Lexica

Lexica is an AI image search and generation company founded by Sharif Shameem. It began as a way to browse Stable Diffusion images and their prompts, then added generation directly within the site and developed its own text-to-image model, Lexica Aperture. User behavior shaped both transitions: people copied prompts into separate generation tools, and their preference for art helped redirect Aperture’s training beyond photorealism.[2][1]

Lexica’s search origins preceded its model development. While working on Debuild in 2022, Shameem spent hours in the Stable Diffusion Discord, where people shared generated images and elaborate prompts. Discord’s search interface showed only a few images at a time. In his May 2023 interview, he recalled scraping approximately ten million images, building the website in about two days, and launching roughly 24 hours after Stable Diffusion’s weights were open-sourced. The gallery made examples accessible without requiring visitors to run the model locally.[1]

Shameem initially treated Lexica as a side project and set a million-user threshold for considering it full time. He said it reached that threshold in roughly its first week and a half, prompting the company to pivot from Debuild to Lexica and subsequently raise a seed round. In the interview published May 8, 2023, he reported raising about $5 million from Daniel Gross and also participating in AI Grant.[1]

Lexica’s original search used PostgreSQL full-text search to match words in prompts. Inspired by the visual search engine Same Energy, Shameem added an index of CLIP image embeddings, allowing retrieval by image content rather than prompt wording alone. The implementation used FAISS for nearest-neighbor search. In his 2023 account, GPUs computed the embeddings while CPUs served retrieval; he said an eight-core server could handle it. The same approach enabled visitors to upload an image and find similar images.[1]

The Lexica Search API documents GET requests with a q parameter containing either a search term or an image URL for reverse image search. Its documented response contains 50 image results, with metadata including prompts, image URLs, dimensions, seeds, models and guidance values.[3]

Shameem watched users browse Lexica, copy a prompt, switch to a separate Stable Diffusion interface, and paste or slightly revise it. Experiencing the same friction himself led the team to integrate generation into Lexica. The feature had launched by his May 2023 interview.[1]

Lexica used website preference signals to help select training data. Shameem described comparing images through pairwise rankings and Elo-style scores, alongside trained aesthetic models. For example, choosing one image over another displayed beside it could supply a preference signal. Those rankings helped train aesthetic models and filter datasets for diffusion-model training.[1]

Aperture v1 emphasized photorealism, but Shameem said some users preferred Stable Diffusion 1.5 for art and unusual abstract images. For v2, the team used art images from Lexica to train aesthetic rankers, then used those rankers to filter larger training sets. He described v3 as an improvement using much more data. Version numbers generally represented different base models; fine-tunes could be released without incrementing the version.[1]

In the May 2023 interview, Shameem reported over five billion images served per month and tens of millions of people visiting Lexica each month. These were historical founder-reported figures: images served measure image delivery, rather than searches or newly generated images.[1]

Dylan Molina lists his tenure as a Lexica ML engineer from April 2024 to June 2025 and says he joined to help train a video model. He describes building internal labeling tools, writing distributed jobs across an H100 cluster, and running data collection, LLM filtering, CLIP jobs and Elo tournaments. Molina credits Shameem as founder and mentor. His account describes training work, without identifying a public video-product release.[6]

On September 18, 2026, Lexica’s account page displayed annual-billing prices of $8 per month for Starter, $24 for Pro and $48 for Max, with respective monthly allowances of 1,000, 3,000 and 7,000 fast generations. Each prompt submission consumed four image-generation credits. Pro and Max included slow generations subject to a daily cap after the fast allowance was exhausted. Generation required a subscription.[4]

Starter and Pro images appeared in Lexica search; Max images were private unless shared. The terms tied commercial permissions for Aperture images to paid plans and team size, while limiting direct resale on stock-photo sites or other marketplaces to the original creator. They also specified an important model exception: Stable Diffusion 1.5 images were CC0-licensed and could be used commercially without a paid plan.[4][5]