Designing systems that attribute a generated image to its source model has lately become a major area of AI research. Most existing datasets and systems related to image source attribution are targeted towards the detection of architectural signatures of the source models. Distinct from this, focusing on stable diffusion, we introduce a novel task, IntrAA-SD, Intra-Architecture source Attribution for Stable Diffusion, where the source models to be differentiated share the same architectural backbone and differ only due to fine-tuning. The proposed task aligns with the emergence of fine-tuned generators having the same architecture, like Juggernaut XL and RealVis XL, which are fine-tuned versions of Stable Diffusion XL (SDXL). We provide a first-of-its-kind 140K-image benchmark dataset containing images generated from different fine-tunes of SDXL, along with in-domain and out-of-distribution prompts. We also provide a baseline framework for IntrAA-SD, which leverages a style-conditioned Mixture-of-Experts network. It uses a frozen CLIP backbone to capture the image style and gate the MoE, which processes fine-tuned trace features from a high-resolution LoRA-adapted DINOv2 network. The fine-tuned models used, the prompting strategies considered, and the formation details employed in the novel dataset are elaborated. Our baseline intra-architecture source attribution method is evaluated in terms of attribution accuracy on our dataset, and ablation studies are also performed. Code and dataset is available at https://github.com/mallikagarg/IntrAA-SD.
Figures & tables
Figure 1: Sample images from our IntrAA-SD benchmark dataset generated using identical prompts by SDXL fine-tunes represented by M1: SDXL Base, M2: Juggernaut, M3: RealVis, M4: DreamShaper, M5: Animagine, M6: Counterfeit, M7: ZavyChroma, M8: Copax, M9: Proteus, M10: AAM Anime. P1: “ volcanic landscape with lava flow ”, P2: “ electric car charging at a station ”. Despite sharing the same backbone, each fine-tune produces outputs with subtle visual distinctness with some pairs (like M2 & M3) more similar than others. Note: We use M1-M10 to denote the corresponding fine-tuned models throughout the paper.
“ bowl of ramen with egg, nori, chopsticks, steam rising ”
Table 2: In-Domain (ID) prompt categories in our ISDB dataset with example prompts.
#
Category
Example Prompt
Adversarial Strategies (7 categories)
1
Cross-Domain
“ studio portrait, soft lighting, 85mm lens ” → fed to Animagine XL; “ flat anime character, cel shading ” → fed to RealVis XL
2
Void Prompts
“ pure white background, nothing else”; “complete darkness, empty frame ”
3
Nested Representation
“ a photograph of a charcoal sketch pinned to a cork board ”; “ a watercolor painting lying on a wooden desk ”
4
Text-Heavy
“ a handwritten letter on aged parchment, cursive text filling the page ”
5
Hands & Anatomy
“ a pianist playing a grand piano, fingers on keys, concert hall ”
Table 3: Out-of-Distribution (OOD) prompt categories in our ISDB dataset with example prompts.
Figure 2: Our IntrAA-SD Baseline Framework. A frozen CLIP backbone captures the image style (Sec. 4.1 ) to compute gating weights across N experts. A LoRA-adapted DINOv2 backbone extracts the fine-tunes’ related patterns fed into experts (Sec. 4.2 ). The expert outputs are combined via style-conditioned Mixture-of-Experts (Sec. 4.3 ) and concatenated with the style context for the attribution (Sec. 4.4 ).
Source Model
ID
OOD
SDXL Base (M1)
99.42
91.50
Juggernaut XL (M2)
99.79
70.50
RealVis XL (M3)
98.67
82.50
DreamShaper XL (M4)
99.83
99.00
Animagine XL (M5)
99.92
98.00
CounterfeitXL (M6)
98.29
98.00
Table 4: Intra-architecture source attribution accuracies (%) of our IntrAA-SD baseline framework on our ISDB dataset, for the test images generated with In-Domain (ID) and Out-of-Distribution (OOD) prompts.
Components
Variants
Acc %
Gating
A1: Equal weighting (no gating)
98.40
strategy
A2: Structure-based gating (DINOv2 → gate)
98.42
Style concat
A3: No style concatenation (after MoE) before classification
98.21
Style or
A4: Image Style-only (Frozen CLIP + linear head)
87.59
trace feature
A5: Source trace-only (LoRA DINOv2 + linear head)
98.52
All
IntrAA-SD framework
99.03
Table 5: IntrAA-SD Ablation: significance of components. All variants use LoRA r=16, 518px, 4 experts (unless removed).
Table 7: Ablation on the loss function ( Ltotal ) parameter λbal
Figure 3: t-SNE visualization of model trace feature embeddings.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Method
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
Overall
ID
Spec [ 15 ]
18.13
12.96
11.86
46.45
40.83
36.51
9.31
20.49
32.01
21.32
24.99
UnivFD [ 30 ]
24.46
27.79
30.32
70.52
76.47
29.73
33.85
35.62
56.71
61.41
44.69
DE-FAKE [ 41 ]
62.96
65.37
61.32
89.58
93.91
76.24
62.28
67.97
87.87
86.22
75.37
NPR [ 46 ]
88.27
85.53
86.78
97.97
96.56
94.03
86.51
89.81
94.42
95.08
91.50
CNNDetect [ 48 ]
87.26
83.93
88.96
98.32
98.40
95.38
87.01
90.27
96.34
96.28
92.21
Ours ∗
99.42
99.79
98.67
99.83
99.92
98.29
99.67
97.88
98.58
98.21
99.03
Appendix
Table 1: Intra-architecture source attribution accuracy (%) on our ISDB test set. Models - M1: SDXL Base, M2: Juggernaut, M3: RealVis, M4: DreamShaper, M5: Animagine, M6: Counterfeit, M7: ZavyChroma, M8: Copax, M9: Proteus, M10: AAM Anime. ∗ Ours is the only algorithm (& the first) in the list which is designed for intra-architecture source attribution. All others listed are existing architecture-level source attribution methods.
Figure 1: t-SNE projection of frozen CLIP ViT-L/14 features of input images generated using a suitable shared prompt across all 10 SDXL fine-tunes.
Experts
Photorealistic
Stylized
Anime
M4
M2
M1
M3
avg.
M8
M7
M9
avg.
M6
M10
M5
avg.
Expert 1
35.5
36.7
37.1
37.6
36.7
37.2
36.6
34.9
36.2
33.5
28.1
15.8
25.8
Expert 2
24.9
24.8
23.0
25.2
24.5
26.3
26.8
27.2
26.8
27.2
29.2
27.4
27.9
Expert 3
23.7
21.6
23.6
21.2
22.5
18.8
19.5
19.6
19.3
17.8
17.4
13.3
16.2
Expert 4
15.9
16.9
16.2
16.1
16.3
17.7
17.1
18.3
17.7
21.6
25.3
43.6
30.2
Appendix
Table 2: Mean Mixture of Expert (MoE) gating weights (%) per model, grouped & averaged by style type.
Training Set
Test Set Accuracy%
ISDB-ID
ImageNet
COCO
ImageNet
80.59
98.35
96.47
COCO
76.16
90.31
98.91
ISDB-ID (Ours)
99.03
84.58
85.95
Appendix
Table 3: Cross-Dataset Performance of Our IntrAA-SD baseline
True ↓ / Pred →
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M1 (SDXL Base)
923
3
2
0
1
5
11
46
5
4
M2 (Juggernaut)
22
708
149
8
0
14
49
28
16
6
M3 (RealVis)
16
87
816
4
0
12
35
18
9
3
M4 (DreamShaper)
0
5
0
985
0
0
4
0
1
5
M5 (Animagine)
4
0
0
0
984
1
3
0
0
8
M6 (Counterfeit)
1
3
0
1
2
978
5
0
3
7
Appendix
Table 4: Confusion Matrix related to performance of our IntrAA-Diff on ISDB-ODD set. Photorealistic: M1-4 , Stylized: M7-9 , Anime: M5-6, M10 .
Figure 2: Significant changes (Red: increase, Blue: decrease) at patch-level feature norm after LoRA adaptation of DINOv2 for images from four different fine-tuned models.
Attributing a generated image to its source diffusion model is a fundamental challenge in provenance verification and intellectual property protection. This problem is particularly difficult because diffusion models trained on different datasets can converge to similar score functions and thus similar output distributions, making the generated images themselves unreliable as attribution evidence. Existing non-invasive methods either fail on architecturally similar variants or rely on signals that vanish when models share the same autoencoder. We propose Spectral Denoising Signatures (SDS), a non-invasive attribution method that identifies the source model by fingerprinting each candidate model's denoising behavior. Our key insight is that a model's denoising score function exhibits a distinctive spectral geometry, reflected in how it redistributes energy across spatial frequency bands during denoising. By probing this behavior with frequency-controlled perturbations, SDS extracts a stable signature that is intrinsic to the model, requiring only standard forward passes with no inversion, optimization, or generation-time enrollment. Our results demonstrate that SDS achieves approximately 99.9% accuracy across eight diverse diffusion models and 96.2% under cross-domain prompt shift, outperforming non-invasive baselines across variations in training data, architecture, and training procedure, establishing spectral geometry as a principled and practical basis for diffusion model attribution. Code is available at: https://github.com/Pragati-Meshram/SGS
Pragati Shuddhodhan Meshram, Varun Chandrasekaran
Department of Electrical and Computer Engineering University of Illinois Urbana-Champaign
The rapid evolution of image generation has produced numerous within-family variants, making source-model attribution of suspect images increasingly important for digital forensics. Existing proactive methods rely on watermark embedding or model modification, which may degrade visual quality and limit deployment flexibility. Passive methods often rely on large-scale supervised training or a single reconstruction signal, limiting their ability to handle unknown sources and distinguish highly similar within-family variants. We observe that attribution signals in latent generative models are naturally stratified across architectural levels: VAE-level cues reflect family-shared information, whereas backbone-level cues capture variant-specific behaviors. Motivated by this insight, we propose Dual-stage Native Attribution (DNA), a coarse-to-fine framework that follows this hierarchy without additional neural-network training. The coarse-grained stage uses Autoencoder Double-Reconstruction (AEDR) for efficient open-set family-level screening. The fine-grained stage performs closed-set model-level attribution with Native Prediction Consistency (NPC), which compares native prediction errors of within-family variants across multiple noise levels under semantic conditioning and attributes the source via normalized calibrated scores. To enable systematic evaluation, we construct DNA-30K, a benchmark for within-family variant attribution under open-set family-level evaluation. It comprises 30,000 images generated by 24 candidate models across six families spanning both denoising diffusion and flow matching, plus non-candidate generated and natural images as unknown sources. Experiments show that DNA achieves 89.11% end-to-end attribution accuracy on a task where random guessing accuracy is below 1% and outperforms the strongest baseline by 33.81% even when AEDR is used as the coarse-grained stage.
Chao Wang, Kejiang Chen, Zijin Yang +4
University of Science and Technology of China, Hefei, China · Hefei University of Technology, Hefei, China
As AI generative models evolve at unprecedented speed, image attribution has become a moving target. New diffusion, adversarial and autoregressive generators appear almost monthly, making existing watermark, classifier and inversion methods obsolete upon release. The core problem lies not in model recognition, but in the inability to adapt attribution itself. We introduce IncreFA, a framework that redefines attribution as a structured incremental learning problem, allowing the system to learn continuously as new generative models emerge. IncreFA departs from conventional incremental learning by exploiting the hierarchical relationships among generative architectures and coupling them with continual adaptation. It integrates two mutually reinforcing mechanisms: (1) Hierarchical Constraints, which encode architectural hierarchies through learnable orthogonal priors to disentangle family-level invariants from model-specific idiosyncrasies; and (2) a Latent Memory Bank, which replays compact latent exemplars and mixes them to generate pseudo-unseen samples, stabilising representation drift and enhancing open-set awareness. On the newly constructed Incremental Attribution Benchmark (IABench) covering 28 generative models released between 2022 and 2025, IncreFA achieves state-of-the-art attribution accuracy and 98.93% unseen detection under a temporally ordered open-set protocol. Code will be available at https://github.com/Ant0ny44/IncreFA.
Haotian Qin, Dongliang Chang, Yueying Gao +3
School of Artificial Intelligence, Beijing University of Posts and Telecommunications, China · Tsinghua University, China