AI Image to Prompt — reverse prompt generator

Most "image to prompt" sites push your picture to a private API and hand back a black-box string. This page is transparent about which parts are exact maths and which part is a learned model, and it keeps the heavy lifting in your browser. Four layers combine into one prompt.

1. Aspect ratio — exact integer maths

The tool reads your image's real pixel dimensions and divides both by their greatest common divisor using Euclid's algorithm. A 1920×1080 image reduces to 16:9. When the reduced ratio is unwieldy — either side larger than 32, as with an odd 1000×1237 crop — it snaps to the nearest standard ratio (1:1, 4:3, 3:2, 16:9, 9:16, 3:4, 2:3, 21:9) by the smallest difference in the width-to-height decimal. That value becomes the Midjourney --ar parameter. This step is deterministic and never leaves your browser.

2. Colour palette — client-side quantisation

The image is drawn to a small canvas (longest side ≤ 100 px), and each pixel is quantised into a fixed 216-bucket colour grid (six levels per RGB channel). The buckets are ranked by how many pixels they hold, and each dominant bucket is mapped to the nearest named colour from a curated vocabulary by squared Euclidean distance in sRGB. The share-weighted warm-versus-cool balance decides whether the palette reads warm or cool. These colour terms feed straight into the prompt — also entirely offline.

3. Subject caption — BLIP on the Hugging Face Inference API

The one learned step. The image bytes are POSTed once to a server-only route that forwards them to Salesforce/blip-image-captioning-base on the free Hugging Face Inference API (Li et al., ICML 2022). The returned caption becomes the editable subject. Running BLIP in the browser would mean downloading a 150–250 MB model to every phone, so it runs server-side; if the service is not configured, you simply type the subject yourself and the rest of the tool is unaffected.

4. Style vocabulary and assembly

The medium, style, lighting, composition, quality, and artist chips follow the clip-interrogator vocabulary. The selected phrases are joined in a fixed, readable order — subject, medium, style, lighting, colour, composition, quality, artist — then formatted per target: Midjourney appends the --ar / --stylize / --v parameters, SDXL adds a default negative-prompt block, and DALL·E 3 receives a single descriptive sentence because it prefers prose. The assembly is deterministic: the same image and the same selections always produce the same prompt.

The aspect-ratio maths is exact and reproducible. Each example below is cross-checked live against the same detectAspectRatio() function the tool uses.

• Hugging Face model card — Salesforce/blip-image-captioning-base (subject caption)
• Hugging Face model card — openai/clip-vit-base-patch32 (interrogation vocabulary)
• pharmapsychotic/clip-interrogator — the medium / movement / artist vocabulary method (MIT)
• Midjourney — parameter reference (--ar, --stylize, --v)
• Hugging Face Inference API — documentation
• Euclidean algorithm — the GCD reduction behind the aspect-ratio detection

Model cards, the clip-interrogator vocabulary approach, and the Midjourney parameter syntax were last cross-checked on 2026-07-11. The aspect-ratio and palette maths are exact and versioned with this page; the subject caption tracks whatever BLIP build Hugging Face serves.