How to detect Stable Diffusion images
Stable Diffusion is the most widely deployed open-source image generator, running on millions of local machines and hundreds of hosted services. Its images can look wildly different depending on the model checkpoint, LoRA weights, and post-processing applied — but certain patterns remain consistent.
Why Stable Diffusion is harder to detect
Unlike Midjourney (cloud-only) or DALL-E (C2PA-signed), Stable Diffusion images:
- Are generated locally, so there is no central server to add metadata.
- Can be run through any of hundreds of community checkpoints (Realistic Vision, DreamShaper, SDXL, Flux, etc.), each with a distinct visual style.
- Are frequently post-processed with upscalers and face-fix tools (CodeFormer, GFPGAN), which alter the pixel-level fingerprint.
- May be generated by third-party hosted services (Dream Studio, Civitai, Clipdrop) that strip or never add metadata.
Metadata traces
When images are exported directly from AUTOMATIC1111, ComfyUI or similar interfaces, the PNG metadata block often contains:
- The full prompt and negative prompt
- The sampler, CFG scale, seed, and model hash
- The checkpoint filename and any LoRA weights
Upload the original PNG to AiPicDetect and check the metadata panel. A chunk with sampler parameters ("Steps:", "CFG scale:", "Seed:") is conclusive evidence of Stable Diffusion generation.
JPEG exports and hosted-service outputs typically strip this. If the PNG metadata is absent, fall back to the detector score and visual tells.
Visual tells by model family
There is no single "Stable Diffusion look", but these patterns appear commonly:
- SD 1.5 / 2.x checkpoints: softer, slightly smeared textures; characteristic skin smoothing; hands and teeth often distorted.
- SDXL: sharper detail, more photorealistic, harder to distinguish from photos. Face quality significantly improved but complex scenes still show distortions.
- Flux (SD3-era): extremely photorealistic, better prompt adherence, very hard to spot by eye. Metadata check is essential.
- Anime/illustrative checkpoints: consistent flat cel-shading; uniform skin; eyes with a characteristic highlight pattern.
Quick reference
| Signal | Weight |
|---|---|
| PNG metadata with sampler/seed/CFG present | Definitive |
| High AiPicDetect score on original file | Strong |
| Distorted hands/text (SD 1.x–2.x) | Supporting |
| Uniform skin + characteristic eye highlights (anime) | Supporting |
| JPEG or hosted output, no metadata | Detector score only |
Check a Stable Diffusion image →
Обновлено 2026-09-13 · исходный код на GitHub