Inspect Stable Diffusion & SDXL
SD 1.5, SDXL, and ComfyUI workflows frequently leave generation parameters, prompt chunks, or latent VAE decoding signatures. Verify any diffusion image instantly.
- Common Models
- SD 1.5, SD 2.1, SDXL 1.0, Pony Diffusion
- PNG Metadata Chunks
- tEXt / iTXt containing prompts, seeds, samplers
- VAE Factor
- 8× Latent Downsampling Compression
- Native Training Grids
- 512×512 (SD 1.5), 1024×1024 (SDXL)
Drop an image here
or choose a file · paste with ⌘V / Ctrl V
The Open-Source Diffusion Landscape: SD 1.5, SDXL & ComfyUI
Unlike closed-source models, Stable Diffusion runs locally on personal GPUs through interfaces like Automatic1111 WebUI, ComfyUI, and Fooocus. Creators use custom checkpoints, LoRAs, and upscalers.
Crucially, default web interfaces save generation metadata directly into PNG container chunks (`parameters` or `workflow`). When metadata is absent, our detector checks for VAE tiling boundaries and high-frequency latent anomalies.
Embedded PNG Parameter Sniffing
Raw PNG files from Automatic1111, ComfyUI, and Fooocus contain plaintext parameter blocks including model hash, prompt, CFG scale, and sampler. Check AI Free detects these blocks immediately in-browser.
VAE Decoding Boundary Artifacts
Because Stable Diffusion operates in a compressed latent space decoded by an 8× VAE, subtle block-boundary artifact patterns can be detected in flat color fields like skies and plain walls.
Anatomical Reduplication & Joint Glitches
In complex poses without ControlNet, SD models often produce doubled collarbones, unnatural finger joints, or limbs that vanish into clothing folds without realistic anatomical anchors.
Latent Upscaler Micro-Fractals
Images enhanced via latent upscalers often display repeating fractal noise patterns in out-of-focus background areas rather than natural photographic sensor noise.
Questions about Stable Diffusion.
Looking for technical verification guidance or have a specific question about Stable Diffusion artifacts?
Can this detector read ComfyUI workflow JSON?
Yes. If an image was saved with embedded ComfyUI or Automatic1111 parameters in its PNG header, our client-side sniffer identifies the presence of diffusion metadata instantly.
What if someone stripped the PNG metadata or converted to JPG?
If the metadata is scrubbed, our pixel-level analysis examines VAE reconstruction artifacts, anatomical consistency, and frequency domain distributions to assess AI likelihood.
Does it support anime and illustrated styles (e.g. Pony / NAI)?
Yes. While anime styles lack real-world photographic textures, latent diffusion geometry and prompt chunk signatures still reveal synthetic generation.