Lexica
Lexica is an AI image search engine and generation platform founded in 2022 by Sharif Shameem, Dylan Molina, and Subhash Ramesh.1 The company launched 24 hours after the open-source release of Stable Diffusion's model weights as a searchable database of AI-generated images and their prompts, reaching one million users within its first ten days.3 Lexica later developed Aperture, a proprietary text-to-image model trained on high-resolution photographs, and raised a $5 million seed round led by Daniel Gross.4 The company is in the Founders, Inc. portfolio.1
Origins
Shameem built the initial version of Lexica in approximately two days while active in the Stable Diffusion Discord server.3 He scraped roughly 10 million images from the Stable Diffusion Discord along with the prompts and parameters used to generate each one, stored them in a database, and built a basic search interface.38 Shameem has said he intended the tool to make "Stable Diffusion prompting a bit less of a dark art and more of a science."7
The site launched on August 22, 2022, the day after Stability AI open-sourced the Stable Diffusion weights.2 Within 24 hours, Lexica handled 51,000 search queries and served 2.2 terabytes of bandwidth; by the second day those figures doubled to 111,000 queries and 22 million images served.3 Shameem originally built Lexica for his own use because the Discord search bar was limited to displaying five images at a time, and he found the process of discovering prompts through it frustrating.3
Before Lexica, Shameem had founded two earlier companies: Vectordash, a Y Combinator-backed GPU cloud that aggregated gaming GPUs from cryptocurrency miners for machine learning researchers, and Debuild, a code-generation tool that used OpenAI's GPT-3 to let users build web applications from natural-language descriptions.610 A viral tweet demonstrating Debuild in July 2020 received over 20,000 likes and coverage in Business Insider, Fast Company, and MIT Technology Review.310 Shameem studied computer science at the University of Maryland but did not graduate, choosing to drop out after joining Y Combinator with Vectordash.3
Search technology
The original search engine relied on PostgreSQL full-text search, matching user queries against the text content of stored prompts.3 In September 2022, Shameem replaced this with semantic search powered by OpenAI's CLIP model, indexing the CLIP image embeddings of every image in the database.9 Searches became k-nearest-neighbor lookups against the CLIP embedding index rather than keyword matches, which allowed users to search by the visual content of images rather than prompt text alone.3
The CLIP embedding index was served using FAISS, Meta's library for fast approximate nearest-neighbor search, running on CPUs.3 Embeddings were computed on GPUs, but the serving infrastructure ran on an eight-core server.3 By September 2022, the index contained over 10 million images.8
Lexica also offered reverse image search: users could upload a photograph and receive the most visually similar Stable Diffusion images along with their prompts.7 The platform exposed a public search API at lexica.art/api/v1/search, returning up to 50 results per query with metadata including the prompt, seed, guidance scale, dimensions, and model used for each image.12
Funding and AI Grant
On September 28, 2022, Lexica announced a $5 million seed round led by Daniel Gross.4 As part of the financing, Lexica joined AI Grant, Daniel Gross's program for AI startups.43 Shameem wrote that the company's goal was to "give everyone the ability to create whatever they imagine, and ultimately invent new mediums of expression and entertainment."4
Before the seed round, Shameem had told himself he would consider focusing on Lexica full-time if it reached one million users, expecting that milestone to take years.3 The platform passed that threshold in roughly ten days, prompting him to pivot entirely from Debuild.3 By May 2023, Lexica was serving over 5 billion image searches per month.3
Aperture
Lexica introduced Aperture, its proprietary image generation model, in December 2022.5 Shameem observed that users were tabbing back and forth between Lexica's search results and separate Stable Diffusion interfaces, copying prompts from one into the other.3 Aperture brought generation directly into the Lexica interface.3
Aperture v1 was released in December 2022, designed to produce photorealistic images from text prompts.5 The model was almost certainly a fine-tuned version of Stable Diffusion trained on high-resolution photographs; Shameem described it as "a photorealistic 3D engine you can prompt via text."5 The web interface allowed users to adjust resolution between 512 and 768 pixels, set an orientation scale between 4 and 13, enter negative prompts to exclude elements, and toggle a setting to prevent double heads in AI portraits.5 Generated images could be upscaled to a maximum of 3,072 by 2,048 pixels.5
Aperture v2 followed in late December 2022, and Aperture v3 shipped in April 2023.3 Each version in the Aperture line improved in the direction Shameem emphasized in interviews: data quality over architectural novelty, on the principle that diffusion models learn the distribution of whatever data they are trained on.3
Product and API
Each image in Lexica's database is accompanied by the prompt and parameters used to generate it, allowing users to learn from and iterate on existing creations.1 Aperture supports adjustable generation speeds, customizable resolutions, and negative prompts.1
The Lexica search API accepts GET requests with a query parameter and returns JSON arrays of up to 50 image results, each including the image ID, gallery URL, source URL, prompt, dimensions, seed, guidance scale, model name, and an NSFW flag.12 Reverse image search is available through the same endpoint by passing an image URL as the query parameter.127
Founders, Inc. and later developments
Lexica is listed in the Founders, Inc. portfolio, classified under AI/ML and Consumer.1 The F.inc portfolio page lists the team size as four people.1
Shameem later joined the Labs team at OpenAI, announcing the move on his X profile.11 His personal website describes his trajectory as having "founded Lexica, one of the first AI image generation tools, where I trained diffusion models that served millions of people."6