GAN-Blot: A Controllable Structure-Style Synthesis Benchmark for Western Blot Forensics

Hao-Chiang Shao; Fong-Yi Lin; Te-An Chien; TianYu Chen; Da-Jhong Chen

cs.CVSep 6, 2026

GAN-Blot: A Controllable Structure-Style Synthesis Benchmark for Western Blot Forensics

Authors: Hao-Chiang Shao, Fong-Yi Lin, Te-An Chien, TianYu Chen, Da-Jhong Chen

Organizations: Institute of Data Science and Information Computing, National Chung Hsing University, Taiwan · Department of Electrical Engineering, National Taiwan University of Science and Technology, Taiwan

Abstract

Western blot (WB) images are widely used as key evidence in biomedical research. Recent scientific misconduct cases reveal that WB imagery is increasingly fabricated, making WB forensics a major concern for research integrity. However, while the progress of forensic detection techniques often relies on advances in forgery-generation techniques, the development of WB forensic techniques has been hindered by the lack of standardized appearance attribute definitions, image datasets, and controllable generation frameworks for WB imagery. To address this limitation, we present a controllable WB image synthesis framework, named GAN-Blot, for generating realistic synthetic WB images. We introduce a formulation that decomposes a WB image into a structure component and a style-reference component, enabling independent control over local protein-band geometry and the global visual appearance of a synthetic WB image. GAN-Blot integrates a dual-path autoencoding design with several style-alignment loss terms to enable implicit control over structure-style synthesis without predefined semantic appearance attributes. We further contribute a synthetic WB dataset containing more than 46K images and propose four evaluation protocols for controllable WB synthesis. Extensive experiments show that GAN-Blot can generate WB images with high fidelity in both protein-band structure and visual style. Under blind inspection, the generated images can fool domain experts and are not reliably distinguished from authentic WB images by existing detectors and screening platforms. These results demonstrate their utility as challenging controlled cases for validating and developing WB forensic methods.

Figures & tables

Fig. 1: Authentic or Synthetic? Western blot (WB) images are digital recordings of molecular signal deposition rather than visual scenes. Consequently, faint bands, smooth backgrounds, and blurred band boundaries can arise in authentic WB imaging and should not be interpreted as evidence of image synthesis. Applying natural-image priors to WB images can lead to incorrect forensic judgments, motivating a generation-first benchmark for scientific-image forensics. Readers are invited to identify each of the 16 rows as authentic or synthetic. The answer is provided on the next page.

Fig. 2: Domain-specific visual attributes of a Western blot (WB) image. Key WB attributes include i) the numbers of lanes, bands, and rows; ii) foreground band morphology; iii) lane-trace contour and intensity; and iv) background characteristics such as intensity, noise patterns, and gel surface artifacts/textures. A lane denotes a vertical electrophoretic channel where a sample is separated by molecular weight. A band is a horizontal intensity region within a lane representing proteins of similar molecular weight. The trace , indicated by yellow arrows, refers to the faint vertical boundary or intensity profile outlining a lane. The lack of standardized definitions for these attributes makes controllable WB synthesis difficult and motivates our structure–style formulation.

Fig. 3: Training framework of GAN-Blot. The model is trained with triplet training samples, each consisting of a randomly sampled style reference $\mathcal{S}$ , a WB image $\mathcal{I}$ , and its band-contour mask $\mathcal{J}$ in a dual-path manner. While the synthesis path learns structure–style synthesis, the self-supervision path enforces accurate reconstruction of $\mathcal{I}$ and consistency of its corresponding band mask, facilitating stable learning.

Fig. 4: Architecture of the WB-Synthesizer in GAN-Blot. (A) The WB-Synthesizer extends StyleGAN2 by integrating a band-contour encoder $\phi$ and a style encoder $\Theta$ . (B) Architecture of $\phi$ . (C) Modified ConvNeXt block within $\phi$ .

Fig. 5: Examples of WB images generated by GAN-Blot. The leftmost column shows the input protein-band mask $\mathcal{J}$ , and the top row shows the input style reference $\mathcal{S}$ . GAN-Blot is capable of generating WB images with independently controllable band structures and appearance styles.

Fig. 6: WB images synthesized by (A) StyleGAN3 [ 56 ] , (B) ADM [ 57 ] , (C) InstructPix2Pix [ 32 ] , and (D) ControlNet [ 31 ] . Results from StyleGAN3 and ADM show that random noise inputs cannot achieve structure or style control. In contrast, results from InstructPix2Pix and ControlNet suggest that text- or instruction-based methods cannot reliably synthesize WB images, because WB appearance lacks well-defined semantic attribute descriptions and these approaches rely primarily on textual instructions without explicit style-reference inputs for controlling WB appearance.

(A) Module Ablation: Performance Comparison among Different Model Architectures
Model	MSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	IoU ( $\tilde{\mathcal{J}}$ ) $\uparrow$	Perc. Loss $\downarrow$
GAN-Blot	.0047 $\pm$ .0039	24.57 $\pm$ 3.55	.8322 $\pm$ .0969	.1131 $\pm$ .0814	.9978 $\pm$ .0188	.1101 $\pm$ .0972
StyleGAN2-JS	.0250 $\pm$ .0230	17.17 $\pm$ 2.97	.6894 $\pm$ .1042	.2835 $\pm$ .1108	.8694 $\pm$ .1720	.0661 $\pm$ .1537
StyleGAN2-J	.0518 $\pm$ .0419	14.11 $\pm$ 3.40	.6036 $\pm$ .1693	.4518 $\pm$ .1588	.4456 $\pm$ .1476	.3562 $\pm$ .4278
CycleGAN	.1214 $\pm$ .0807	9.91 $\pm$ 2.53	.2792 $\pm$ .1814	.5837 $\pm$ .1434	.0130 $\pm$ .0198	.3117 $\pm$ .3117
(B) Loss-term Ablation: Impact of Each Loss

TABLE I: Performance comparison and ablation study via WB re-synthesis. (A) Performance comparison among network architectures and (B) Loss-term ablation. All values are computed on images whose intensities are normalized to the [0,1] range

Fig. 7: Visualization of Protocol-P3 evaluation. This figure presents the reconstruction results corresponding to Table I (A). Since StyleGAN2-JS and GAN-Blot’s WB-Synthesizer are architecturally identical, these examples clearly illustrate that the dual-path training scheme and loss formulation adopted in GAN-Blot play a crucial role in WB image synthesis with controllable band contours and appearance style, enabling faithful synthesis of the desired target WB image.

Method	#Img.	Fooling Rate	Struc. Plaus.	Style Plaus.	Realism Rate
GAN-Blot	75	81.3% (88.2%)	76.0%	80.0%	56.0%
StyleGAN2-JS	50	74.0% (57.0%)	61.3%	56.6%	38.7%
CycleGAN [ 47 ]	50	10.0% (22.0%)	7.3%	6.0%	0%
ControlNet [ 31 ]	52	9.6% (47.5%)	5.8%	7.7%	3.8%
InstructPix2Pix [ 32 ]	50	0 (0)	0	0	0

TABLE II: Expert visual evaluation of synthesized WB images. Fooling Rate is computed under blind inspection. Structure Plausibility (Struct. Plaus.), Style Plausibility (Style Plaus.), and Realism Rate measure structural validity, stylistic validity, and overall realism, respectively. Parenthesized values indicate fooling rates from image-processing specialists without WB-domain expertise.

	(A) GAN-Blot-Synthesized						(B) Non-AI-generated
	(A-1) Manually inspected			(A-2) Randomly selected			(B-1) Lab-Cropped			(B-2) Lab-full			(B-3) publication-cropped
Method	n=360			n=1640			n=151			n=195			n=1654
	TP	FN	TPR(%)	TP	FN	TPR(%)	TN	FP	FPR(%)	TN	FP	FPR(%)	TN	FP	FPR(%)
Effort [ 16 ]	68	292	18.89	296	1344	18.05	99	52	34.44	133	62	31.79	1182	472	28.54
RINE [ 14 ]	41	319	11.39	505	1135	30.79	151	0	0	175	20	10.26	1477	177	10.70
UnivFD [ 13 ]	3	357	0.83	415	1225	25.30	141	10	6.62	140	55	28.21	1289	365	22.07

TABLE III: Performance of existing AI-generated image detectors on GAN-Blot synthetic WB images and non-AI-generated WB images.

Fig. 8: Response distributions of Manjunath et al.’s method [ 24 ] . Histograms show the final softmax score distributions for (A) Manjunath et al.’s official real-image subset, (B) Manjunath et al.’s official fake-image subset, (C) the randomly selected GAN-Blot subset, and (D) the cropped authentic laboratory WB subset. The detector separates the official real and fake subsets, but does not reliably respond to WB images synthesized by GAN-Blot and lacks a stable response to authentic laboratory images.

Fig. 9: Balanced accuracy of Cardenuto et al.’s method [ 25 ] under official configurations. Heat maps show the balanced accuracy (%) of Cardenuto et al.’s 3 $\times$ 5 $\times$ 8=120 official configurations on GAN-Blot images, using (A) Isolation Forest, (B) PCA, and (C) One-Class SVM. Most configurations remain close to 50%, indicating near-chance detection performance.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 469 papers

Apr 30, 2026

AEGIS: A Holistic Benchmark for Evaluating Forensic Analysis of AI-Generated Academic Images

We introduce AEGIS, A holistic benchmark for Evaluating forensic analysis of AI-Generated academic ImageS. Compared to existing benchmarks, AEGIS features three key advances: (1) Domain-Specific Complexity: covering seven academic categories with 39 fine-grained subtypes, exposing intrinsic forensic difficulty, where even GPT-5.1 reaches 48.80% overall performance and expert models achieve only limited localization accuracy (IoU 30.09%); (2) Diverse Forgery Simulations: modeling four prevalent academic forgery strategies across 25 generative models, with 11 yielding average forensic accuracy below 50%, showing that forensics lag behind generative advances; and (3) Multi-Dimensional Forensic Evaluation: jointly assessing detection, reasoning, and localization, revealing complementary strengths between model families, with multimodal large language models (MLLMs) at 84.74% accuracy in textual artifact recognition and expert detectors peaking at 79.54% accuracy in binary authenticity detection. By evaluating 25 leading MLLMs, nine expert models, and one unified multimodal understanding and generation model, AEGIS serves as a diagnostic testbed exposing fundamental limitations in academic image forensics.

Bo Zhang, Tzu-Yen Ma, Zichen Tang +18

Beijing University of Posts and Telecommunications

Image Forgery Detection Digital Image Forensics Image Forgery Localization AI-Generated Image Detection

Ask PDF

Sep 5, 2026

FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

AI-generated image detection is increasingly open-world: new image generators produce highly realistic images that make visual artifacts harder to identify. Existing detectors usually rely on a fixed set of forensic cues, so a detector that works well for one generator family may fail on another. We introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), which learns an image-conditioned tool-use policy for forensic analysis. Instead of applying a fixed detector, FACT decides which forensic tools to call, interprets the returned evidence, and stops when sufficient evidence has been collected. FACT follows an Evolve--Distill--Refine pipeline: it evolves an execution-verified forensic skill, compiles the skill into action--observation tool-use trajectories, distills them into a compact agent, and refines the policy with cost-aware GRPO. Across two internal and four public benchmarks, FACT achieves the best performance among all compared methods, including on recent unseen generators, deepfakes, and manipulated images.

Jiaoyang Chen, Bin Hu, Jingyu Hu +3

Lingnan University, Hong Kong · Harbin Institute of Technology, Shenzhen · The University of Hong Kong +2

AI-Generated Image Detection Image Forgery Detection Digital Image Forensics Deepfake Detection

Ask PDF

Apr 9, 2026

SciFigDetect: A Benchmark for AI-Generated Scientific Figure Detection

Modern multimodal generators can now produce scientific figures at near-publishable quality, creating a new challenge for visual forensics and research integrity. Unlike conventional AI-generated natural images, scientific figures are structured, text-dense, and tightly aligned with scholarly semantics, making them a distinct and difficult detection target. However, existing AI-generated image detection benchmarks and methods are almost entirely developed for open-domain imagery, leaving this setting largely unexplored. We present the first benchmark for AI-generated scientific figure detection. To construct it, we develop an agent-based data pipeline that retrieves licensed source papers, performs multimodal understanding of paper text and figures, builds structured prompts, synthesizes candidate figures, and filters them through a review-driven refinement loop. The resulting benchmark covers multiple figure categories, multiple generation sources and aligned real--synthetic pairs. We benchmark representative detectors under zero-shot, cross-generator, and degraded-image settings. Results show that current methods fail dramatically in zero-shot transfer, exhibit strong generator-specific overfitting, and remain fragile under common post-processing corruptions. These findings reveal a substantial gap between existing AIGI detection capabilities and the emerging distribution of high-quality scientific figures. We hope this benchmark can serve as a foundation for future research on robust and generalizable scientific-figure forensics. The dataset is available at https://github.com/Joyce-yoyo/SciFigDetect.

You Hu, Chenzhuo Zhao, Changfa Mo +2

Zhejiang University, Hangzhou, China · Independent Researcher, Beijing, China · University of Oulu, Oulu, FINLAND

AI-Generated Image Detection Digital Image Forensics Synthetic Benchmark Generation

Ask PDF

Abstract

Image Generation

Explore similar work

AEGIS: A Holistic Benchmark for Evaluating Forensic Analysis of AI-Generated Academic Images

FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

SciFigDetect: A Benchmark for AI-Generated Scientific Figure Detection