the chroma magnitude. The paper neither cites nor discusses these two works.

Second, introducing frequency or color cues into CLIP also has precedents: the adapter in FatFormer (CVPR 2024) already combines spatial and frequency cues, and MCAN (AAAI 2026) combines RGB, chromaticity, and high-frequency cues on a frozen CLIP. And Table 5 shows that the gain from adding DCTHFA alone is comparable to that from adding the Chroma-RGB interaction alone, which makes the overall method closer to an engineering combination of existing components.

Therefore, the main difference of this paper seems to lie in using the chroma magnitude as the auxiliary representation; whether this choice constitutes an independent contribution depends on the hypothesis discussed in the next point, which has not yet been validated.

2. The core semantic-debiasing hypothesis is insufficiently validated.

The central claim of the paper is that chroma guidance reduces the detector's reliance on semantic shortcuts, but this claim lacks support. The NMI analysis that serves as the starting point lacks computational details and is not systematically validated: Table 6 compares three representations without any high-NMI representation as a control, and DCT, which has a lower NMI, actually yields a lower average ACC than DWT. And, table 6 lacks controls with inputs such as RGB or grayscale, so it cannot be determined whether the advantage of chroma comes from its weak semantic correlation. In addition, the evaluation scope is narrow; see Q3 for details.

Questions for authors

Q1. On the NMI analysis and the choice of chroma

  • The paper states that NMI is estimated over the 20 semantic categories of the AIGCDetectBenchmark training set. Could the authors provide more analysis details? Does the analysis also apply to the chroma magnitude that is actually used?
  • DCT has a lower NMI than DWT, yet a lower ACC. How do the authors explain this? Could the authors add, under the same architecture, controls that use high-NMI components (e.g., Cepstrum, LBP) as well as RGB and grayscale/luma as the auxiliary input?

Q2. On whether semantic reliance is actually reduced

  • The backward interaction (Eq. 2) writes CLIP features back into the chroma branch. Why is it necessary? Could the authors provide semantic probing of the chroma features before and after the interaction and of the final features, to directly verify that the model reduces semantic reliance?

Q3. On the evaluation scope

  • Could the authors report results on Chameleon, WildRF, and additional benchmarks to further validate the effectiveness of the proposed method?
  • Prior work (e.g., DDA) has pointed out that AIGCDetectBenchmark exhibits a real/fake format bias and mitigates it by applying JPEG compression (Q=96Q=96) to its synthetic images. Could the authors re-report the main results under this setting?
  • The robustness evaluation currently only covers JPEG compression on the GAN test sets. Could the authors add blur and resizing, and also cover the diffusion subsets?
Reviewer_C2xdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper identifies using semantic shortcut as a problem for synthetic image detection. Their analysis claims that YCbCr chroma is weakly related to image semantics, and therefore they propose a multi-branch architecture using RGB and chroma components. They further introduce a DCT high-frequency adapter for modeling frequency domain artifacts in the RGB branch. Their results demonstrate improved average performance across GAN-based and Diffusion-based generator detection using their approach, and they further check JPEG robustness and semantic sensitivity training across 1,2, or 4 classes.

Critical strengths
  1. The paper shows DCT High-Freq Adapter added to the baseline alone can be quite strong (Table 5 shows adapter-only Cb + Cd reaches 93.41). It’s an interesting result that might provide additional insights into the classification process, though more explanation is needed on why it performs this well.
  2. The semantic sensitivity evaluation (Table 2) shows performance gain over the closest competitor Rine. This demonstrates stable mean accuracy as the number of training classes change, which also suggests that the proposed methods are worth examining further to determine the reason for their strong performance.
Critical weaknesses
  1. One of the key claims of the paper is that the YCbCr chroma is weakly related to image semantics, and therefore it can complement traditional RGB image classification. But the analysis provided to establish this is not adequate, specially given color can plausibly carry class information depending on how its processed. The analysis provided calculates normalized mutual information, but further information is needed to establish how this is calculated, how this reasonably proves the weak semantic information in the YCbCr chroma, and whether this is actually being used by the multi-branch architecture. Establishing these ideas are central to the paper’s claim. The proposed method also utilizes a DCT-high-frequency adapter in their RGB branch, so it further raises the question about the necessity of the Chroma information. Table 5 results establish that the DCT adapter by itself is quite strong.
  2. On Table 5, the inclusion of Ci and Cd creates a strong boost in Midjourney (19.76 points over Cb + Ci) and DALLE2 (7.35 points over Cb + Ci), which significantly contributes to the performance of the combined model getting 1.95 points higher mean accuracy over the closest single-component variant. It is important to analyze why their combined approach is giving such a strong boost and increasing the gap between the individual components, and demonstrate how this is different from seed/variance noise.
Questions for authors
  1. Can the authors clarify how the NMI is computed and how it supports YCbCr chroma has weak semantic information? Comparing it with the traditional RGB image would also help with establishing the difference between these two types of input.
  2. Can the authors demonstrate whether the trained model actually uses Chroma information for its classification or how it makes it less reliant on semantics? What happens if the Chroma input is zeroed, or replaced with another image’s Chroma?
  3. Can the authors provide the mean +- std of Table 5 across multiple seeds?

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

Related papers

Discussion (0)

Sign in to comment.