Mediarun Glossary
Artificial intelligence
Models, how to run them and their limits
20 terms in this section
AIA set of methods and systems performing tasks typically requiring human reasoning, learning, perception, or content generation. In HEXX AI CMS, this includes language models, document analysis, and automation.ASRTechnical term for systems that automatically recognize speech; in practice, often used interchangeably with STT. ASR analyzes audio and generates searchable or further-processable transcripts.CLIPA model trained on hundreds of millions of image-caption pairs. It places images and text in a shared numerical space, enabling it to assess how well a given description matches a given image.DIFFUSERSAn open library for running diffusion models — those that generate images by gradually removing noise. It provides a unified interface for models from different vendors.ENCODERPart of a model that converts human-readable data — such as text, images, or audio — into a sequence of numbers representing their meaning. The model does not process words directly, but works with these numerical represFLUXA family of image generation models from the team that previously created Stable Diffusion. The schnell variant is accelerated and generates an image in four steps.GGUFA model file format where weights are quantized — stored with lower precision, for example using four bits instead of sixteen. The model becomes several times smaller, with only a minor drop in quality.KIMIA language model from Chinese company Moonshot AI, known for handling very long context — capable of processing input equivalent to hundreds of pages of text at once.LLAMAA family of open-source language models from Meta. The release of model weights in 2023 triggered a wave of open models, providing a starting point previously available only to large companies.LLMAn AI model trained on large text datasets, capable of understanding and generating natural language. Performs tasks such as summarization, translation, writing, classification, and programming assistance.MLOpsA set of practices ensuring repeatable deployment, versioning, monitoring, and maintenance of machine learning models. It integrates work on data, models, infrastructure, and production performance quality.NLPA field of AI focused on analyzing, understanding, and generating human language. Includes classification, data extraction, sentiment analysis, translation, search, and language models.OCRTechnology that reads text from images, scans, and PDF documents. OCR converts visual content into searchable, copyable, classifiable, and analyzable text for AI models.QWENA family of open models from Chinese Alibaba Cloud, including language and image models. The Qwen-Image version generates graphics from text descriptions and, unlike most models, accurately renders text on images.RAGAn AI pattern in which a large language model (LLM) retrieves relevant data from an external knowledge base before generating a response. The retrieved snippets are added to the model’s context, reducing hallucinations aSAMPLING STEPSThe number of passes in which the image generator gradually removes noise from a random starting point until the final image is ready. Each step represents one full model computation.SDXLAn image generator by Stability AI, for years the foundational tool in the open-source AI graphics ecosystem. The model has approximately 3.5 billion parameters and a vast number of community-created add-ons.STTTechnology that converts recorded or live speech into text. It can recognize language, punctuation, timestamps, and sometimes separate utterances from different speakers.TRANSFORMERA neural network architecture that underpins virtually every large model today — whether language, image, or audio. Its core is the attention mechanism: the model looks at all other parts simultaneously while processing TTSTechnology that converts text into synthetic speech. TTS models can adjust language, pronunciation, pace, intonation, style, and in some cases clone a voice from a sample.