AI image metadata viewer
Most AI image tools quietly embed the exact recipe used to create a picture inside the file itself. Stable Diffusion writes the full prompt, seed, steps, CFG, and sampler into a PNG text chunk; some tools use JPEG EXIF fields. That means you can often recover the precise parameters of any AI image you saved — to reproduce it, remix it, or learn from it. This viewer reads those fields locally, so even private images stay on your device.
What the parameters mean
When the viewer extracts generation metadata, here is what each field tells you:
Prompt — the full positive text instruction sent to the model. This is the most valuable thing to recover if you want to reproduce or extend an image.
Negative prompt — the terms the model was asked to avoid. Common negatives
like blurry, lowres, extra limbs, bad anatomy suppress typical artifacts;
custom negatives refine the style further.
Model / checkpoint — the base model used. The same prompt can produce
dramatically different results across models, so this is essential context for
reproduction. Names like v1-5-pruned-emaonly.ckpt or dreamshaper_8.safetensors
identify the exact checkpoint.
Seed — the random initialisation number. Running the same seed with the same other parameters in the same model will produce a nearly identical image. Changing seed by even one digit will produce a completely different composition.
Steps — how many denoising iterations were run. Typical values are 20–50. More steps generally improve quality up to a point, then plateau.
CFG scale — how strongly the model adheres to the prompt. Values around 7 are a common default; higher values push the model to follow the prompt more strictly but can introduce artifacts; lower values give the model more creative latitude.
Sampler — the algorithm used for the denoising process. Common samplers include Euler a, DPM++ 2M Karras, and DDIM. Different samplers produce visibly different textures and convergence speeds.
How it works
When you select a file, the browser reads its bytes with FileReader. For PNGs
the tool walks the chunk structure looking for tEXt and iTXt chunks — the
one keyed parameters holds the Stable Diffusion string. For JPEGs it scans the
EXIF block for UserComment and ImageDescription. The raw string is then
parsed into labeled fields: everything before “Negative prompt:” is the
prompt, and the trailing key/value list yields steps, sampler, CFG
scale, seed, and model. Nothing is uploaded — all parsing happens in
your browser.
Tips and limitations
- Use originals. Re-encoding an image (screenshots, social uploads, format conversion) usually strips the metadata chunk — keep the file your tool produced.
- Midjourney embeds little. Unlike SD, Midjourney does not write full parameters into the file, so expect sparse or empty results.
- Seed is the key to reproduction. With the seed plus the other parameters you can regenerate a near-identical image in the same tool.
- No metadata is normal. Many tools and exports simply do not embed parameters; an empty result does not mean the image is not AI-generated.