How to tell if an image is Stable Diffusion 3.5
Stable Diffusion 3.5 is Stability AI's open-weight image model family, released in October 2024 in three variants: Large (8.1 billion parameters), Large Turbo (a distilled version that generates in four steps), and Medium (2.5 billion parameters). Because the weights are public, most SD 3.5 images are produced locally or through third-party services, so what ends up inside the file is decided by the software that wrote it, not by the model. That is also why SD 3.5 is one of the harder generators to identify from pixels alone, and our own benchmark reflects that: we recall about 62% of SD 3.5 images, against roughly 85% on Midjourney, DALL-E, SDXL and Flux.
Visual tells
Garbled lettering is a weak tell here. The SD3 architecture was designed for text: the paper behind it reports improved typography, and the model card lists typography as an improvement over earlier Stable Diffusion. Short words, signage and labels frequently render correctly, so clean text does not rule out SD 3.5. Check long strings, repeated words, and lettering that wraps around curved or angled surfaces instead.
Several similar subjects get blended or dropped. SD 3.5 uses the MMDiT (Multimodal Diffusion Transformer) architecture, and published research on MMDiT models documents subject neglect and subject mixing when a prompt names multiple subjects of similar semantics or appearance (arXiv:2411.18301). In an image this shows up as two similar animals or people where one has borrowed features from the other, or where one named subject is simply missing.
The blotchy decoder look of older Stable Diffusion is largely absent. SD 3.5 decodes from a 16-channel latent at one-eighth resolution (the reference implementation allocates a latent of shape [batch, 16, height/8, width/8]), and the SD3 paper selected 16 channels because higher channel counts significantly improve reconstruction. Against the 4-channel latents of SD 1.5 and SDXL, small faces, eyes and fine texture come back much cleaner, so the mushy-patch artefact people learned to spot on older Stable Diffusion images is unreliable on this one.
Anatomy errors, when they occur, cluster in the same places. SD 3 Medium was widely reported in June 2024 for mangled limbs on prompts involving people lying down, and SD 3.5 followed that October. Hands, feet, joints, and figures that are reclining, foreshortened or partly occluded remain the highest-yield areas to inspect on any image from this family.
Dimensions tend to sit on the 1-megapixel grid. Stability describes SD 3.5 Large as intended for 1 megapixel output, the official reference script defaults to 1024x1024, and SD 3.5 Medium covers 0.25 to 2 megapixels. An image that is exactly 1024x1024, or another crop at about 1 megapixel, with no camera EXIF, proves nothing on its own, but it is consistent with SD 3.5 and inconsistent with an unedited phone photo.
Metadata & provenance
SD 3.5 writes no provenance data of any kind. The official reference implementation (github.com/Stability-AI/sd3.5) saves a bare PNG with a sequential filename and calls no metadata or watermarking code: no EXIF, no XMP, no PNG text chunk. The Hugging Face diffusers pipeline for SD3 contains no watermarking code either, which is a deliberate change from SDXL, whose diffusers pipeline switches on an invisible DWT-DCT watermark by default when the invisible-watermark package is installed. The invisible-watermark check that sometimes works on SD 1.5 and SDXL images therefore does not apply to SD 3.5. There is no C2PA Content Credentials support in the released model or reference code, and neither the model card nor the reference repository mentions C2PA, Content Credentials or provenance metadata. If an SD 3.5 image does carry a C2PA manifest, it was attached by the hosting service or editor that wrote the file, so read the signer to see who vouched for it. SynthID is not involved: it is Google DeepMind's watermark, applied to Google's own models, and nothing in SD 3.5 references it. Anything you do find in an SD 3.5 file was written by the interface rather than the model, for example a "parameters" PNG text chunk from AUTOMATIC1111 or "prompt" and "workflow" JSON from ComfyUI, and all of it disappears on a re-save, a screenshot, or an upload to any platform that strips metadata. Absence of metadata is not evidence that an image is real.
How well WeDetect handles Stable Diffusion 3.5
On our internal held-out benchmark, wedetect recalls about 62% of Stable Diffusion 3.5 images. That is well below the roughly 85% recall we measure on Midjourney, DALL-E, SDXL and Flux, and below our approximately 89% overall figure on modern generators. In plain terms, close to four in ten SD 3.5 images currently come back as likely real. The reasons are structural rather than a tuning problem: the 16-channel latent removes much of the decoder-level noise that detectors trained on older Stable Diffusion relied on, and there is no watermark and no metadata to fall back on. Research on detecting SD 3.5 specifically reports that the usable signal sits in frequency-domain statistics rather than anything a person can see (arXiv:2507.02995). Practically, treat a "likely AI" verdict on a suspected SD 3.5 image as meaningful, and treat a "likely real" verdict as weak evidence rather than a clearance. Every result is probabilistic, and anything consequential should be corroborated by other means.
Check an image now
Drop the image in below — free, no sign-up, deleted after analysis.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
Frequently asked questions
Does Stable Diffusion 3.5 watermark its images?
No. The released SD 3.5 code contains no watermarking, and the diffusers pipeline for SD3 has none either. This is a change from SDXL, whose pipeline applied an invisible DWT-DCT watermark by default when the invisible-watermark library was installed. Because SD 3.5 never adds a mark, finding no watermark in an image tells you nothing about whether it is real.
How do I check if a photo was made with Stable Diffusion 3.5?
Start with the file itself. Local interfaces often leave generation settings behind, such as a "parameters" text chunk from AUTOMATIC1111 or "prompt" and "workflow" JSON from ComfyUI, and those usually name the checkpoint. Our metadata checker reads them, along with any C2PA manifest and camera EXIF. If the file has been screenshotted, re-saved or uploaded to a platform that strips metadata, none of that will survive, and you are left with the pixels: run the image through the detector, and inspect hands, reclining or occluded figures, long text strings, and any scene containing two or more similar subjects.
Why did the detector say my AI image is real?
SD 3.5 is one of our weakest cases. Our measured recall on it is about 62%, so roughly four in ten SD 3.5 images are missed. The model is built on a 16-channel latent that reconstructs fine detail cleanly, which erases much of the low-level residue detectors key on, and it leaves no watermark or metadata as a backup signal. A "likely real" verdict on an image you already suspect is generated should not settle the question.
Can you tell which model made an AI image, SD 3.5 versus Flux or SDXL?
Not from the pixels, and we do not claim to. A detector estimates whether an image is generated, not which system generated it. The only reliable way to attribute an image to SD 3.5 is leftover generation metadata naming the checkpoint, or a signed C2PA manifest from the service that produced it, and both are trivially removed. The visual and resolution patterns on this page are consistent with SD 3.5 but are not proof of authorship.
Other generators: Midjourney · DALL·E 3 · GPT-4o · Stable Diffusion · Flux · Grok · Firefly · Gemini · SDXL · Nano Banana · GPT Image 2 · Flux 2 · Sora · all generators.