Rethinking supervision for AI image forensics

Less Supervision,
Better Generalization:Weakly Supervised Fake Region Localization
in Diffusion-Edited Images

Junhee Lee1 Donghyeon Jeon1 Taeoh Kim2 Beomyoung Kim2 MyeongAh Cho1,*
1Kyung Hee University2NAVER Cloud

* Corresponding author

PaperarXivCode

What matters is not just how much we supervise,
but what our supervision tells the model.

Start with the question

01 / The changing annotation landscape

More labels used to mean more effort.
What happens when that changes?

The conventional choice

Annotate in detail.

Dense annotations provide explicit spatial targets for localization. But as data grows, producing these annotations becomes costly and difficult to scale.

Detailed labels · substantial annotation effort
The practical alternative

Learn with less.

Weakly supervised learning reduces this burden by learning spatial predictions from coarser labels, such as one real/fake label for an entire image.

Coarse labels · more accessible training data
The new question

What if labels are automated?

MLLMs, VLMs, and other AI tools are changing how annotations can be produced. Even when dense labels become easier to obtain, are they necessarily the right training targets?

Label availability ≠ label suitability

The central question

If dense supervision is easy to obtain,
do we still need weak supervision?

Should we always prefer full supervision in the age of automatic annotation?

A. Not necessarily. More detailed labels can still teach the wrong thing.

02 / The answer starts with the labels

More supervision is not
always better supervision.

In diffusion-based image editing, artifacts can extend beyond the intended edit area, meaning dense supervision can guide the model with misleading signals and hurt generalization. We introduce ReGFLoW, the first weakly supervised framework for diffusion-edited fake region localization, using only image-level real/fake labels while achieving stronger performance across unseen generators.

ReGFLoW is introduced as the first weakly supervised framework for diffusion-edited fake region localization in generic visual content. The strongest gains appear when evaluation includes both partially edited and fully synthetic images, and in cross-dataset transfer.

Figure 2: guidance mask, VAE encoding and decoding, and boundary blending can produce traces outside the intended edit region. View full resolution
Figure 2 The intended edit region is not the full extent of diffusion artifacts. Latent processing and blending can introduce weak global or boundary-localized traces beyond the guidance mask. This is a conceptual illustration, not an exhaustive artifact ground-truth map. Paper, p. 4 ↗
What the mask describes

The region we intended to edit.

What the image can contain

Diffusion traces inside and outside that region.

What dense labels can enforce

Out-of-mask artifacts are trained as authentic.

Sometimes, less supervision can lead to better generalization.

Automatic annotation can change the cost of obtaining labels. It does not, by itself, resolve a mismatch between what a label represents and what the model needs to learn.

03 / Why a new formulation is needed

Less supervision.
Not just the same assumptions.

Removing pixel masks is not enough. Diffusion artifacts may be non-semantic, weakly bounded, and spatially diffuse—so object-based expansion and boundary-aligned pseudo-masks are not reliable defaults.

A first step for this task

The first weakly supervised framework for diffusion-edited fake region localization in generic visual content.

ReGFLoW learns from image-level real/fake labels and asks where the artifact evidence is, rather than assuming an object-shaped target. To the best of our knowledge; prior weakly supervised work addresses conventional manipulations or face-centric diffusion edits.

Figure 1: conventional segmentation follows object boundaries, diffusion edits may not, and ReGFLoW instead uses a diffusion prior and artifact-centric MIL.View full resolution ↗
Figure 1 A different task calls for different spatial guidance. Conventional localization targets often follow objects or boundaries. ReGFLoW instead uses reconstruction-guided multiple instance learning to identify diffusion artifact evidence. Paper, p. 2 ↗
From semantic similarity as the default priorTo reconstruction-guided artifact evidence

04 / The framework

Meet ReGFLoW.

Reconstruction-Guided Fake Localization under Weak supervision

ReGFLoW uses diffusion reconstruction cues to guide localization in both feature and score spaces, then learns artifact evidence with an artifact-centric MIL objective.

Figure 4: frozen VAE reconstruction, global and local encoders, gated feature fusion, a reconstruction logit bias, artifact-centric MIL, and a localization-guided classifier.Explore the framework ↗
Figure 4 Dense spatial guidance without dense target masks. Reconstruction cues enter both the feature and score spaces. Artifact-centric MIL trains the patch map, which supports fake-region localization and image-level classification. Paper, pp. 5–7 ↗
1

Reconstruction prior

A frozen diffusion VAE reconstructs the image, and the reconstruction error becomes a dense spatial prior for where artifact evidence may lie.

r = |I - D(E(I))|
2

Feature and score guidance

The reconstruction prior is fused with local features and also biases the patch logits, so the model is nudged toward transferable artifact cues rather than object boundaries.

ᵝ = Floc + tanh(γ) · Δ
3

Artifact-centric MIL

Weak supervision is applied through a MIL objective that highlights suspicious patches and ranks fake-image evidence above real-image evidence.

ŝfake > ŝrealSee the paper for the full objective and calibration details.
Where to look for more details

The main paper introduces the framework at a high level. The appendix provides additional details on mask availability in real-world diffusion editing, auxiliary losses and inference calibration, implementation details, additional qualitative results, limitations, and related work.

For full equations, training details, and calibrated inference, please refer to the paper and appendix.

Read the abstract

Localizing AI-edited regions is essential for interpretable forensic analysis, but remains challenging due to subtle and spatially distributed artifacts that are misaligned with semantic or object boundaries. Existing approaches rely on pixel-level supervision from controlled editing pipelines, which is difficult to scale and can introduce misleading signals: artifacts frequently extend beyond annotated regions, while out-of-mask pixels are treated as authentic. This limits models’ ability to capture transferable evidence and generalize across generators and datasets. To address these issues, we propose ReGFLoW, a Reconstruction-Guided Fake Localization framework under Weak supervision, which is the first weakly supervised approach for diffusion-edited fake region localization. ReGFLoW requires only real/fake labels at the image-level and uses diffusion reconstruction errors as dense spatial guidance to inject them into both feature and score spaces. Furthermore, by artifact-centric multiple instance learning, ReGFLoW utilizes localized diffusion evidence without relying on semantic-affinity or boundary-based pseudo-mask priors. Extensive experiments show competitive cross-generator localization, while ReGFLoW outperforms all evaluated fully supervised baselines when evaluation includes both partially edited and fully synthetic images and in cross-dataset tests, without target-domain adaptation.

Original abstract ↗

05 / Evidence, not just an intuition

Less supervision.
Better generalization where it matters.

Train on SD1.5 and evaluate on unseen generators and datasets. ReGFLoW never uses pixel-level training masks or target-domain adaptation.

P+F · OOD localization F131.3%+2.2 points over MaskCLIP (29.1)Table 2 ↗ COCO-GLIDE · pixel F142.01%+16.84 points over TruFor (25.17)Table 3 ↗ Image-level · OOD detection F176.22%Best among the evaluated methodsTable 4 ↗
Read the settings. P: partially edited images only.P+F: partially edited + fully synthetic images.OOD Avg.: average over SD2.1, SDXL, SD3, and Flux.1.

Table 1 / OpenSDID · P setting

Competitive cross-generator localization.

Pixel-level F1 / IoU ↑

ReGFLoW achieves 38.6 OOD average F1, outperforming the adapted weakly supervised baselines and several fully supervised methods. On Flux.1 it obtains 29.2 F1, compared with MaskCLIP’s 18.4. MaskCLIP retains the best P-setting OOD average F1 (41.7); ReGFLoW does not lead every generator or metric.

Table 1. Partially edited images (P): pixel-level F1 and IoU
MethodIn-domainCross-domainAverage
SD1.5SD2.1SDXLSD3Flux.1Avg.OOD Avg.
F1IoUF1IoUF1IoUF1IoUF1IoUF1IoUF1IoU
Fully supervised methods
PSCC-Net[52] 64.2 54.7 44.8 36.7 26.1 19.7 37.3 29.3 11.6 8.2 36.8 29.7 29.9 23.5
TruFor[19] 71.0 63.4 61.9 54.7 31.9 26.5 38.5 32.2 9.7 7.6 42.6 36.9 35.5 30.3
IML-ViT[18] 73.6 66.5 50.6 44.8 26.0 21.5 28.3 23.6 7.9 6.1 37.3 32.5 28.2 24.0
MaskCLIP[9] 75.8 68.6 63.2 56.1 35.2 29.5 49.8 42.7 18.4 14.8 48.5 42.4 41.7 35.8
Weakly supervised methods
WSCL[21] 25.6 16.7 25.9 17.0 26.0 17.0 25.4 16.5 25.3 16.4 25.6 16.7 25.7 16.7
DOLOS[22] 21.5 13.4 19.1 11.9 19.8 12.2 20.3 12.5 19.2 12.0 20.0 12.4 19.6 12.2
BoxPromptIML[44] 11.5 7.0 10.9 6.5 10.9 6.6 13.8 8.4 10.3 6.2 11.5 6.9 11.5 6.9
ReGFLoW(Ours) 49.3 36.9 46.3 33.8 34.9 23.6 43.9 30.9 29.2 19.0 40.7 28.8 38.6 26.8
Shaded row: ReGFLoW. Bold: best in each numeric column across all displayed methods. All values are as reported.Paper Table 1 ↗

Table 3 / Cross-dataset evaluation

Transfer across datasets. Without adaptation.

Pixel-level F1 ↑

ReGFLoW obtains the highest F1 in all three reported cross-dataset settings: COCO-GLIDE (P), DOLOS (P), and DOLOS (P+F). These tests extend the evaluation beyond OpenSDID to different generation and content distributions.

Table 3. Cross-dataset localization: pixel-level F1
MethodCOCO-GLIDE[53]DOLOS[22]
PPP+F
Fully supervised methods
TruFor[19] 25.17 8.21 12.70
IML-ViT[18] 15.23 1.16 1.74
MaskCLIP[9] 8.34 13.35 14.65
Weakly supervised methods
ReGFLoW(Ours) 42.01 21.09 21.19
Methods are rows; datasets and settings are columns. COCO-GLIDE reports P only; DOLOS reports P and P+F.Paper Table 3 ↗

06 / Understanding the evidence

What makes it work—and what scales?

Image-level detection, module ablations, and source-domain diversity provide complementary evidence for reconstruction-guided weak supervision.

Table 4 / Image-level detection

Localization evidence helps detection, too.

F1 / Accuracy ↑

By combining high-response regions with a global image embedding, ReGFLoW achieves 76.22 OOD F1 and 80.93 OOD accuracy—the highest averages among the evaluated detectors.

Table 4. Image-level detection: F1 and accuracy
MethodIn-domainCross-domainAverage
SD1.5SD2.1SDXLSD3Flux.1Avg.OOD Avg.
F1ACCF1ACCF1ACCF1ACCF1ACCF1ACCF1ACC
CNNDet[54] 84.60 85.04 71.56 75.94 59.70 68.72 56.27 67.08 35.72 57.57 61.57 70.87 55.81 67.33
GramNet[55] 80.51 80.35 74.01 76.66 65.28 70.76 64.35 70.29 52.00 63.37 67.23 72.29 63.91 70.27
FreqNet[56] 75.88 77.70 60.97 68.37 53.15 64.02 53.50 64.37 38.47 57.08 56.39 66.31 51.52 63.46
NPR[57] 79.41 79.28 81.67 81.84 72.12 74.28 73.43 75.47 67.62 71.36 74.85 76.45 73.71 75.74
TruFor[19] 90.12 97.73 35.93 55.62 58.04 66.41 59.73 67.51 49.12 61.62 58.59 69.78 50.70 62.79
IML-ViT[18] 94.47 75.73 69.70 61.19 40.98 49.95 44.69 51.25 18.20 43.62 53.61 56.35 43.39 51.50
MaskCLIP[9] 91.28 91.47 83.36 85.28 72.90 77.98 67.79 74.95 53.08 67.12 73.68 79.36 69.28 76.33
ReGFLoW(Ours) 93.02 93.08 91.34 91.73 79.89 82.63 76.87 80.62 56.78 68.76 79.58 83.36 76.22 80.93
Numeric values and reference numbers follow the paper. The original “Approach” grouping is omitted here; ReGFLoW is trained with image-level labels only (Section 4.1).Paper Table 4 ↗

Table 5 / Ablation study

Both the prior and the learning objective matter.

Pixel-level F1 ↑

Removing the reconstruction logit bias lowers F1 from 49.38 to 46.06. Removing the reconstruction prior entirely lowers it to 38.08; score-pooling BCE yields 32.34. The full design combines feature guidance, score guidance, and artifact-centric MIL.

Table 5. Ablation study: pixel-level F1
MethodBias
Predictor		Feature-aligned
Fusion		Artifact-Centric
MIL		Pixel F1
ReGFLoW(Ours) ✓ ✓ ✓ 49.38
w/o Bias Predictor × ✓ ✓ 46.06
w/o Reconstruction Prior × × ✓ 38.08
Score-Pooling BCE × × × 32.34
Values retain the precision reported in this ablation. They are not substituted for the differently reported Table 1 values.Paper Table 5 ↗

Table 6 / Broader training domains

Broader sources. Not just more images.

Fixed training-set size

Adding one source domain to SD1.5 improves mean F1 on the remaining unseen domains while holding the total training size fixed. Weak supervision can therefore support broader-source training without requiring pixel masks from every source.

Table 6. Source-domain diversity at fixed total training size
TrainSD2SD3SDXLFLUXAvg. Δ
SD1.5 40.96 38.09 29.36 24.12 –
+ SD2 – 42.34(+4.25) 33.73(+4.37) 28.71(+4.60) +4.40
+ SD3 42.75(+1.79) – 34.48(+5.12) 30.96(+6.84) +4.58
+ SDXL 40.88(-0.08) 41.84(+3.75) – 31.00(+6.88) +3.52
+ FLUX 41.57(+0.61) 43.82(+5.73) 34.47(+5.11) – +3.82
“–” excludes a domain used for training. Parentheses show changes from the SD1.5-only row. Reported Avg. Δ values and source domain names are preserved without recomputation.Paper Table 6 ↗

07 / Robustness

Beyond pristine images.

ReGFLoW degrades gradually under stronger Gaussian blur and remains competitive under JPEG compression on SD3 and SDXL.

Figure 5: four line plots compare pixel F1 under Gaussian blur and JPEG compression on SD3 and SDXL.View plots at full resolution ↗
Figure 5 Robustness to image degradations. Baseline curves are digitized from the robustness results reported in OpenSDI; ReGFLoW is evaluated under the same degradation settings. Paper, p. 10 ↗

08 / Qualitative results

Look for artifacts.
Not just object-shaped regions.

Out-of-domain comparisons include both boundary-aligned edits and boundary-unaligned changes—background edits, partial object changes, and multiple manipulated regions.

Figure 6: inputs, benchmark ground-truth masks, MaskCLIP, TruFor, and ReGFLoW predictions for boundary-aligned and boundary-unaligned edits.Inspect the comparisons ↗
Figure 6 Localization when boundaries offer limited guidance. MaskCLIP and TruFor are fully supervised; ReGFLoW is weakly supervised. The GT row shows the benchmark masks used for evaluation. Paper, p. 10 ↗

These qualitative examples illustrate the paper’s hypothesis. They do not establish a universally correct ground-truth map of every diffusion trace outside an intended edit mask.

09 / Appendix

The paper PDF also includes an appendix.

Beyond the main paper, the appendix discusses why pixel masks are hard to scale in real diffusion-editing settings, details the auxiliary losses and inference calibration, provides implementation details, summarizes assets and licenses, shows additional qualitative results, and discusses limitations and related work.

Appendix in the PDF ↗

10 / The larger takeaway

Before asking for more supervision,
ask whether it is the right supervision.

The best choice depends on the task, the data, and what the labels actually represent. A dense label can be precise about the intended edit while remaining incomplete about the artifacts a detector should learn.

Diffusion-edited localization offers a concrete case study: weak supervision is not merely a cheaper substitute for pixel masks. With an appropriate artifact-centric formulation, it can be a better route to generalization.

Sometimes, less supervision
can lead to better generalization.

A task-dependent lesson—not a claim that weak supervision always wins.

11 / Citation

Build on this work.

Use the citation below. Update the archive or publication fields when the final public links are available.

BibTeX
@misc{lee2026less,
  title  = {Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images},
  author = {Junhee Lee and Donghyeon Jeon and Taeoh Kim and Beomyoung Kim and MyeongAh Cho},
  year   = {2026},
  note   = {Manuscript}
}

Paper figure
Enlarged paper figure

Open original ↗