Guess the Unified Model: How Much Can We Recover from Generated Images?

Addison J. Wu; Jasin Cekinmez; Ryo Mitsuhashi; Yida Yin

arxiv: 2605.25254 · v1 · pith:Y4VQ3KXE · submitted 2026-05-24 · cs.CV · cs.AI

Guess the Unified Model: How Much Can We Recover from Generated Images?

Jasin Cekinmez , Ryo Mitsuhashi , Addison J. Wu , Yida Yin This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-30 11:51 UTCgrok-4.3pith:Y4VQ3KXE record.json open to challenge →

classification cs.CV cs.AI

keywords model attributiongenerated imagesimage forensicsunified modelsprovenanceAI-generated contentdiffusion modelssource identification

0 comments

The pith

Unified AI image models produce outputs with consistent visual signatures that allow near-perfect source attribution from roughly 20,000 images per model.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether images from different unified generative models carry distinguishable traits tied to their origin. It measures attribution performance across image corruptions, domain shifts, and prompt languages using outputs from seven models. Results show that a classifier reaches near-perfect accuracy under these conditions, with semantic content helping but not dominating the signal and prompt language adding little. This establishes that model-specific visual characteristics persist in the outputs and could support tracing of generated content.

Core claim

Images generated by seven unified models exhibit separable visual characteristics sufficient for a classifier to recover the source model with near-perfect accuracy from around 20K images per model. Corruptions and structural perturbations affect performance only modestly, cross-domain tests indicate that semantic content contributes to separability without being the main driver, and prompt-language attribution stays near chance levels for most models.

What carries the argument

An image classifier trained to predict the generating model, evaluated on held-out images under controlled variations in corruption, domain, and prompt language.

If this is right

Model attribution becomes practical for auditing generative image pipelines at scale.
Consistent model-specific visual traits survive common image corruptions and domain changes.
Semantic content aids separability but is not required for high performance.
Prompt language leaves minimal detectable visual signatures in most models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the signatures remain stable under real-world distribution shifts, attribution tools could be deployed on public image collections without retraining.
The approach could be extended to detect fine-tuned or merged models by checking whether their outputs cluster with base-model signatures.
Large-scale provenance tracking might become feasible once a small reference set of labeled images exists for each model family.

Load-bearing premise

The observed separability arises from intrinsic model-specific visual characteristics rather than from shared generation pipelines, prompt distributions, or other uncontrolled factors.

What would settle it

Retrain the attribution classifier on a new set of images where all models receive identical prompts, identical generation parameters, and identical post-processing, then measure whether accuracy falls to chance level.

Figures

Figures reproduced from arXiv: 2605.25254 by Addison J. Wu, Jasin Cekinmez, Ryo Mitsuhashi, Yida Yin.

**Figure 2.** Figure 2: Attribution accuracy scales rapidly, surpassing [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Figure 4: Images generated by Janus, MMaDA, and Show-o2 with prompts in English, Spanish, Japanese, Turkish, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

With unified model-generated images now widespread online, attributing their model of origin offers a path toward transparency and deeper insight into the characteristic behaviors of individual models. Prior work has explored provenance in LLM-generated text, diffusion model images, and datasets, but the separability of unified model-generated images remains an underexplored area. We address this gap by examining separability across corruption, domains, and prompt languages using images generated by seven unified models. We show that model attribution is highly feasible as our model achieves near-perfect accuracy with around 20K images per model. Corruptions and structural perturbations have only a modest effect on attribution performance, and cross-domain generalization reveals that semantic content contributes to separability but is not the dominant signal. Finally, we observe that for most models, prompt language attribution is around chance levels, suggesting minimal language-specific visual signatures. These findings highlight consistent model-specific visual characteristics in unified models outputs and open new directions for tracing and auditing generative image pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows unified model images are highly separable by classifier with near-perfect accuracy that holds up under corruptions and cross-domain tests.

read the letter

The main takeaway is that images from seven different unified generative models carry consistent visual signatures a classifier can exploit for attribution at near-perfect levels using roughly 20k images per model.

What is new is the focus on unified models rather than just diffusion or text generators, plus the controls for corruptions, domain shifts, and prompt language. The cross-domain results indicate semantic content contributes but does not dominate, while language effects stay near chance for most models. These tests directly check whether the signal is model-intrinsic instead of prompt or pipeline artifacts, which strengthens the central claim.

The work does well at showing modest degradation under perturbations and structural changes, giving evidence that the attribution signal is stable. That is useful empirical grounding for the idea of model-specific characteristics in outputs.

The soft spots are mostly in the abstract's lack of detail on the classifier architecture, exact baselines, data splits, and any statistical tests. Without those it is hard to judge how much the result depends on standard techniques versus something more involved. The full paper presumably fills this in, but the summary leaves the methods somewhat opaque.

This is for people working on generative forensics or auditing pipelines. A reader who needs concrete numbers on separability in this setting will get value from the controlled experiments.

It deserves peer review. The described controls line up with the main potential objections, and the topic fills a stated gap even if the accuracy numbers need the methods section to land fully.

Referee Report

0 major / 2 minor

Summary. The paper examines separability of images generated by seven unified models. It reports near-perfect attribution accuracy using a classifier trained on around 20K images per model. Experiments test robustness under corruptions and structural perturbations, cross-domain generalization (showing semantic content contributes but is not dominant), and prompt-language attribution (near chance for most models), concluding that consistent model-specific visual signatures exist in the outputs.

Significance. If the empirical results hold under the reported controls, the work provides concrete evidence that model attribution for unified generative models is feasible and driven by intrinsic visual characteristics rather than prompt distributions or shared pipelines. The cross-domain and corruption tests directly address potential confounds, strengthening the central claim and opening directions for auditing generative pipelines.

minor comments (2)

[Methods] Methods section: the classifier architecture, training procedure, exact data splits, and any baseline comparisons should be described with sufficient detail (including hyperparameters and statistical tests) to support reproducibility of the near-perfect accuracy figures.
[Results] Results: tables or figures reporting the accuracy numbers under each corruption/domain/language condition would benefit from explicit error bars or p-values to quantify the 'modest effect' claims.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, significance assessment, and recommendation of minor revision. No major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity; empirical classification results are self-contained

full rationale

The paper reports empirical model attribution accuracies from classifiers trained on generated images across seven unified models, with experiments on corruptions, cross-domain generalization, and prompt languages. No derivation chain, equations, or first-principles predictions exist that could reduce to inputs by construction. Central claims rest on observed classification performance under controlled conditions rather than any self-definitional, fitted-input, or self-citation load-bearing steps. This is the expected outcome for an empirical ML attribution study with no mathematical reduction to fitted quantities.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract; the central claim rests on the domain assumption that different unified models produce images with distinguishable visual features. No free parameters or invented entities are described.

axioms (1)

domain assumption Different unified models produce images with distinguishable visual features
This assumption directly enables the attribution task and is invoked by the claim of high separability.

pith-pipeline@v0.9.1-grok · 5708 in / 1262 out tokens · 37660 ms · 2026-06-30T11:51:31.216668+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages · 2 internal anchors

[1]

Emu3.5: Native Multimodal Models are World Learners

On the detection of synthetic images generated by diffusion models. InIEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jinsheng Wang, Wenxuan Wang, Yueze Wang, Chengyuan Wang, Fan Zhang, Yingli Zhao, Ting Pan, Xian...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[2]

Emerging properties in unified multimodal pretraining.arXiv:2505.14683v3. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An image is worth 16x16 words: Transformers for image recognition at...

work page internal anchor Pith review Pith/arXiv arXiv 2021
[3]

Below are k reference example(s) per model so you can see each model’s visual style,

Depth anything v2. InAdvances in Neural Information Processing Systems, volume 37, pages 21875–21911. Curran Associates, Inc. Ling Yang, Ye Tian, Bowen Li, Xinchen Zhang, Ke Shen, Yunhai Tong, and Mengdi Wang. 2025. Mmada: Multimodal large diffusion language mod- els. InAdvances in Neural Information Processing Systems, volume 38, pages 138867–138907. Cur...

2025

[1] [1]

Emu3.5: Native Multimodal Models are World Learners

On the detection of synthetic images generated by diffusion models. InIEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jinsheng Wang, Wenxuan Wang, Yueze Wang, Chengyuan Wang, Fan Zhang, Yingli Zhao, Ting Pan, Xian...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[2] [2]

Emerging properties in unified multimodal pretraining.arXiv:2505.14683v3. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An image is worth 16x16 words: Transformers for image recognition at...

work page internal anchor Pith review Pith/arXiv arXiv 2021

[3] [3]

Below are k reference example(s) per model so you can see each model’s visual style,

Depth anything v2. InAdvances in Neural Information Processing Systems, volume 37, pages 21875–21911. Curran Associates, Inc. Ling Yang, Ye Tian, Bowen Li, Xinchen Zhang, Ke Shen, Yunhai Tong, and Mengdi Wang. 2025. Mmada: Multimodal large diffusion language mod- els. InAdvances in Neural Information Processing Systems, volume 38, pages 138867–138907. Cur...

2025