acceptodds
Withdrawn from ICLR 2027

Relation Conflict Estimation for Generalizable Face Forgery Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generalizable face forgery detection remains challenging because models trained on known manipulations often overfit to specific appearance artifacts, which may not transfer to unseen forgery mechanisms. However, the discriminative evidence carried by facial relation conflicts remains insufficiently explored. We observe that authentic facial videos tend to preserve stronger compatibility between facial patch representations and their surrounding contexts, whereas forgery synthesis may introduce subtle contextual discrepancies even when local appearances remain visually plausible. Motivated by this observation, we propose a Relation Conflict Estimation framework for generalizable face forgery detection, which exploits contextual discrepancies among facial regions as transferable forgery evidence. Lightweight distributional guidance is further introduced to regularize relation representations in terms of frame-level conflict magnitude and cross-frame distribution. Extensive experiments under cross-dataset, cross-manipulation, and cross-generation protocols demonstrate that the proposed method consistently improves generalization across both manipulation-based and AIGC-generated forgeries, with consistent gains observed on ViT-B/16 and CLIP ViT-L/14 backbones. Ablation and visualization analyses further show that pairwise relation-conflict separation provides the primary regularization effect, while cross-frame distribution guidance offers complementary improvements.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Face Forgery Detection, Generalizable Deepfake Detection, Relation Conflict Estimation, Cross-Domain Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_iEqYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Relation Conflict Estimation framework for generalizable face forgery detection. The main idea is to model the discrepancy between facial patch representations and their context-induced relation responses, together with pairwise relation-conflict separation and cross-frame distribution regularization. However, the experimental comparison is not sufficiently up to date for an ICLR 2027 submission. In particular, the main cross-dataset evaluation compares only against methods published up to 2025. The cross-operation evaluations only include a TIFS2026 baseline, which is actually work from one or two years ago based on the publication speed of the journal. As a result, the paper does not convincingly establish the effectiveness of the proposed method against the current state of the art. Furthermore, the paper lacks theoretical justification or direct empirical comparisons to prove that the claimed 'relation conflict' captures anything fundamentally different from ordinary patch-level feature discrepancy, which has already been extensively explored in prior works.

Critical strengths

1.The paper addresses the important problem of generalizable face forgery detection under unseen datasets, manipulation methods, and generation mechanisms.

2.The idea of using contextual discrepancies between facial patches and their surrounding regions is intuitive and potentially complementary to local appearance artifacts.

Critical weaknesses

1.The main experimental comparisons use outdated and incomplete baselines. In Table 1, the cross-dataset evaluation does not include any method published after 2025. In Table 2, the cross-manipulation evaluation includes only one TIFS2026 baseline, MAP-Mamba, while most other baselines are from 2024 or earlier.

2.The core novelty of Relation Conflict Estimation is not sufficiently distinguished from existing contextual and token-level modeling methods. The proposed framework combines a 3×3 local neighborhood, the class token, cross-attention, cosine discrepancy, a margin-based separation loss, and entropy regularization. These components are relatively standard individually.

3.The paper repeatedly states that the method operates “without extra inference-time computation.” However, the relation estimation and relation-aware prediction pathways are still retained during inference. Table 6 in the appendix also shows increases in both the number of parameters and FLOPs.

Questions for authors

1Why does the paper not compare against more recent baselines?

2The pape fundamentally relies on computing cosine similarity and applying a separation loss to measure differences. Both of these are common solutions. What is the strict theoretical or direct empirical distinction between the claimed "relation conflict" and the already well-explored "patch discrepancy"?

Reviewer_jMyrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This manuscript present a generalizable face forgery localization work by modeling facial relation conflict estimation. The motivation is mainly based on the observation that real facial videos have stronger compatibility between facial patch representations and their surrounding contexts than fake videos. Accordingly, the authors proposed a Relation Conflict Estimation framework to exploit contextual discrepancies among facial regions to perform more generalizable face forgery detection. Experiments demonstrate the effectiveness of the designed components under various settings.

Critical strengths

None

Critical weaknesses

The proposed method is motivated by the observation that real facial videos exhibit stronger compatibility between facial patch representations and their surrounding contexts than fake videos. However, no statistical evidence is provided to support this observation, making the motivation behind the proposed framework less convincing. Furthermore, the proposed relation conflict estimation and distributional guidance modules lack sufficient methodological novelty.

Some experimetal limitations have been identified. Tables 1 and 2 appear to report video-level face forgery detection performance. However, the authors compare their video-level detection method with several frame-level detection baselines, potentially compromising the fairness of the comparisons. Moreover, the baseline methods are not sufficiently up to date, with the most recent baseline in Table 1 appearing to have been published at ICML 2025. In the image-level evaluation presented in Table 7, the proposed method is not compared with state-of-the-art approaches. Additionally, Tables 1 and 2 do not report the performance of the vanilla ViT-B and CLIP-L baselines for reference. Further ablation studies on hyperparameter selection are also expected to better justify the design choices and strengthen the experimental validation. Overall, the manuscript lacks sufficient methodological novelty and rigorous experimental comparisons. Therefore, I do not believe that the manuscript meets the acceptance standards of ICLR.

Questions for authors

Please see the critical weaknesses.

Reviewer_Uy9TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RCE. For each patch token, a local–global context is formed from its 3×3 neighborhood and the class token; cross-attention produces a context-induced relation response, injected back into the class token via a second attention layer for classification. Two training-time regularizers are added: a pairwise separation loss encouraging authentic videos to exhibit lower frame-averaged relation conflict than forged ones by a margin, and a cross-frame distribution regularizer maximizing the normalized entropy of frame-level conflicts for authentic videos only. Trained on FF++ (c23) and evaluated under cross-dataset (CDFv2/DFDC/DFDCP), cross-manipulation (eight unseen face-swapping pipelines), and cross-generation (two GAN, four diffusion) protocols, it reports consistent gains on ViT-B/16 and CLIP ViT-L/14, reaching 93.98% average video-level AUC with CLIP-L.

Critical strengths
  1. Broad evaluation coverage. The experiments cover cross-dataset, cross-manipulation, and cross-generation generalisation, reporting both video-level and frame-level AUC. In particular, the inclusion of four DF40 diffusion subsets—DDIM, RDDM, DiT, and SiT—extends the evaluation to challenging generation methods that are only partially covered by many contemporaneous studies.
  2. Largely consistent numerical reporting and transparent computational costs. The paper explicitly reports the additional computational cost: for CLIP-L, Table 6 shows an increase of 17.86M parameters (5.86%) and 4.17G FLOPs (5.15%).
  3. A straightforward, visual-only design. The method uses only the visual encoder of CLIP ViT-L/14, without text supervision or additional modalities. This keeps the approach relatively simple and supports comparisons with Effort and Forensics Adapter using comparable visual backbones.
Critical weaknesses
  1. Insufficient differentiation from the closest prior work. The central idea of modelling contextual inconsistency among visual tokens appears closely related to UIA-ViT (Zhuang et al., ECCV 2022), which performs unsupervised inconsistency-aware modelling over ViT patches, as well as self-consistency learning (Zhao et al., ICCV 2021) and IID (Huang et al., CVPR 2023). However, these methods appear only in grouped citations in Section 2.1, without a substantive methodological comparison. Given that contextual inconsistency modelling is central to the claimed contribution, the paper should clarify what RCE adds beyond these approaches—particularly UIA-ViT—in terms of context construction, relation modelling, and training objectives. Without this comparison, it is difficult to assess whether RCE introduces a substantively different mechanism or primarily refines an existing line of work. 2.The sign convention in Fig. 1 contradicts Eq. (7). The figure states L_sep = max(0, m + E_f − E_r), which drives the fake samples' confusion below that of the real samples — exactly the opposite of what the paper argues throughout.
Questions for authors

1.For the already-trained w/o L_sep variant, do the E^r and E^f distributions remain separated? If not, how is the claim that relation conflict constitutes intrinsic, transferable evidence supported?

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

Related papers

Discussion (0)

Sign in to comment.