Learning to Falsify:
Hierarchical Multiple Instance Learning for
High Resolution AI-Generated Image Detection

ACM Multimedia 2026

Haoran Chen¹Daiqi Gao¹Xinyuan Li¹Yongwei Wang²Lei Zhu³Rui Cai¹Jianfeng Dong^1,4,*

¹ Zhejiang Gongshang University² Zhejiang University³ HKUST (Guangzhou)⁴ Zhejiang Key Laboratory of Big Data and Future E-Commerce Technology

* Corresponding author

Paper arXiv Code

Demo

The LTF framework: frozen CLIP, patch-level and tile-level MIL, and Cross-Shuffle during training. — **Overview of LTF.** Local evidence is aggregated at two spatial levels. Cross-Shuffle augments instance bags during training.

News

2026-11-05Robustness Lab and Meet the Heads are available on the project page.
ACM MM 2026 Our paper has been accepted to ACM Multimedia 2026!
2026-11-04Code, the GenImage checkpoint integration, and the interactive demo are available.

Abstract

Recent advances in generative models have made synthetic images increasingly photorealistic, posing growing challenges to reliable and generalizable detection of AI-generated images. Existing methods typically rely on holistic image-level classification, which may overlook local artifacts, especially in high-resolution images. To address this limitation, we formulate this task as a falsification-driven multiple instance learning problem, where a single fake patch can falsify the whole image, while a real image requires all patches to be authentic. Based on this formulation, we propose Learning to Falsify (LTF), a hierarchical multiple instance learning framework that models an image as multiple bags of instances across different spatial levels and aggregates their predictions under the falsification-inspired principle. We further introduce Cross-Shuffle during training to encourage the model to learn local artifact evidence rather than image-level shortcuts. Furthermore, we propose the Falsification-driven Test-time Aggregation Strategy (FATAS), a plug-and-play module that extends the same falsification-inspired principle to existing detectors by aggregating patch-level predictions. Extensive experiments demonstrate that LTF achieves strong generalization and robustness across multiple benchmarks, while FATAS substantially improves representative detectors in high-resolution scenarios without retraining.

Method

LTF formulates image detection as a multiple instance learning problem: local evidence of synthesis can falsify an image, while an authentic image requires its instances to be authentic.

Hierarchical multiple instance learning. A frozen CLIP ViT-L/14 backbone encodes local image tiles. Patch-level and tile-level bags aggregate evidence at different spatial scales.
Cross-Shuffle. Real-image tokens replace a subset of instances during training. This changes bag context while retaining original instances and preserving the training label.
FATAS. For an independent tile scorer, the Falsification-driven Test-time Aggregation Strategy averages the k highest generated probabilities. It requires no retraining.

The released GenImage profile uses CLIP block 24. See the checkpoint and reproduction notes for implementation details.

Experimental Results

Selected results reported in the paper. Accuracy is shown in percent.

Selected benchmark accuracy reported in the LTF paper
Benchmark	Metric	LTF	Reference
GenImage	Mean accuracy	98.4	Table 1
DRCT	Mean accuracy	98.62	Table 2
Chameleon	Accuracy	72.94	Figure 3
HiRes-50K	Average over 8 bins	81.43	Table 3

These are paper-reported results. The available weights and online demo use the GenImage checkpoint; they do not establish reproduction of every benchmark setting. Release notes.

HiRes-50K: accuracy by resolution bin

Unweighted mean across the eight bins: 81.43%. Values from Table 3.

Paper Figure 3: comparison on Chameleon. — **Cross-generator detection.** Results on Chameleon (Figure 3).

Paper Figure 4: FATAS accuracy across values of k. — **FATAS aggregation.** Effect of the choice of k (Figure 4).

Paper Figure 5: robustness under JPEG compression and Gaussian blur. — **Robustness.** Detection under common image transformations (Figure 5).

See the paper for full experiments · Figure attribution

Interactive Illustrations

Follow RGB instances through Cross-Shuffle, then explore top-k aggregation with measured tile scores.

Loading RGB examples…

Replace instances with real-image donors

RGB crops represent feature-token instances. The training operation acts on embeddings; the mosaics below show their provenance.

Replacement ratio50%

8 of 16 instances per bag; at least 1 original remains.

Sample donors with replacement

Real-image donor pool

The real target excludes its own image. The generated target can draw from all three real images.

Real target bag Real

Real image A

16 original · 0 inserted

Generated target bag Generated

Stable Diffusion 1.4

16 original · 0 inserted

Inserted real instanceRetained original instanceBorders indicate provenance, not artifact locations.

Original bags. Click any crop to inspect its source.

Explore LTF · 01

Robustness Lab

Compress, blur, or downsample an image. See how the pixels and the detector’s score change together.

Loading measured examples…

RGB exampleModel branch

Current transformed RGB example — **Current image**

JPEG qualityOriginal

Generated score—

Original: —
Change: —
Local tiles + global: —

Score across transformation levelsFusion

Click a point to select its image. The dashed line is the 0.5 decision threshold.

How this experiment works

Each setting is a separate, measured whole-image LTF prediction with the released GenImage checkpoint and CLIP block 24. Local 224 × 224 blocks do not overlap; only final right/bottom boundary blocks may overlap. The global view is included. Standard LTF aggregation is used.

One transformation is applied to the original at a time. JPEG uses the displayed quality; blur uses the displayed Pillow Gaussian radius; resize uses LANCZOS and passes the smaller image directly to the model. The original is the reference for every curve. These individual examples do not replace a benchmark robustness evaluation.

· All measured data · Sources and protocol

Explore LTF · 02

Meet the Heads

Eight L1 aggregation heads, one image. Explore how each head weights the same local tiles.

Loading measured attention…

RGB exampleInspect tile

Overlay opacity45%

Linked crop inspector

Select a tile

Relative to uniform attention

0.25×4×

1× is neutral · same scale for all heads

Hover, focus, or click a tile in any map. The selected crop stays linked across all eight heads; use the tile selector to reach overlapping crops.

Reading these maps

These are the eight learned L1 aggregation heads across local tiles. Each head’s weights sum to one. A weight of 1/N is uniform attention over N local tiles. The separate global view belongs to L2 and is excluded from these maps.

All maps share a fixed log₂ color scale from 0.25× to 4× uniform attention. Overlapping regions show the mean weight of covering tiles. Similar weights have similar colors. Attention shows relative model weighting, not a pixel-level forgery probability or a causal explanation.

· All measured data · Inspect your own image in the demo ↗

Interactive Demo

Upload an original image to inspect its prediction, local tiles, L1/L2 attention, individual aggregation heads, and patch weights. Attention represents relative model weighting, not pixel-level forgery probabilities.

Try the demo

For local use, see the demo source and instructions.

TODO

☐Release checkpoints and evaluation profiles for additional benchmarks.
☐Complete full training and benchmark reproduction checks.
☐Add data preparation scripts and more demo examples.

BibTeX

@inproceedings{chen2026ltf,
  title={Learning to Falsify: Hierarchical Multiple Instance Learning for High Resolution AI-Generated Image Detection},
  author={Chen, Haoran and Gao, Daiqi and Li, Xinyuan and Wang, Yongwei and Zhu, Lei and Cai, Rui and Dong, Jianfeng},
  booktitle={Proceedings of the 34th ACM International Conference on Multimedia},
  year={2026},
  doi={10.1145/3767308.3836365},
  eprint={2611.04529},
  archivePrefix={arXiv}
}