Img2prompt

Img2prompt is a Replicate-hosted AI model, built by Methexis Inc, that works in reverse of a typical text-to-image tool: it takes an image as input and outputs a text prompt describing it, formatted for use with Stable Diffusion. Internally it combines OpenAI's CLIP model with BLIP image captioning, testing the input image against a large set of known artists, mediums, and visual styles to infer which words would most likely have produced a similar-looking result. The resulting prompt typically names a likely subject, artistic medium, and style descriptors drawn from that comparison. The model runs on Nvidia T4 GPU hardware through Replicate's hosted playground and API, with a typical run taking a few minutes and costing a small, usage-based fee per generation. It is open source, based on the CLIP Interrogator project by pharmapsychotic, and can also be run locally through Docker for users who prefer not to rely on the hosted API. Because img2prompt is a single-purpose research model rather than a polished consumer app, there is no dedicated website, pricing plan, or support beyond what Replicate provides; it works best as a starting point for reverse-engineering an image's style rather than producing an exact prompt match, since CLIP-based interrogation is inherently approximate.