Inkrata

Do you store the images I upload?

No. We do not store any images. When you use the upload endpoint, the image is processed in memory by our hashing, OCR, and VLM models, and then immediately discarded. Our database only contains cryptographic hashes (SHA-256, PDQ) and vectorized identifiers, none of which can be used to reconstruct the original image. If you are still uncomfortable, you can use the /check/pdq endpoint instead, which requires no image upload at all.

Is Inkrata actually free? Are there rate limits?

Yes, Inkrata is completely free for everyone. We have reasonable rate limits in place to prevent abuse and keep the service stable, but standard bot and moderation usage is well within those bounds. If you are building something large-scale and are worried about limits, reach out to us on Discord and we can sort it out.

Why should I use PDQ instead of SHA-256?

SHA-256 is a cryptographic hash, meaning if even a single pixel changes or the image is compressed during transit (which happens automatically on Discord, Telegram, etc.), the resulting hash will be completely different. PDQ is a perceptual hash, it compares the visual similarity of images rather than their raw bytes. This makes it highly resilient to compression, resizing, and minor edits.

What is the difference between uploading an image and sending a hash?

Sending a hash via /check/pdq or /check/sha256 is fast and lightweight, but it can only identify images that are already in our database. Uploading an image to /upload allows our VLM and OCR models to actually analyze the content of the image. This means we can detect new, previously unseen scam images based on the text or visual patterns within them. Uploading also helps us automatically grow our hash database for future lookups.

What if Inkrata flags a legitimate image, or misses a scam?

No system is perfect. Scam images often reuse identical templates (like the fake MrBeast screenshot), which makes hash matching extremely reliable for known campaigns. For unknown images, our VLM and OCR models can catch obvious scam text, but they may occasionally produce false positives or miss something cleverly disguised. Anything above score: 0.30 is very likely to be scam. If you encounter false positives, please report it in our Discord server so we can tune the models and update the database.