Synthetic-media work shifts from spotting fakes to naming the generator

A benchmark built for attribution

In a preprint, researchers introduce DeepfakeGenome, a benchmark with 100 facial forgery algorithms and 2M images. The shift it proposes is in the question being asked. Instead of binary classification, real or fake, the benchmark is built for deepfake attribution: identifying which forgery method produced an image. Its retrieval-based evaluation protocol is designed to mimic real-world blacklist defense, where a suspect image is checked against a registry of known forgery sources. Retrieval-based attribution metrics are, according to the paper, more aligned with real-world active defense and blacklist registration mechanisms than open-set evaluation metrics. m:…4279

This suggests the field's bottleneck is no longer whether a detector can pass a test set but whether its verdict is useful to someone defending a platform. A binary label tells a moderator to remove an image; an attribution tells them which generator to block, which campaign the image belongs to, and whether it matches something already seen. The benchmark's scale, 2M images across 100 methods, is what makes that kind of matching testable at all, since a small dataset cannot cover the variety of forgery tools now in circulation. m:…4279

Detectors are reading the wrong signal

A second preprint argues that many current detectors would fail that harder test because they are not really detecting manipulation. The paper introduces the Inpainting Exchange diagnostic, which restores the original background pixels of an edited image, and finds that it collapses state-of-the-art detector performance to chance levels. The reason, according to the authors, is that VAE-based inpainting introduces a pervasive global spectral shift in unedited regions, and 97% of background spectral shift is caused by the autoencoder round-trip. In other words, detectors are largely picking up the fingerprint of the autoencoder that processed the whole image, not the content that was actually changed. m:…4257

The diagnostic collapses state-of-the-art detector performance to chance levels, which means a detector that looked strong in the lab can be near useless on an edit where the background was restored. The reason is the autoencoder: VAE-based inpainting introduces a pervasive global spectral shift in unedited regions, and 97% of background spectral shift is caused by the autoencoder round-trip. In other words, detectors are largely picking up the fingerprint of the autoencoder that processed the whole image, not the content that was actually changed. Detecting such inpainting edits remains an open problem, according to the paper's own unresolved list. m:…4257

Identity, separated from the picture

A third preprint tackles the supply side of synthetic faces. IDPrior decouples identity generation from image synthesis by learning a diffusion prior in the embedding space of a pretrained face recognition model. The generated identity embeddings then condition a pretrained diffusion model to synthesize consistent facial images, which means new identities can be created without collecting large-scale identity-labeled face datasets, with their cost and privacy risks. The framework generates diverse and well-separated identities while maintaining a closer correspondence to the distribution of authentic identities. m:…4273

Face recognition models trained on IDPrior-generated identities achieve competitive or improved performance compared with existing synthetic-data-based methods on face verification benchmarks. The implication cuts both ways. The same technique that lets researchers train face recognition without harvesting real faces also makes it cheaper to produce consistent synthetic identities at scale, which is exactly the capability an attribution system would need to trace. Whether the identities stay stable in large-scale training remains to be fully explored, the paper notes. m:…4273

Teaching models to see what is actually there

A fourth preprint in the same batch tackles the other side of the same problem: not whether an image is fake, but whether a vision-language model can be trusted to describe what it actually sees. DreamRealm is a post-training framework that mines a target model's latent hallucinations and uses image editing to create corrected image-response pairs as supervision signals. Those pairs are then used to train the model with bidirectional contrastive losses, a step the authors call RealmMatch, which they say effectively suppresses perceptual mismatches. m:…4261

In a preprint, the paper reports that DreamRealm improves HallusionBench accuracy by 7.26 percentage points and MMVP pair accuracy by 11.34 percentage points compared with the base Qwen3.5-9B model. This suggests the same discipline attribution needs applies inside the model as well: a verdict grounded in the content of the image, rather than in the fingerprint of whatever processed it. Whether the gains hold beyond the two benchmarks reported is not shown in the abstract, and no public code or dataset URL was provided there. m:…4261