MIMLv2: Webly-Supervised Image Manipulation Dataset

• MIMLv2 is a large-scale, webly supervised dataset for image manipulation localization with pixel-level annotations gathered from authentic forged images.
• It employs a category-aware auto-annotation pipeline (CAAAv2) combining SPG/SDG classification, DASS, and Corr-DINO to generate accurate pixel masks.
• The dataset enhances forensic model training through Quality Evaluation Score filtering and Object Jitter augmentation, significantly boosting IoU and F1 performance.

Searching arXiv for the primary paper to ground the article in the cited source. MIMLv2 is a large-scale, webly-supervised dataset for image manipulation localization (IML) built to directly address the chronic data scarcity that has limited progress in localizing forged regions at the pixel level. It is constructed from manually forged images collected from the web and paired with automatically derived pixel-level mask annotations produced by a category-aware constrained image manipulation localization (CIML) pipeline. In the reported formulation, MIMLv2 contains 246,212 forged images with pixel-level mask annotations and 63,847 originals, and is presented together with CAAAv2, the Quality Evaluation Score (QES), Object Jitter, and the Web-IML model as components of a web-scale supervision framework for real-world forgery localization (Qu et al., 28 Aug 2025).

1. Motivation and problem setting

The dataset is motivated by two linked constraints in IML. First, traditional IML datasets are tiny because crafting realistic manipulations and labeling pixels is labor-intensive. Second, synthetic data often exhibits a domain gap relative to real forgeries. The consequence, as stated in the source description, is that models overfit, struggle to generalize, and are brittle to real-world image degradations (Qu et al., 28 Aug 2025).

MIMLv2 is framed as a response to this bottleneck through webly supervised learning. Rather than relying primarily on handcrafted datasets, it harvests manually forged images circulating on the web and distills pixel-level supervision from a simpler auxiliary task, constrained image manipulation localization. The stated objective is not merely to enlarge the training pool, but to teach forensic models to learn semantic-agnostic manipulation cues rather than overfit shallow semantics or synthetic artifacts.

This positioning distinguishes MIMLv2 from older handcrafted datasets such as CASIA and from smaller modern resources such as IMD20. The source explicitly states that MIMLv2 is approximately 120× larger than IMD20, which contains 2,010 forged images, and roughly doubles MIMLv1, which contains 123,150 forged images. It is also described as far more modern and diverse than older handcrafted datasets.

2. Dataset composition, provenance, and annotation format

MIMLv2 contains manually forged images collected from imgur.com, paired with corresponding originals where available. The images are described as authentic human manipulations created in the wild, including splicing, copy-move, removal, subtle edits, and AIGC-style manipulations. The resolution range is broad, with height and width in $[(45, 120), (13846, 9200)]$, and the released annotations are binary pixel-level masks aligned to image coordinates (Qu et al., 28 Aug 2025).

Item	Value
Forged images	246,212
Originals	63,847
Source	imgur.com
Annotation type	Per-image pixel-level binary masks
Resolution range	Height, width $\in [(45, 120), (13846, 9200)]$
Manipulation coverage	Copy-move, splicing, removal, and AIGC-style manipulations

The dataset is intended primarily for web-scale supervision and pretraining rather than as a benchmark with fixed canonical splits. The authors do not define fixed official splits for MIMLv2; models sample uniformly in training. Evaluation is instead conducted on external benchmarks, including CASIAv1/v2, Coverage, NIST16, IMD20, CocoGlide, CIMD, and MISD. To avoid leakage, images appearing on those benchmarks are removed from MIMLv2.

Quality control includes removal of duplicates via MD5/pHash, exclusion of evaluation-set overlaps, and QES filtering of auto-annotations. This suggests that dataset curation is treated as an explicit component of the supervision pipeline rather than an incidental preprocessing step.

3. CAAAv2 and the category-aware auto-annotation pipeline

The central annotation mechanism is Category-Aware Auto-Annotation v2 (CAAAv2), an upgraded, category-aware CIML annotation system designed around two classes of original–manipulated image pairs. The Shared Donor Group (SDG) denotes cases where the forged image contains foreground content copied from the original image, so the shared content is the foreground object. The Shared Probe Group (SPG) denotes cases where the forged image is a direct modification of the original, so the shared content is largely the background and the images are roughly spatially aligned (Qu et al., 28 Aug 2025).

CAAAv2 has three stages.

First, it performs self-supervised SPG/SDG classification. A pair-wise classifier is trained without manual labels by constructing SPG pairs through mild manipulations of the original image, including copy-paste within the same image, splicing, removal, color jitter, and compression, and constructing SDG pairs by copy-pasting 1–3 random regions from one image into a different image. The classifier decides whether the two images are mostly identical, corresponding to SPG, or visually dissimilar, corresponding to SDG.

Second, it applies Difference-Aware Semantic Segmentation (DASS) to SPG pairs. The rationale is that, in SPG pairs, the absolute difference between forged and original typically highlights manipulated regions, but network-transmission noise and camera pipeline variations make raw differences noisy. DASS concatenates authentic image $I_a$, forged image $I_b$, and the absolute difference map $|I_a - I_b|$ along the channel dimension and feeds the result into a semantic segmentation network. A VAN encoder extracts semantics, and the decoder includes a Multi-Aspect Denoiser to suppress noise while retaining genuine difference signals relevant to the tampered area. The stated outcome is a clean pixel-level mask of manipulated regions for SPG pairs.

Third, it applies Correlation DINO (Corr-DINO) to SDG pairs. This branch addresses cases in which correlation must recover shared foreground objects across diverse scales and complex backgrounds. The pipeline uses a frozen DINOv2 ViT backbone to encode both images and produce last-layer features. It then computes cross-image and self-correlation features, uses learnable 1×1 convolutions plus channel-wise statistics for aggregation, performs feature super resolution by fusing aggregated correlations with DINOv2 features from the last 2–4 layers and interpolating to multiple scales, and finally applies a Multi-Aspect Denoiser with top-down feature fusion and dilated convolutions. Training uses end-to-end pixel-wise cross-entropy on synthetic data, since ground-truth masks are available only for synthetic pairs and small handcrafted sets, not for web images.

The description explicitly states that CAAAv2 produces more accurate masks than the CVPR-2024 conference version, CAAA, enabling stricter quality filtering while retaining more samples and thereby increasing dataset size and diversity.

4. Quality control through QES and supervision enrichment through Object Jitter

After auto-annotation, SPG predictions are described as typically sharp and confident, whereas a subset of SDG predictions requires filtering. The proposed Quality Evaluation Score captures confidence concentration and edge sharpness without ground truth. Let $p_{i,j}$ be the predicted probability at pixel $(i,j)$, normalized to $[0,1]$. QES is defined as the ratio between the number of pixels with $p > 1 - T_h$ and the number of pixels with $p > T_l$, with $\in [(45, 120), (13846, 9200)]$0, and a sample is kept if $\in [(45, 120), (13846, 9200)]$1 (Qu et al., 28 Aug 2025).

The reported effect of this filtering is explicit: increasing the QES threshold raises average IoU while reducing retained sample count. At threshold $\in [(45, 120), (13846, 9200)]$2, QES filtering improves SDG mask IoU from $\in [(45, 120), (13846, 9200)]$3 to $\in [(45, 120), (13846, 9200)]$4; higher thresholds can push IoU beyond $\in [(45, 120), (13846, 9200)]$5 while dropping more samples. The threshold study is described as monotonic in this trade-off, and threshold $\in [(45, 120), (13846, 9200)]$6 is presented as a strong balance.

Object Jitter is introduced as a complementary source of supervision. Instead of synthetic splicing that creates unrealistic content, it generates additional training samples with subtle, semantically plausible artifacts using authentic images. Object masks are obtained via SAM or EfficientViT-SAM, and 1–3 random operations are applied per image: Size Jitter, defined as slight enlargement of the object region with scale approximately $\in [(45, 120), (13846, 9200)]$7–$\in [(45, 120), (13846, 9200)]$8; Exposure Jitter, defined as mild overexposure of the object region; and Texture Jitter, defined as slight texture alteration via JPEG compression, reverse JPEG, and blur. Edge transitions are then smoothed to minimize obvious seams.

The released description states that these operations complement MIMLv2 by reducing over-reliance on exposure and texture cues, with size jitter particularly impactful, and that the approach is robust to image source choice, specifically COCO versus SA-1B. A plausible implication is that Object Jitter functions as an artifact-space regularizer rather than as a substitute for real manipulations.

5. Web-IML and empirical performance with web-scale supervision

Web-IML is the forensic model introduced to exploit MIMLv2 and Object Jitter. Its architecture consists of a ConvNeXt-Base encoder, a Multi-Scale Perception (MSP) module that fuses global-to-local features through pooling at multiple scales and top-down fusion with lower layers, and a Self-Rectification (SR) module that iteratively refines masks by analyzing initial predictions, mapping them to rectification features, and applying nested channel attention with multi-dilation filtering (Qu et al., 28 Aug 2025).

Training uses pixel-wise cross-entropy loss:

$\in [(45, 120), (13846, 9200)]$9

The reported training details are: input size $I_a$0, batch size $I_a$1, AdamW optimization, learning rate linearly decayed from $I_a$2 to $I_a$3, and $I_a$4 iterations. Baseline training sets are CASIAv2 and tampCOCO, and web supervision is added by incorporating MIMLv2 and Object Jitter samples with uniform sampling across datasets while keeping total volume constant when comparing different supplements.

Evaluation uses Intersection-over-Union and binary F1. IoU is defined as

$I_a$5

where $I_a$6 is the predicted binary mask and $I_a$7 is the ground-truth mask. Benchmarks include CASIAv1, Coverage, NIST16, IMD20, CocoGlide, CIMD, and MISD, together with robustness tests under resizing, blur, and JPEG compression, and downstream document IML on SACP.

The headline result is that Web-IML trained with web supervision, specifically MIMLv2 plus Object Jitter, achieves average IoU $I_a$8 and F1 $I_a$9, representing a $I_b$0 improvement over the baseline training without web supervision and surpassing TruFor by $I_b$1 average IoU points. The source further states that, compared to the conference model APSC-Net, Web-IML is smaller and consistently better, and that web supervision further lifts both models.

Ablation findings are also specified. MSP yields $I_b$2 IoU points over baseline, SR yields an additional $I_b$3, and nested attention together with a second rectification pass further improves accuracy. MIMLv2 outperforms MIMLv1 as a training supplement. DASS outperforms difference baselines on SPG, and Corr-DINO with a frozen DINOv2-B backbone beats trainable CNN and ViT backbones because overfitting is reduced. Robustness is reported as high AUC under resizing at $I_b$4 and $I_b$5, Gaussian blur with $I_b$6 and $I_b$7, and JPEG compression with quality $I_b$8–$I_b$9, with AUCs up to $|I_a - I_b|$0 on NIST16 JPEG 100.

6. Limitations, practical use, and significance

The limitations are stated in operational rather than speculative terms. Remaining annotation noise persists: despite QES, some SDG predictions can be imperfect in extreme cases, including tiny objects, severe occlusions, or unusual composites. A domain gap may remain because web data may underrepresent certain forgeries, such as highly technical forensic attacks or extremely subtle color manipulations with no structural cues. The type distribution is also incompletely specified, since precise counts by manipulation type are not provided; AIGC may be less frequent, and the model might underperform on rare manipulations without Object Jitter supplementation (Qu et al., 28 Aug 2025).

Practical usage guidance is explicit. The repository URL is given as https://github.com/qcf-568/MIML, and license and usage terms are to be consulted there. Recommended use is to pretrain an IML model on CASIAv2 plus tampCOCO plus MIMLv2 and, if possible, integrate Object Jitter as an online augmentor. Uniform sampling across datasets is recommended to maintain balanced exposure to different artifact regimes. Input resizing to $|I_a - I_b|$1 with standard image augmentations and binarization of predictions at threshold $|I_a - I_b|$2 are also specified for evaluation. Web-IML is reported to have approximately $|I_a - I_b|$3 parameters, and training for $|I_a - I_b|$4 iterations with batch size $|I_a - I_b|$5 is said to require modern GPU resources.

The broader significance assigned to MIMLv2 is that it shifts IML practice away from exclusive reliance on small, expensive handcrafted datasets toward abundant and growing web resources, while still providing high-quality pixel annotations through CAAAv2 and QES. The same description extends this significance to downstream transfer, including document tampering, and emphasizes that the construction pipeline is scalable: SPG/SDG classification, DASS/Corr-DINO annotation, QES filtering, and Object Jitter augmentation can continually expand the dataset as new forgeries appear online. This suggests a model of IML data construction in which annotation quality control and supervision scaling are treated as a single integrated system rather than separate stages.