Detect Stable Diffusion Images
A Stable Diffusion detector identifies images generated by the Stable Diffusion model family - including SD 1.5, SDXL, SD3, and thousands of community fine-tunes - by analyzing the characteristic VAE decoder artifacts, latent-space frequency signatures, and denoising grid patterns that latent diffusion models embed in every generated image. FauxLens detects official Stable Diffusion releases and community variants including DreamBooth, LoRA, and ControlNet outputs.
SCAN IMAGE NOW - FREEHow every scan runs
Read the methodProvenance
C2PA content credentials and EXIF metadata, checked first
Visual Analysis
Anatomy, lighting and composition examined for inconsistencies
Evidence Chain
Every finding itemised and graded, not reduced to a yes or no
- Files deleted after analysis
- No account required
How to detect a Stable Diffusion image
Stable Diffusion does its work in a compressed latent space and only decodes to pixels at the very end, through a Variational Autoencoder. That last step is where it gives itself away. This two-stage process leaves two distinct sets of forensic artifacts. The first set comes from the diffusion process itself: the denoising schedule produces characteristic noise patterns in the frequency domain that differ from both camera noise and other generator architectures. The second set comes from the VAE decoder: the 8x spatial upscaling from latent to pixel space introduces a subtle but consistent grid-like artifact pattern at the pixel level, particularly visible in high-frequency detail regions. FauxLens analyzes both artifact classes simultaneously. Additionally, the choice of sampler affects artifact patterns: images generated with DDIM have different spectral characteristics than those generated with DPM++ or Euler samplers. Our detection engine is trained on outputs from all major samplers across AUTOMATIC1111, ComfyUI, and InvokeAI frontends, and maintains high accuracy across sampler variations. Detection accuracy for clean txt2img outputs without post-processing exceeds 94% for SD 1.5, 91% for SDXL, and 89% for SD3.
Base models, fine-tunes and LoRAs
The forensic challenge of Stable Diffusion detection is not the base model. It is the vast ecosystem of fine-tuned models built on top of it. Platforms like Civitai and HuggingFace host over 100,000 community-fine-tuned models. DreamBooth fine-tunes teach the base model to generate a specific person, object, or style. LoRA (Low-Rank Adaptation) adapters add style characteristics without replacing base model weights. Checkpoint merges combine multiple models to produce hybrid capabilities. Each fine-tuned model inherits the base model's latent diffusion architecture and VAE decoder, which means it retains the core forensic fingerprint, while adding its own style-specific artifacts on top. Photorealistic fine-tunes like epiCRealism, Deliberate, and RealisticVision are designed to produce outputs closer to real photographs, which makes them the hardest SD variants to detect. Anime fine-tunes like Anything V5 and AbyssOrangeMix produce stylistically distinctive outputs that are visually obvious but still carry the latent diffusion fingerprint. FauxLens detects both categories by targeting the underlying VAE decoder artifacts and frequency-domain signatures that all SD fine-tunes share, regardless of their visual style.
SDXL, SD3, and Stability AI's Newer Models
Stability AI's model lineup has expanded significantly since the original SD 1.x releases, and each generation has distinct forensic characteristics. SD 1.5 (2022) remains the most widely deployed base model for fine-tuning. It uses a relatively small latent space (4 channels, 8x spatial compression) and shows the strongest VAE decoder artifacts - detection accuracy approaches 96% for unmodified outputs. SDXL (2023) introduced a larger latent space and two-stage generation pipeline with a base model and refiner model. The increased model capacity produces higher-quality outputs with subtler artifacts, reducing detection accuracy to approximately 91%. The refiner stage introduces its own characteristic artifact pattern at detail boundaries that our frequency analysis detects. SD3 (2024) adopted a Multimodal Diffusion Transformer (MMDiT) architecture - a fundamental shift away from the U-Net backbone used in SD 1.x and SDXL. The transformer architecture produces different artifact patterns than U-Net-based generation, particularly in how attention heads produce texture at fine scales. FauxLens maintains specific detection models for each architecture variant rather than relying on a single cross-version classifier, which is why accuracy remains high across the Stability AI model family.
Why this is the most misused generator
Stable Diffusion has a fundamentally different threat profile than Midjourney or DALL-E. It is free, open-source, and runs entirely locally without any API, moderation layer, or usage terms enforcement. This makes it the primary tool for AI image generation that bypasses all platform controls. The consequences are significant. Stable Diffusion has been used to generate illegal content targeting minors by removing the safety filters present in the official release - a use case that has led to documented prosecutions in multiple jurisdictions. It is used to generate disinformation imagery without the platform attribution that Midjourney's Discord bot leaves. It is the primary tool for non-consensual intimate imagery (NCII) creation, where a real person's face is composited into synthetic content. In fraud, SD is used for synthetic identity document photos, fake product imagery, and romance scam profiles where the user wants images that return no results in reverse image search. The detection arms race is real: as FauxLens and other tools improve detection of SD outputs, the community develops post-processing pipelines designed to reduce forensic signal - adding grain, applying slight color grading, or passing outputs through additional AI upscalers. FauxLens continuously updates its detection models to maintain accuracy against these evasion techniques.
Getting a reliable answer
Stable Diffusion detection accuracy is most affected by post-processing applied after generation. Original PNG files exported directly from AUTOMATIC1111 or ComfyUI with no post-processing yield the highest detection confidence. PNG is the preferred format because it uses lossless compression that does not introduce additional JPEG artifacts on top of the generation artifacts. JPEG exports at quality 90 or above retain strong forensic signals. JPEG exports below quality 85 begin to degrade the VAE decoder artifacts significantly. Images that have been run through AI upscalers (Topaz Gigapixel, Real-ESRGAN, the AUTOMATIC1111 Hires Fix upscaler) show reduced confidence because the upscaler introduces its own artifact patterns that partially mask the original SD fingerprint. Images with heavy post-processing (film grain overlays, color grading, significant sharpening), similarly reduce confidence. For images where you suspect SD involvement but receive an inconclusive result, examine whether the image has distinctive smooth skin texture without the subtle skin pore detail that real photographs contain, and check whether background elements have the characteristic over-smooth foreground-background separation that SD photorealistic models produce. If the image is a suspected inpainted composite, the inpainted region will show different ELA characteristics than the surrounding area even when the full-image confidence is moderate.
Ready to verify an image?
Try It FreeFrequently Asked Questions
It can. Our engine detects the underlying latent diffusion process artifacts, VAE decoder grid patterns and frequency-domain signatures, shared by all SD variants, including custom fine-tuned models, LoRAs, and community checkpoints. Detection accuracy is highest for unmodified txt2img outputs and remains strong even with post-processing.
FauxLens detects AI involvement across the generation spectrum. Pure txt2img outputs are detected with highest confidence. Img2img outputs show mixed signals - the AI-modified regions carry SD artifacts while the original real photo regions show authentic camera noise. Inpainted images are flagged with confidence levels reflecting the degree of AI modification and the proportion of the image that was AI-generated.
ControlNet is a conditioning mechanism that guides Stable Diffusion generation using reference images (depth maps, edge maps, pose skeletons). The output is still generated through the full SD latent diffusion pipeline and inherits the same VAE decoder artifacts and frequency signatures. ControlNet does not mask or remove the generation fingerprint.
Slightly. Different frontends apply different default settings, samplers, and optional post-processing steps. AUTOMATIC1111 and ComfyUI both use the same base SD models, so the core forensic fingerprint is the same. But AUTOMATIC1111 applies a face restoration pass by default for portrait images (GFPGAN/CodeFormer) which can introduce additional artifacts. ComfyUI typically outputs more raw images. FauxLens is trained on outputs from both frontends.
Yes. Anime fine-tunes (Anything V5, Counterfeit, AbyssOrangeMix) use the SD latent diffusion backbone and VAE decoder, so they carry the same core forensic fingerprint. Their visual style is highly distinctive, but the mathematical detection relies on pixel-level statistics rather than visual appearance. Detection accuracy for anime fine-tunes is comparable to the base SD 1.5 detection rate.
SD-based face swap techniques, particularly those using inpainting to replace faces in real photos, are detected with moderate-to-high confidence. The face region shows SD generation artifacts while the surrounding body and background show real camera noise. FauxLens flags the image as AI-involved and the per-layer analysis shows which forensic signals fired in which regions.
It does. There is no forensic difference between images generated by the SD model running locally versus via an API. The mathematical artifacts are embedded by the model itself, not by the serving infrastructure. Running SD locally without network connectivity does not affect the forensic signature in any way.
SD3 uses a Multimodal Diffusion Transformer (MMDiT) architecture rather than the U-Net backbone used in SD 1.x and SDXL. This architectural difference produces different artifact patterns - particularly in how attention heads generate fine-scale texture. FauxLens maintains separate detection models for each SD architecture rather than a single cross-version classifier.
FauxLens detects AI-generated images regardless of their content. The forensic analysis examines pixel-level mathematical patterns, not the semantic content of the image. Detection accuracy is the same for NSFW and SFW SD outputs - the generation artifacts are content-independent.
The web tool processes one image per submission. For batch processing of multiple images, the FauxLens API supports multiple images per request with optimized throughput. Contact support@fauxlens.com for API access details.
Learn More
How to Detect Stable Diffusion Images: A Forensic Deep Dive
Stable Diffusion is open-source, infinitely customizable, and runs locally. That makes it the hardest AI model to detect. Here is how forensic analysis identifies SDXL and its fine-tuned variants.
The Science of Deception: How AI Detection Works
A detailed guide to the digital forensics behind Faux Lens. From Error Level Analysis (ELA) and JPEG compression artifacts to Photo Response Non-Uniformity (PRNU).
Every AI Image Has a Hidden Fingerprint - Here's How Forensics Finds It
Every AI model leaves an invisible fingerprint in the images it generates. GAN fingerprints are the forensic equivalent of ballistic markings-unique to the weapon, invisible to the eye, detectable by science.
Can You Tell the Difference? Midjourney vs DALL-E vs Real Photos [Forensic Test]
A technical showdown between the major AI engines. We analyze the specific 'digital fingerprints' left by Midjourney v6, DALL-E 3, and Stable Diffusion.
Detect while browsing - try the Chrome Extension
Right-click any image · 4 free detections · No account required