CSF: Black-box Fingerprinting via Compositional Semantics for Text-to-Image Models

Junhoo Lee; Mijin Koo; Nojun Kwak

cs.CRMar 20, 2026

CSF: Black-box Fingerprinting via Compositional Semantics for Text-to-Image Models

Authors: Junhoo Lee, Mijin Koo, Nojun Kwak

Organizations: Seoul National University

Abstract

Text-to-image models are commercially valuable assets often distributed under restrictive licenses, but such licenses are enforceable only when violations can be detected. Existing methods require pre-deployment watermarking or internal model access, which are unavailable in commercial API deployments. We present Compositional Semantic Fingerprinting (CSF), the first black-box method for attributing fine-tuned text-to-image models to protected lineages using only query access. CSF treats models as semantic category generators and probes them with compositional underspecified prompts that remain rare under fine-tuning. This gives IP owners an asymmetric advantage: new prompt compositions can be generated after deployment, while attackers must anticipate and suppress a much broader space of fingerprints. Across 6 model families (FLUX, Kandinsky, SD1.5/2.1/3.0/XL) and 13 fine-tuned variants, our Bayesian attribution framework enables controlled-risk lineage decisions, with all variants satisfying the dominance criterion.

Figures & tables

Figure 1: Comparison of model identification scenarios. (a) Watermarking requires pre-deployment access to inject a trigger into the base model. (b) Traditional Fingerprinting relies on white-box or gray-box ‘Internal Access’ (e.g., weights or activations), which is not available in commercial API. (c) Our approach (CSF) is designed for the most restrictive ‘Query Only’ black-box setting, where the defender only has access to the final T2I generation API, reflecting real-world infringement scenarios.

Figure 2: The “Name That Dataset” game [ 52 ] in Diffusion Fingerprinting. Image (a) is from a fine-tuned model (SD1.5-DreamShaper). One of the images (b, c, d) is its base model. This figure illustrates how difficult it is to identify the base model using a naive prompt (right column , randomly sampled from LAION-2B [ 46 ] ), compared to our CSF prompt (left column) . All images are uncurated results generated with different random seeds. Can you specify which model (b, c, or d) is the base model for (a)? Answer: (a) SD1.5-DreamShaper (fine-tuned), (b) SD2.1-DPO (fine-tuned), (c) SD1.5 Base, and (d) SDXL Base. The correct base model for (a) is (c).

Panel of Figure 3 — Figure 3: Challenges in naive fingerprinting approaches. (a) Visual space : t-SNE visualization of CLIP embeddings shows no family clustering. (b) Text space : Even when images are converted to captions via I2T models, style information leaks into the text, causing models from different families (e.g., SD1.5 DPO and SD2.1) to cluster together due to similar style descriptors.

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.932*	0.023	0.023	0.023	0.023	0.068
Flux-Turbo-Alpha	0.977*	0.023	0.023	0.023	0.023	0.023
Kandinsky Family
Kandinsky-Naruto	0.023	0.977*	0.023	0.023	0.023	0.023

Table 1: Posterior Mean of the Derived Models. In this table, xx indicates significance a (Confidence Interval (CI) low

>

0.167), xx indicates Not significant (CI includes 0.167), and xx indicates Sig. below chance (CI high

<

0.167). * indicates the model meets Dominance test (CI low

>

0.5)

Figure 4: User study results for the “Name That Dataset” game ( Fig. 2 ). Participants attempted to identify the correct base model using naive prompts (LAION-2B [ 46 ] ) vs. our CSF prompts.

Figure 5: Generated category distributions vary substantially with scene context. Compared with the dimmed-studio setting, the dish and wooden-floor settings produce broader and systematically shifted category mixtures. Each donut summarizes 40 samples, and the same color denotes the same category across scenes.

Fine-tuned Model	Wasserstein (Ours)	JSD (Baseline)	Confidence Gap
Flux-LoRA	93.2%	77.3%	+15.9%
Kandinsky-Naruto	97.7%	43.2%	+54.5%
SD3-Reality-Mix	70.5%	56.8%	+13.7%
SDXL-DPO	97.7%	70.5%	+27.2%

Table 2: Attribution confidence: Wasserstein vs. JSD

	Base Models
Model	Flux	Kandinsky	SD1.5	SD2.1	SD3-Med	SDXL
LoRA	0.714	0.143	0.143	0.143	0.143	0.286
Turbo-Alpha	0.857	0.143	0.143	0.143	0.143	0.143
Naruto	0.143	0.857	0.143	0.143	0.143	0.143
Pokemon-LoRA	0.143	0.857	0.143	0.143	0.143	0.143
1.2-Base	0.143	0.143	0.857	0.143	0.143	0.143

Table 3: Attribution results under adversarial concept removal. Posterior mean attribution scores computed from 9 animal-specific probes after removing animal-related concepts using UCE.

Figure 6: Failure cases of CSF. Top: prompts for baked goods in front of a brick wall. Bottom: prompts for tropical flowers on a pot. In both cases, models failed to generate identifiable objects within the target categories.

Model Name	HuggingFace Path
Flux Family
Flux-Base	black-forest-labs/FLUX.1-dev
Flux-LoRA	XLabs-AI/flux-RealismLora
Flux-Turbo-Alpha	alimama-creative/FLUX.1-Turbo-Alpha
Kandinsky Family
Kandinsky-Base	kandinsky-community/kandinsky-2-2-prior

Table A1: Model details for fingerprinting evaluation. All models are available at https://huggingface.co/[model-path] .

	AVERAGE ACROSS ALL PROMPTS
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.34	0.25	0.71	0.71	0.71	0.85	0.84	0.84	0.77	0.81	0.83	0.83	0.63	0.65	0.66	0.78	0.79	0.73
F-LoRA	0.33	0.00	0.41	0.71	0.71	0.75	0.86	0.83	0.82	0.81	0.80	0.81	0.81	0.68	0.73	0.73	0.74	0.74	0.70
F-Turbo	0.24	0.41	0.00	0.73	0.74	0.73	0.78	0.78	0.79	0.77	0.76	0.77	0.77	0.67	0.69	0.69	0.78	0.79	0.73
K-Base	0.67	0.70	0.73	0.00	0.14	0.34	0.79	0.83	0.83	0.78	0.81	0.79	0.79	0.78	0.78	0.78	0.65	0.65	0.63

Table A2 : Average Normalized Wasserstein Distance Matrix across all prompts. Each value is the mean of column-normalized distances across all 42 prompts. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

Panel of Figure 12 — Figure B1: Qualitative results (Part 1): Flux, Kandinsky, and SD1.5 families.

Panel of Figure 13 — Figure B2: Qualitative results (Part 2): SD2.1, SD3, and SDXL families.

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.541	0.004	0.004	0.004	0.004	0.004
Flux-Turbo-Alpha	0.541	0.004	0.004	0.004	0.004	0.004
Kandinsky Family
Kandinsky-Naruto	0.004	0.541	0.004	0.004	0.004	0.004

Table A3 : 95% Confidence Interval Lower Bound of the Derived Models with base method

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.857	0.143	0.143	0.143	0.143	0.143
Flux-Turbo-Alpha	0.857	0.143	0.143	0.143	0.143	0.143
Kandinsky Family
Kandinsky-Naruto	0.143	0.857	0.143	0.143	0.143	0.143

Table A4 : 95% Confidence Posterior mean of the Derived Models with base method

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.842	0.001	0.001	0.001	0.001	0.015
Flux-Turbo-Alpha	0.918	0.001	0.001	0.001	0.001	0.001
Kandinsky Family
Kandinsky-Naruto	0.001	0.918	0.001	0.001	0.001	0.001

Table A5 : 95% Confidence Interval Lower Bound of the Derived Models with our method

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.640	0.001	0.015	0.006	0.026	0.026
Flux-Turbo-Alpha	0.842	0.006	0.001	0.001	0.001	0.006
Kandinsky Family
Kandinsky-Naruto	0.015	0.291	0.118	0.053	0.015	0.068

Table A6 : 95% Confidence Interval Lower Bound of the Derived Models in jsd

	Base Models
	Flux-Base	Kandinsky-Base	SD1.5-Base	SD2.1-Base	SD3-Medium-Base	SDXL-Base
Flux Family
Flux-LoRA	0.773	0.023	0.068	0.045	0.091	0.091
Flux-Turbo-Alpha	0.932	0.045	0.023	0.023	0.023	0.045
Kandinsky Family
Kandinsky-Naruto	0.068	0.432	0.227	0.136	0.068	0.159

Table A7 : Posterior mean of JSD

Figure B5 : An instruction of human study

Figure B6 : Example of human study for comparison

Object Category	Adjectives	Locations
Baked good	savory, cheesy, sweet	dimmed studio, dark wood surface, brick wall
Animal	dangerous, wild, peaceful	grassland, forest, dimmed studio
Flower	vibrant, tropical	pot, dimmed studio, vase
Bird	peaceful, dangerous, flightless	grass, savana, dimmed studio
Fruit	sweet, frozen, savory	dish, wooden floor, dimmed studio

Table A8: Compositional structure of fingerprinting prompts. Each category systematically varies semantic attributes and visual contexts, yielding 42 total prompts (9+9+6+9+9). The dimmed studio setting appears across all categories to enable cross-category comparison.

	AVERAGE ACROSS ALL PROMPTS
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.34	0.25	0.71	0.71	0.71	0.85	0.84	0.84	0.77	0.81	0.83	0.83	0.63	0.65	0.66	0.78	0.79	0.73
F-LoRA	0.33	0.00	0.41	0.71	0.71	0.75	0.86	0.83	0.82	0.81	0.80	0.81	0.81	0.68	0.73	0.73	0.74	0.74	0.70
F-Turbo	0.24	0.41	0.00	0.73	0.74	0.73	0.78	0.78	0.79	0.77	0.76	0.77	0.77	0.67	0.69	0.69	0.78	0.79	0.73
K-Base	0.67	0.70	0.73	0.00	0.14	0.34	0.79	0.83	0.83	0.78	0.81	0.79	0.79	0.78	0.78	0.78	0.65	0.65	0.63

Table A9 : Average Normalized Wasserstein Distance Matrix across all prompts. Each value is the mean of column-normalized distances across all 42 prompts. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Cheesy Baked Good Against A Brick Wall
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.15	0.15	0.96	0.96	0.96	0.47	0.30	0.36	0.29	0.93	0.94	0.94	0.35	0.39	0.42	0.50	0.42	0.37
F-LoRA	0.14	0.00	0.19	0.86	0.87	0.83	0.46	0.29	0.35	0.32	0.75	0.79	0.79	0.26	0.32	0.34	0.40	0.36	0.29
F-Turbo	0.15	0.21	0.00	0.98	0.99	0.97	0.40	0.27	0.33	0.30	0.93	0.92	0.93	0.34	0.39	0.43	0.49	0.40	0.35
K-Base	1.00	1.00	1.00	0.00	0.05	0.25	1.00	1.00	1.00	1.00	0.87	0.55	0.59	1.00	1.00	1.00	1.00	1.00	1.00

Table A10 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Cheesy Baked Good On A Dark Wood Surface
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.14	0.14	1.00	1.00	1.00	0.68	0.26	0.31	0.22	0.74	0.78	0.77	0.27	0.54	0.55	1.00	1.00	1.00
F-LoRA	0.12	0.00	0.19	0.86	0.86	0.82	0.65	0.27	0.31	0.28	0.67	0.67	0.67	0.29	0.43	0.44	0.84	0.82	0.86
F-Turbo	0.12	0.19	0.00	0.85	0.85	0.86	0.58	0.27	0.31	0.27	0.67	0.68	0.68	0.32	0.51	0.52	0.83	0.82	0.85
K-Base	1.00	1.00	1.00	0.00	0.04	0.27	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.38	0.50	0.71

Table A11 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Cheesy Baked Good On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.43	0.45	0.74	0.69	0.61	0.80	0.68	0.70	0.56	0.53	0.64	0.62	0.50	0.65	0.65	0.55	0.59	0.55
F-LoRA	0.42	0.00	0.61	0.63	0.61	0.65	0.74	0.71	0.72	0.64	0.54	0.58	0.58	0.57	0.71	0.71	0.55	0.58	0.60
F-Turbo	0.45	0.63	0.00	0.74	0.70	0.67	0.70	0.72	0.69	0.64	0.65	0.76	0.74	0.57	0.65	0.65	0.65	0.70	0.59
K-Base	1.00	0.87	1.00	0.00	0.20	0.69	0.77	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00

Table A12 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Dangerous Animal In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.60	0.25	0.88	0.90	0.90	0.95	0.96	0.95	0.98	0.61	0.56	0.57	0.45	0.43	0.44	0.88	0.89	0.79
F-LoRA	0.57	0.00	0.66	0.76	0.75	0.83	0.85	0.86	0.87	0.89	0.92	0.90	0.91	0.66	0.83	0.83	0.79	0.75	0.65
F-Turbo	0.24	0.68	0.00	0.85	0.87	0.85	0.91	0.92	0.92	0.92	0.50	0.50	0.50	0.46	0.44	0.44	0.91	0.93	0.80
K-Base	0.88	0.80	0.87	0.00	0.10	0.32	0.51	0.53	0.56	0.52	0.89	0.95	0.94	0.87	0.89	0.89	0.70	0.56	0.58

Table A13 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Dangerous Animal In A Forest
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.70	0.32	0.77	0.70	0.77	0.90	0.86	0.74	0.98	0.86	0.98	0.98	0.78	0.57	0.57	0.45	0.58	0.35
F-LoRA	0.77	0.00	1.18	0.81	0.87	0.81	1.05	0.99	1.03	1.03	1.04	1.08	1.07	1.05	0.84	0.86	0.47	0.44	0.51
F-Turbo	0.30	1.01	0.00	0.89	0.79	0.87	0.90	0.88	0.76	0.95	0.85	0.93	0.93	0.81	0.89	0.88	0.76	0.89	0.67
K-Base	0.73	0.70	0.89	0.00	0.14	0.27	0.71	0.59	0.67	0.70	0.72	0.79	0.78	1.00	0.93	0.94	0.62	0.62	0.55

Table A14 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Dangerous Animal In A Grassland
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.28	0.18	0.88	0.84	0.97	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.93	0.93	0.74	0.69	0.78
F-LoRA	0.36	0.00	0.50	0.98	0.93	1.19	1.19	1.08	1.09	1.12	1.17	1.18	1.17	1.17	1.32	1.33	0.71	0.66	0.69
F-Turbo	0.16	0.35	0.00	0.87	0.85	0.97	0.88	0.88	0.89	0.90	0.87	0.90	0.90	0.85	0.79	0.79	0.75	0.74	0.81
K-Base	0.85	0.74	0.94	0.00	0.09	0.27	0.66	0.63	0.61	0.61	0.64	0.63	0.62	0.92	1.00	1.00	0.47	0.53	0.73

Table A15 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Frozen Single Fruit On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.18	0.20	0.48	0.50	0.48	1.00	1.00	1.00	0.99	0.92	0.92	0.91	0.83	0.81	0.80	0.73	0.81	0.55
F-LoRA	0.16	0.00	0.25	0.46	0.45	0.50	0.94	0.91	0.89	0.96	0.89	0.85	0.83	0.79	0.87	0.86	0.67	0.74	0.48
F-Turbo	0.17	0.24	0.00	0.44	0.48	0.49	0.88	0.90	0.85	0.86	0.80	0.79	0.78	0.79	0.82	0.82	0.62	0.69	0.47
K-Base	0.34	0.37	0.37	0.00	0.14	0.17	0.75	0.70	0.70	0.64	0.80	0.75	0.74	0.35	0.49	0.50	0.41	0.48	0.37

Table A16 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Frozen Single Fruit On A Dish
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.28	0.21	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.90	1.00	1.00	0.94	1.00	1.00	1.00	0.97	1.00
F-LoRA	0.22	0.00	0.33	0.81	0.79	0.78	0.80	0.84	0.80	0.83	0.89	0.94	0.95	0.81	0.84	0.84	0.81	0.83	0.80
F-Turbo	0.20	0.41	0.00	0.96	1.03	1.10	0.97	0.96	0.99	0.89	0.90	1.03	1.02	1.10	1.15	1.15	1.08	1.07	1.06
K-Base	0.65	0.66	0.64	0.00	0.27	0.42	0.53	0.54	0.54	0.55	0.83	0.81	0.79	1.00	0.85	0.85	0.83	0.90	0.70

Table A17 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Frozen Single Fruit On A Wooden Floor
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.18	0.26	0.32	0.34	0.30	0.86	0.87	0.88	0.67	0.88	0.87	0.86	0.76	0.77	0.76	0.38	0.39	0.38
F-LoRA	0.18	0.00	0.30	0.32	0.31	0.32	0.96	0.94	0.93	0.66	0.91	0.89	0.89	0.67	0.86	0.86	0.38	0.39	0.38
F-Turbo	0.25	0.27	0.00	0.37	0.39	0.37	0.98	0.97	1.02	0.69	0.85	0.82	0.82	0.66	1.05	1.04	0.42	0.44	0.40
K-Base	0.37	0.36	0.45	0.00	0.13	0.18	1.00	0.95	1.00	0.61	1.00	1.00	1.00	0.61	1.00	1.00	0.32	0.32	0.39

Table A18 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Peaceful Animal In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.43	0.27	0.71	0.74	0.72	0.92	1.00	1.00	0.97	0.69	0.69	0.69	0.66	0.70	0.69	0.85	0.81	0.88
F-LoRA	0.44	0.00	0.43	0.67	0.69	0.73	0.77	0.88	0.90	0.88	0.70	0.70	0.70	0.77	0.83	0.83	0.72	0.67	0.79
F-Turbo	0.27	0.42	0.00	0.71	0.76	0.74	0.91	1.03	1.02	0.99	0.69	0.67	0.68	0.69	0.70	0.69	0.82	0.80	0.83
K-Base	0.86	0.79	0.88	0.00	0.11	0.51	0.68	0.86	0.90	0.69	1.00	1.00	1.00	0.95	0.94	0.94	0.59	0.54	0.81

Table A19 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Peaceful Animal In A Forest
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.65	0.28	0.57	0.57	0.58	0.94	1.00	0.96	0.86	0.92	1.00	1.00	0.91	0.79	0.79	0.84	0.89	0.84
F-LoRA	0.56	0.00	0.58	0.66	0.67	0.81	0.84	0.84	0.82	0.76	0.85	0.83	0.82	0.86	0.71	0.72	0.71	0.68	0.59
F-Turbo	0.26	0.63	0.00	0.60	0.59	0.54	0.85	0.89	0.88	0.82	0.87	0.94	0.93	0.79	0.66	0.66	0.77	0.84	0.78
K-Base	0.69	0.93	0.77	0.00	0.14	0.45	0.91	0.93	0.92	0.80	0.95	0.98	0.97	1.00	1.00	1.00	0.86	0.89	0.88

Table A20 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Peaceful Animal In A Grassland
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.36	0.21	0.80	0.78	0.82	0.98	0.92	0.87	1.00	1.00	1.00	1.00	0.68	0.47	0.49	0.68	0.62	0.66
F-LoRA	0.35	0.00	0.36	0.94	0.93	1.06	1.02	0.85	0.87	1.04	0.98	0.96	0.96	0.77	0.65	0.67	0.73	0.63	0.58
F-Turbo	0.22	0.39	0.00	0.86	0.86	0.85	0.99	0.95	0.88	1.03	1.04	1.03	1.02	0.70	0.46	0.47	0.72	0.67	0.67
K-Base	0.80	0.97	0.83	0.00	0.14	0.54	1.00	1.00	1.00	0.77	0.98	0.99	0.98	1.00	1.00	1.00	1.00	1.00	1.00

Table A21 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Baked Good Against A Brick Wall
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.11	0.11	1.00	1.00	1.00	1.00	0.56	0.66	0.71	0.76	1.00	1.00	0.34	0.48	0.51	0.93	0.91	1.00
F-LoRA	0.13	0.00	0.18	1.16	1.17	1.12	1.16	0.65	0.76	0.94	0.79	1.07	1.06	0.27	0.38	0.41	0.94	0.93	1.08
F-Turbo	0.09	0.14	0.00	0.87	0.87	0.84	0.85	0.46	0.55	0.68	0.64	0.82	0.82	0.32	0.43	0.45	0.81	0.79	0.83
K-Base	1.00	1.00	1.00	0.00	0.06	0.31	0.95	1.00	1.00	0.86	1.00	0.56	0.59	1.00	1.00	1.00	1.00	1.00	0.67

Table A22 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Baked Good On A Dark Wood Surface
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.20	0.18	1.00	1.00	0.93	1.00	0.87	1.00	0.32	1.00	1.00	1.00	0.35	0.81	0.82	1.00	1.00	1.00
F-LoRA	0.16	0.00	0.22	0.84	0.86	0.68	0.81	0.65	0.75	0.30	0.79	0.77	0.76	0.32	0.59	0.59	0.78	0.79	0.86
F-Turbo	0.14	0.22	0.00	0.74	0.77	0.69	0.78	0.69	0.77	0.37	0.80	0.76	0.76	0.42	0.67	0.68	0.79	0.78	0.81
K-Base	0.80	0.85	0.76	0.00	0.09	0.78	0.43	1.00	0.90	0.94	0.68	0.52	0.51	0.96	1.00	1.00	0.32	0.31	0.28

Table A23 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Baked Good On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.43	0.37	1.00	1.00	0.82	1.00	1.00	1.00	0.63	0.82	1.00	1.00	0.38	0.57	0.57	1.00	1.00	1.00
F-LoRA	0.29	0.00	0.41	0.71	0.74	0.56	0.65	0.63	0.65	0.56	0.51	0.61	0.60	0.39	0.48	0.47	0.55	0.57	0.65
F-Turbo	0.26	0.43	0.00	0.63	0.64	0.54	0.69	0.68	0.71	0.61	0.64	0.67	0.68	0.44	0.53	0.53	0.72	0.70	0.67
K-Base	0.95	1.00	0.84	0.00	0.14	0.43	0.47	0.92	0.97	0.91	1.00	0.71	0.70	1.00	1.00	1.00	0.59	0.49	0.39

Table A24 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Single Fruit On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.29	0.23	1.00	1.00	1.00	0.75	0.86	1.00	1.00	1.00	1.00	1.00	0.83	0.74	0.73	1.00	1.00	1.00
F-LoRA	0.25	0.00	0.28	0.78	0.80	0.89	0.73	0.79	0.89	0.90	0.84	0.83	0.84	0.72	0.65	0.66	0.89	0.85	0.92
F-Turbo	0.24	0.34	0.00	1.16	1.20	1.16	0.58	0.77	0.92	1.09	1.20	1.20	1.20	1.13	0.91	0.90	1.03	0.98	0.98
K-Base	0.91	0.83	1.00	0.00	0.21	0.54	1.00	1.00	0.96	0.71	0.56	0.56	0.59	0.69	0.92	0.98	0.71	0.70	0.96

Table A25 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Single Fruit On A Dish
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.28	0.33	0.76	0.90	1.00	0.63	0.98	0.91	0.93	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00
F-LoRA	0.22	0.00	0.29	0.65	0.79	0.79	0.58	0.81	0.79	0.79	0.79	0.75	0.76	0.82	0.77	0.77	0.81	0.80	0.85
F-Turbo	0.27	0.30	0.00	0.60	0.72	0.70	0.54	0.80	0.74	0.79	0.80	0.81	0.80	0.86	0.83	0.83	0.85	0.82	0.88
K-Base	0.51	0.54	0.49	0.00	0.25	0.49	0.43	0.60	0.56	0.63	0.59	0.60	0.60	0.73	0.71	0.71	0.65	0.60	0.69

Table A26 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Savory Single Fruit On A Wooden Floor
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.27	0.28	0.64	0.64	0.50	0.58	0.65	0.80	1.00	0.96	0.77	0.77	0.81	0.78	0.78	0.89	1.00	0.69
F-LoRA	0.27	0.00	0.28	0.72	0.69	0.64	0.64	0.69	0.77	0.96	0.93	0.78	0.78	0.64	0.56	0.56	0.83	0.91	0.66
F-Turbo	0.29	0.30	0.00	0.94	0.91	0.81	0.49	0.56	0.71	1.10	0.89	0.69	0.69	0.86	0.83	0.83	0.85	0.94	0.81
K-Base	0.60	0.67	0.82	0.00	0.11	0.21	0.62	0.62	0.69	0.57	0.78	0.73	0.71	0.66	0.60	0.60	0.93	0.93	0.73

Table A27 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Baked Good Against A Brick Wall
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.22	0.23	1.00	0.70	0.51	1.00	1.00	1.00	0.90	1.00	1.00	1.00	0.41	0.71	0.74	0.88	1.00	0.68
F-LoRA	0.22	0.00	0.34	1.05	0.85	0.64	0.97	1.01	1.01	0.93	0.89	0.88	0.89	0.30	0.52	0.55	0.96	1.03	0.79
F-Turbo	0.19	0.28	0.00	0.78	0.62	0.50	0.84	0.84	0.83	0.83	0.80	0.79	0.79	0.42	0.67	0.69	0.67	0.76	0.51
K-Base	0.68	0.71	0.63	0.00	0.32	0.54	0.59	0.63	0.67	0.83	0.85	0.66	0.66	0.75	0.82	0.81	0.70	0.70	0.61

Table A28 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Baked Good On A Dark Wood Surface
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.13	0.12	1.00	1.00	1.00	1.00	1.00	0.77	0.24	0.60	0.67	0.67	0.27	0.47	0.46	1.00	1.00	1.00
F-LoRA	0.11	0.00	0.14	0.85	0.85	0.82	0.82	0.77	0.63	0.26	0.50	0.56	0.55	0.30	0.37	0.36	0.85	0.84	0.83
F-Turbo	0.10	0.14	0.00	0.77	0.78	0.76	0.79	0.74	0.62	0.25	0.45	0.52	0.52	0.32	0.42	0.41	0.83	0.83	0.80
K-Base	0.52	0.52	0.48	0.00	0.11	0.28	0.43	0.65	0.73	0.61	0.64	0.62	0.63	0.61	0.49	0.48	0.33	0.34	0.31

Table A29 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Baked Good On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.16	0.14	0.77	0.76	0.68	1.00	1.00	0.55	0.38	0.27	0.42	0.41	0.14	0.23	0.23	0.99	1.00	0.96
F-LoRA	0.13	0.00	0.15	0.70	0.72	0.69	0.67	0.65	0.35	0.33	0.30	0.32	0.32	0.19	0.22	0.22	0.77	0.77	0.74
F-Turbo	0.10	0.14	0.00	0.59	0.61	0.55	0.72	0.69	0.39	0.30	0.23	0.31	0.30	0.19	0.26	0.25	0.74	0.75	0.70
K-Base	0.45	0.53	0.46	0.00	0.10	0.23	0.56	0.75	0.61	0.49	0.47	0.55	0.53	0.50	0.59	0.58	0.35	0.36	0.34

Table A30 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Single Fruit On A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.31	0.33	0.27	0.24	0.26	0.80	0.97	0.94	0.61	0.96	1.00	1.00	0.58	0.54	0.54	0.82	0.97	0.80
F-LoRA	0.29	0.00	0.42	0.44	0.39	0.38	0.84	0.88	0.95	0.73	0.75	0.77	0.77	0.59	0.63	0.65	0.75	0.84	0.77
F-Turbo	0.35	0.47	0.00	0.60	0.57	0.57	0.64	0.93	0.88	0.57	1.03	1.06	1.06	0.88	0.76	0.75	0.89	1.04	0.82
K-Base	0.25	0.43	0.52	0.00	0.08	0.17	0.92	0.96	0.93	0.66	0.78	0.85	0.85	0.42	0.46	0.46	0.78	0.91	0.82

Table A31 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Single Fruit On A Dish
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.26	0.49	0.33	0.33	0.37	0.96	0.97	0.94	0.98	0.65	0.78	0.77	0.43	0.49	0.50	0.51	0.52	0.56
F-LoRA	0.27	0.00	0.62	0.43	0.43	0.40	0.96	0.97	0.98	1.03	0.62	0.75	0.74	0.46	0.57	0.57	0.48	0.51	0.63
F-Turbo	0.37	0.46	0.00	0.56	0.55	0.59	0.73	0.76	0.72	0.74	0.68	0.77	0.78	0.62	0.59	0.60	0.70	0.70	0.67
K-Base	0.35	0.44	0.78	0.00	0.12	0.18	1.00	1.00	1.00	1.00	0.47	0.66	0.65	0.27	0.39	0.39	0.31	0.36	0.38

Table A32 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Sweet Single Fruit On A Wooden Floor
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.17	0.35	0.34	0.33	0.36	0.71	0.77	0.74	0.88	0.87	0.84	0.84	0.43	0.34	0.34	0.45	0.51	0.37
F-LoRA	0.21	0.00	0.50	0.36	0.34	0.43	0.86	0.93	0.94	1.14	1.03	0.97	0.98	0.33	0.37	0.36	0.62	0.67	0.40
F-Turbo	0.30	0.34	0.00	0.63	0.60	0.62	0.55	0.63	0.62	1.00	0.88	0.81	0.82	0.55	0.52	0.52	0.55	0.62	0.47
K-Base	0.46	0.39	1.00	0.00	0.06	0.15	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.33	0.42	0.41	0.62	0.70	0.43

Table A33 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Tropical Single Flower In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.39	0.11	0.53	0.63	0.49	0.69	1.00	0.92	0.56	1.00	1.00	1.00	1.00	0.70	0.73	0.64	0.73	0.58
F-LoRA	0.28	0.00	0.28	0.59	0.62	0.62	0.91	1.14	1.00	0.79	0.72	0.71	0.71	0.67	0.74	0.76	0.65	0.68	0.62
F-Turbo	0.11	0.38	0.00	0.57	0.66	0.52	0.70	1.00	0.91	0.57	0.96	0.96	0.96	0.93	0.72	0.75	0.61	0.67	0.57
K-Base	0.39	0.61	0.43	0.00	0.15	0.25	0.56	0.80	0.78	0.54	0.73	0.73	0.74	0.87	0.64	0.66	0.57	0.67	0.39

Table A34 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Tropical Single Flower On A Pot
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.49	0.23	0.80	0.78	0.59	0.93	0.93	0.97	0.91	0.85	0.89	0.89	0.66	0.48	0.49	0.63	0.68	0.70
F-LoRA	0.52	0.00	0.60	1.04	0.93	0.75	1.19	1.03	1.04	1.16	0.81	0.87	0.88	0.47	0.73	0.72	0.77	0.84	0.80
F-Turbo	0.21	0.50	0.00	0.72	0.71	0.55	0.84	0.83	0.88	0.77	0.75	0.78	0.78	0.66	0.44	0.45	0.60	0.64	0.67
K-Base	0.57	0.69	0.57	0.00	0.23	0.41	0.63	0.66	0.65	0.63	0.64	0.61	0.61	0.77	0.65	0.65	0.62	0.63	0.68

Table A35 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Tropical Single Flower On A Vase
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.41	0.27	0.50	0.51	0.44	0.75	0.77	0.70	0.63	0.46	0.47	0.47	0.60	0.51	0.51	0.68	0.72	0.65
F-LoRA	0.52	0.00	0.61	0.59	0.47	0.64	1.08	1.00	0.84	0.94	0.39	0.42	0.42	0.65	0.65	0.64	0.73	0.76	0.75
F-Turbo	0.27	0.48	0.00	0.51	0.53	0.38	0.73	0.74	0.68	0.55	0.51	0.51	0.52	0.57	0.47	0.46	0.64	0.67	0.63
K-Base	0.49	0.46	0.50	0.00	0.22	0.32	0.82	0.79	0.64	0.77	0.41	0.42	0.42	0.52	0.51	0.50	0.50	0.52	0.60

Table A36 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Vibrant Single Flower In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.34	0.15	0.32	0.36	0.35	0.97	0.93	0.94	0.94	0.93	0.90	0.90	0.50	0.39	0.40	0.86	0.91	0.72
F-LoRA	0.30	0.00	0.40	0.58	0.62	0.65	1.19	1.06	1.02	0.92	0.86	0.79	0.78	0.69	0.68	0.70	0.76	0.78	0.83
F-Turbo	0.14	0.42	0.00	0.35	0.38	0.36	0.84	0.81	0.83	0.87	0.87	0.85	0.85	0.50	0.39	0.39	0.80	0.84	0.72
K-Base	0.36	0.73	0.41	0.00	0.08	0.10	0.96	1.00	1.00	1.00	1.00	1.00	1.00	0.42	0.26	0.26	1.00	1.00	0.62

Table A37 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Vibrant Single Flower On A Pot
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.31	0.21	0.34	0.34	0.43	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.47	0.86	0.86	0.86	0.90	0.85
F-LoRA	0.31	0.00	0.39	0.51	0.52	0.60	1.01	1.03	1.00	1.05	1.05	1.05	1.05	0.47	0.81	0.81	0.80	0.87	0.90
F-Turbo	0.19	0.36	0.00	0.33	0.35	0.39	0.95	0.95	0.94	0.95	0.89	0.90	0.90	0.47	0.72	0.73	0.76	0.81	0.69
K-Base	0.31	0.47	0.32	0.00	0.15	0.23	0.95	0.93	0.91	0.91	0.73	0.76	0.77	0.41	0.72	0.72	0.65	0.74	0.65

Table A38 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Vibrant Single Flower On A Vase
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.41	0.20	0.54	0.56	0.89	1.00	1.00	1.00	0.95	1.00	1.00	1.00	0.79	0.76	0.76	1.00	1.00	1.00
F-LoRA	0.38	0.00	0.46	0.77	0.65	1.01	1.02	1.00	0.96	1.01	0.94	0.90	0.90	0.73	0.65	0.65	0.89	0.89	0.91
F-Turbo	0.18	0.45	0.00	0.51	0.55	0.78	0.90	0.90	0.91	0.86	0.91	0.92	0.92	0.70	0.72	0.72	0.90	0.90	0.89
K-Base	0.41	0.62	0.42	0.00	0.17	0.31	0.73	0.69	0.73	0.64	0.75	0.75	0.75	0.65	0.64	0.63	0.60	0.59	0.62

Table A39 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Wild Animal In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.51	0.17	1.00	1.00	0.98	1.00	1.00	1.00	1.00	0.62	0.56	0.55	0.43	0.72	0.70	1.00	1.00	0.98
F-LoRA	0.40	0.00	0.42	0.79	0.71	0.91	0.79	0.79	0.79	0.81	0.53	0.51	0.52	0.49	0.84	0.83	0.65	0.67	0.70
F-Turbo	0.18	0.58	0.00	1.18	1.17	1.08	1.09	1.09	1.09	1.12	0.71	0.64	0.63	0.49	0.80	0.78	1.09	1.14	1.07
K-Base	0.60	0.60	0.65	0.00	0.13	0.50	0.47	0.49	0.53	0.46	0.53	0.58	0.57	0.55	0.53	0.52	0.64	0.51	0.82

Table A40 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Wild Animal In A Forest
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.38	0.25	0.74	0.69	0.53	1.00	0.97	1.00	1.00	1.00	1.00	1.00	0.93	0.86	0.87	0.80	0.72	0.60
F-LoRA	0.39	0.00	0.51	0.63	0.65	0.58	1.08	0.97	0.92	1.12	1.01	1.00	1.01	1.02	1.10	1.11	0.83	0.72	0.54
F-Turbo	0.24	0.48	0.00	0.80	0.74	0.53	0.97	0.93	0.99	0.92	0.95	0.96	0.96	0.88	0.77	0.76	0.77	0.77	0.64
K-Base	0.70	0.57	0.78	0.00	0.17	0.33	0.90	0.81	0.78	0.88	0.86	0.91	0.92	1.00	1.00	1.00	0.94	0.86	0.61

Table A41 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of A Wild Animal In A Grassland
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.23	0.15	1.00	1.00	0.96	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.88	0.92	1.00	1.00	1.00
F-LoRA	0.22	0.00	0.21	1.10	1.04	1.07	1.00	0.96	0.96	1.02	0.88	0.90	0.90	0.98	1.07	1.11	0.93	0.90	0.82
F-Turbo	0.14	0.21	0.00	1.03	1.01	1.01	0.96	0.95	0.95	0.98	0.94	0.95	0.95	0.94	0.86	0.90	0.99	0.99	0.96
K-Base	0.62	0.71	0.68	0.00	0.22	0.31	0.70	0.61	0.60	0.61	0.63	0.57	0.55	0.79	0.75	0.78	0.45	0.45	0.66

Table A42 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Dangerous Bird In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.23	0.20	0.47	0.46	1.00	0.97	0.85	0.89	0.73	0.94	0.91	0.92	0.68	0.56	0.56	0.70	0.74	0.51
F-LoRA	0.38	0.00	0.49	0.58	0.57	1.45	1.41	1.37	1.42	1.17	1.52	1.52	1.53	1.28	1.05	1.06	1.31	1.34	0.98
F-Turbo	0.18	0.26	0.00	0.45	0.43	0.91	0.92	0.79	0.82	0.70	0.84	0.82	0.84	0.62	0.52	0.52	0.62	0.68	0.46
K-Base	0.52	0.38	0.54	0.00	0.12	0.80	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.81

Table A43 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Dangerous Bird On A Grass
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.45	0.32	0.88	0.91	0.96	0.93	0.92	0.93	1.00	0.97	1.00	1.00	0.60	0.55	0.54	0.60	0.56	0.68
F-LoRA	0.41	0.00	0.51	0.81	0.87	1.04	0.94	0.93	0.87	0.98	0.99	0.99	0.99	0.76	0.71	0.71	0.70	0.65	0.76
F-Turbo	0.27	0.47	0.00	0.75	0.78	0.83	0.83	0.85	0.85	0.90	0.85	0.87	0.87	0.56	0.49	0.49	0.58	0.58	0.66
K-Base	1.00	1.00	1.00	0.00	0.21	0.41	1.00	1.00	1.00	0.97	0.80	0.87	0.86	1.00	1.00	1.00	1.00	1.00	1.00

Table A44 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Dangerous Bird On A Savana
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.49	0.27	0.56	0.56	0.55	0.66	0.58	0.72	0.71	0.55	0.52	0.53	0.53	0.58	0.59	0.55	0.53	0.49
F-LoRA	0.51	0.00	0.39	0.60	0.62	0.61	0.47	0.39	0.48	0.51	0.51	0.51	0.52	0.75	0.72	0.73	0.57	0.56	0.54
F-Turbo	0.44	0.61	0.00	0.84	0.82	0.79	0.71	0.57	0.85	0.72	0.45	0.41	0.43	0.61	0.76	0.76	0.87	0.86	0.82
K-Base	0.85	0.87	0.78	0.00	0.12	0.20	0.84	0.85	0.91	0.81	0.79	0.81	0.81	0.71	0.77	0.78	0.27	0.25	0.25

Table A45 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Flightless Bird In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.59	0.17	0.86	0.87	0.81	0.43	0.42	0.68	0.44	0.45	0.39	0.39	0.62	0.81	0.81	0.84	0.84	0.80
F-LoRA	0.36	0.00	0.38	0.53	0.51	0.56	0.44	0.48	0.53	0.43	0.51	0.48	0.48	0.73	0.85	0.87	0.71	0.68	0.61
F-Turbo	0.26	0.90	0.00	1.26	1.29	1.24	0.54	0.52	0.94	0.72	0.37	0.36	0.36	1.11	1.36	1.36	1.37	1.37	1.33
K-Base	1.00	1.00	1.00	0.00	0.12	0.19	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.85	0.93	0.96	0.56	0.59	0.53

Table A46 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Flightless Bird On A Grass
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.22	0.18	0.86	0.87	0.86	0.49	0.53	0.56	0.58	0.38	0.40	0.40	0.83	0.85	0.86	0.92	0.91	0.87
F-LoRA	0.34	0.00	0.20	1.14	1.19	1.15	0.69	0.78	0.81	0.88	0.52	0.55	0.55	1.35	1.25	1.27	1.45	1.41	1.32
F-Turbo	0.26	0.18	0.00	1.05	1.10	1.05	0.63	0.69	0.74	0.78	0.43	0.45	0.45	1.15	1.14	1.14	1.29	1.27	1.21
K-Base	0.85	0.74	0.75	0.00	0.11	0.25	1.00	1.00	1.00	1.00	0.90	0.91	0.90	0.97	0.72	0.77	0.62	0.58	0.59

Table A47 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Flightless Bird On A Savana
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.38	0.28	0.39	0.39	0.38	0.53	0.50	0.45	0.40	0.55	0.54	0.54	0.42	0.46	0.46	0.37	0.37	0.31
F-LoRA	0.38	0.00	0.40	0.53	0.57	0.52	0.62	0.58	0.55	0.52	0.58	0.54	0.54	0.68	0.68	0.69	0.60	0.60	0.49
F-Turbo	0.28	0.40	0.00	0.54	0.54	0.52	0.56	0.56	0.53	0.49	0.48	0.46	0.46	0.51	0.54	0.55	0.52	0.51	0.43
K-Base	0.72	0.98	1.00	0.00	0.08	0.15	0.96	0.74	0.77	0.50	1.00	1.00	1.00	0.61	0.48	0.49	0.29	0.30	0.29

Table A48 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Peaceful Bird In A Dimmed Studio
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.64	0.35	0.86	0.86	0.87	0.88	0.73	0.68	0.66	0.86	0.86	0.86	0.73	0.64	0.65	0.94	0.84	0.86
F-LoRA	0.68	0.00	0.63	0.67	0.72	0.80	0.94	0.86	0.83	0.69	0.96	0.94	0.96	0.75	0.86	0.87	0.66	0.59	0.62
F-Turbo	0.39	0.65	0.00	0.75	0.76	0.82	0.91	0.76	0.67	0.83	0.81	0.86	0.86	0.84	0.62	0.62	0.94	0.92	0.90
K-Base	0.96	0.71	0.77	0.00	0.33	0.46	0.93	0.91	0.85	0.89	0.99	1.00	1.00	1.00	0.66	0.68	0.78	0.77	0.71

Table A49 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Peaceful Bird On A Grass
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.36	0.52	0.67	0.70	0.78	0.74	0.81	0.70	0.95	0.74	0.77	0.77	1.00	0.85	0.85	0.87	0.76	0.66
F-LoRA	0.39	0.00	0.47	0.65	0.71	0.80	0.80	0.87	0.77	1.01	0.84	0.84	0.84	1.06	0.94	0.93	0.85	0.73	0.66
F-Turbo	0.52	0.44	0.00	0.78	0.79	0.79	0.71	0.74	0.72	0.93	0.77	0.73	0.72	0.89	0.84	0.83	0.76	0.66	0.56
K-Base	0.87	0.78	1.00	0.00	0.13	0.26	1.00	1.00	1.00	1.00	1.00	1.00	1.00	0.92	0.84	0.81	1.00	1.00	1.00

Table A50 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

	Prompt: A Photo Of An Peaceful Bird On A Savana
	Flux			Kand			SD1.5				SD2.1			SD3			SDXL
	F-Base	F-LoRA	F-Turbo	K-Base	K-Naru	K-Poke	1.5-v1.2	1.5-v1.4	1.5-Base	1.5-Dream	2.1-DPO	2.1-Base	2.1-Art	SD3-Real	SD3-Base	SD3-Anim	XL-DPO	XL-Base	XL-Light
F-Base	0.00	0.40	0.27	0.37	0.37	0.32	0.83	0.84	0.92	0.90	0.82	0.82	0.81	0.73	0.80	0.80	0.30	0.30	0.37
F-LoRA	0.37	0.00	0.45	0.45	0.46	0.51	0.78	0.78	0.79	0.78	0.96	0.89	0.88	0.91	0.90	0.89	0.36	0.35	0.37
F-Turbo	0.25	0.45	0.00	0.40	0.38	0.40	0.78	0.76	0.84	0.92	0.76	0.72	0.70	0.63	0.72	0.71	0.38	0.36	0.43
K-Base	0.37	0.48	0.44	0.00	0.14	0.28	0.90	0.89	0.90	1.00	0.86	0.88	0.88	0.89	0.98	0.97	0.32	0.31	0.34

Table A51 : Normalized Wasserstein Distance Matrix. Each column normalized by max distance to base models. xx Short, xx Medium-low, xx Medium-high, xx Long distance.

Panel of Figure 67 — Figure B7: Qualitative results (Part 1): Flux, Kandinsky, and SD1.5 families.

Panel of Figure 68 — Figure B8: Qualitative results (Part 2): SD2.1, SD3, and SDXL families.

Panel of Figure 69 — Figure B9: Qualitative results (Part 1): Flux, Kandinsky, and SD1.5 families.

Panel of Figure 70 — Figure B10: Qualitative results (Part 2): SD2.1, SD3, and SDXL families.

Panel of Figure 71 — Figure B11: Qualitative results (Part 1): Flux, Kandinsky, and SD1.5 families.

Panel of Figure 72 — Figure B12: Qualitative results (Part 2): SD2.1, SD3, and SDXL families.

Panel of Figure 73 — Figure B13: Qualitative results (Part 1): Flux, Kandinsky, and SD1.5 families.

Panel of Figure 74 — Figure B14: Qualitative results (Part 2): SD2.1, SD3, and SDXL families.

T2I Diffusion Models

Text-to-image (T2I) diffusion models generate images from prompts by learning to reverse noise under text conditioning.

See all 403 papers

Aug 12, 2026

Fingerprinting Text-to-Image Diffusion Models via Collapsed Generation

Proprietary text-to-image diffusion models are increasingly distributed as hosted services and downloadable checkpoints, making their intellectual property (IP) protection an increasingly critical concern when model leakage, copying, or unauthorized fine-tuning is disputed. In this work, we present a non-invasive model fingerprinting framework based on collapsed generation, a phenomenon where certain input conditions produce highly consistent images across multiple stochastic seeds. We show that collapsed generation is an intrinsic, model-dependent property of the learned generation process. These collapse-prone conditions therefore expose model-specific behavioral signatures, enabling reliable ownership verification without embedding invasive watermarks. After preparing conditions on the source model, the framework verifies a suspect model under two access settings: (1) white-box pipeline access, where optimized continuous embeddings can be injected into the generation process, and (2) black-box API-only access, where natural language prompts are queried through the service interface. In both cases, ownership evidence is measured by whether the suspect model reproduces the source model's collapse behavior across stochastic samplings. Extensive experiments across UNet- and transformer-based diffusion models show that collapsed generation fingerprints can distinguish different source models with low confusion. These fingerprints remain verifiable in fine-tuned derivatives and under common and adaptive model- or query-level obfuscations, while requiring only a modest verification query budget. Together, these results establish collapsed generation as a reliable intrinsic evidence source for non-invasive diffusion model ownership verification.

Yuanmin Huang, Chen Chen, Geng Hong +5

Fudan University, Shanghai, China · East China University of Science and Technology, Shanghai, China · Alibaba Group, Hangzhou, China +1

T2I Diffusion Models Diffusion Models

Ask PDF

Apr 18, 2026

SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models

The public accessibility of large vision-language models (LVLMs) raises serious concerns about unauthorized model reuse and intellectual property infringement. Existing ownership verification methods often rely on semantically abnormal queries or out-of-distribution responses as fingerprints, which can be easily detected and removed by adversaries. We expose this vulnerability through a Semantic Divergence Attack (SDA), which identifies and filters fingerprint queries by measuring semantic divergence between a suspect model and a reference model, showing that existing fingerprints are not semantic-preserving and are therefore easy to detect and bypass. To address these limitations, we propose SIF (Semantically In-Distribution Fingerprints), a non-intrusive ownership verification framework that requires no parameter modification. SIF introduces Semantic-Aligned Fingerprint Distillation (SAFD), which transfers text watermarking signals into the visual modality to produce semantically coherent yet fingerprinted responses. In addition, Robust-Fingerprint Optimization (RFO) enhances robustness by simulating worst-case representation perturbations, making the fingerprints resilient to model modifications such as fine-tuning and quantization. Extensive experiments on LLaVA-1.5 and Qwen2.5-VL demonstrate that SIF achieves strong stealthiness and robustness, providing a practical solution for LVLM copyright protection. Code is available at https://github.com/UCF-ML-Research/SIF-VLM-Fingerprint

Yifei Zhao, Qian Lou, Mengxin Zheng

University of Central Florida

VLM Robustness Language Model Fingerprinting Large Vision-Language Models LLM Watermarking

Ask PDF

Sep 30, 2026

Persistent Watermarking of Text-to-Image Models

Text-to-image (T2I) generation is gaining increasing popularity with the general public, motivating the development of reliable mechanisms for copyrighting such models given their expensive training costs. An adversary may obtain and reuse a pretrained T2I model without authorization, and then serve a modified version through an API service. Such modifications may arise from ordinary downstream adaptation or deliberate attempts to erase ownership, including input-prompt preprocessing, model fine-tuning, and output post-processing. From the model owner's perspective, a key challenge is therefore to embed trigger data that remain persistent under such changes while preserving the model's normal image-generation capabilities. In this work, we propose a contrastive-style watermarking objective with a term that explicitly encourages the watermarked model to behave differently from the original model on trigger inputs. Experiments show substantially stronger trigger-data persistence than prior methods across a wide range of downstream modifications and deliberate attempts to weaken the watermark, resulting in higher detection rates, often approaching 100% TPR@FPR<

10^{-4}

Dixi Yao, Kaiwen Chen, Tahseen Rabbani +1

University of Chicago

Diffusion Model Watermarking Robust Watermarking T2I Diffusion Models AI-Generated Image Watermarking

Ask PDF

Abstract

T2I Diffusion Models

Explore similar work

Fingerprinting Text-to-Image Diffusion Models via Collapsed Generation

SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models

Persistent Watermarking of Text-to-Image Models