# Glossary: AI Image and Stock Photography Terms Источник / Source: https://picwin.co/en/blog/ai-image-glossary Язык / Language: en Рубрика / Topic: Technology Автор / Author: The Picwin editors Опубликовано / Published: 2026-07-08T00:00:00.000Z Обновлено / Updated: 2026-09-03T11:17:26.468Z Слов / Words: 4271 · Чтение / Reading: 22 min Обложка / Cover: https://picwin.co/api/picwin/blog-images/a4f2a6e7afeea6e3f23175377d31876b9356b81f2f34a8142387fa03a297cd5f.webp Переводы / Translations: kk: https://picwin.co/kk/blog/ai-suretter-terminderi-sozdigi, pt: https://picwin.co/pt/blog/glossario-de-imagens-ai, ru: https://picwin.co/blog/glossariy-ai-izobrazheniy, uz: https://picwin.co/uz/blog/ai-tasvirlar-atamalari-lugati Eighty-five terms covering AI images, stock licences, and how Picwin works, each with a short definition and explanation. A reference for anyone reading model descriptions, licence terms, and assignment briefs. --- **In short** This glossary collects eighty-five terms found in generative model descriptions, stock licence terms, and Picwin assignment briefs. Each entry follows the same format: a standalone one-phrase definition, then a short explanation and, where useful, a link to an article that explores the subject in detail. The vocabulary in this field comes from three different languages at once: engineering, used by model researchers; photography, used by photographers; and law, used by stock agencies. That is why generator descriptions and licence terms can seem to describe entirely different worlds. The eight sections below arrange terms from general to specific. **Sections:** - [How Generative Models Work](#how-generative-models-work) - [Prompts and Frame Control](#prompts-and-frame-control) - [Lighting, Composition, and Optics](#lighting-composition-and-optics) - [Files, Formats, and Dimensions](#files-formats-and-dimensions) - [Stock Libraries and Their Rules](#stock-libraries-and-their-rules) - [How Picwin Works](#how-picwin-works) - [Quality and Common Flaws](#quality-and-common-flaws) - [Image Rights and Provenance](#image-rights-and-provenance) Tip Terms are arranged by section rather than alphabetically, so concepts that occur together in practice also appear together here. The article’s table of contents contains the complete consecutive list of terms; on a phone, it is collapsed into the “Contents” bar above the text. ## How Generative Models Work This section explains what happens between submitting a prompt and receiving an image, along with the words model developers use to describe that process. ### Generative model A program that creates new data instead of searching for existing data. An image generator does not select a picture from a database or assemble pieces of other people’s files. It builds a new frame pixel by pixel, based on statistical patterns in its training data. This is why the same prompt produces different results. ### Diffusion model A model that creates an image by gradually removing noise from a random canvas. It is trained to solve the reverse problem: the model sees images progressively covered with noise and learns to predict what the preceding step looked like. During generation, the process begins with pure noise and runs backward until a frame emerges. Most modern generators work this way. ### Latent space A compressed internal representation of an image in which the model performs its work. Running diffusion at full resolution is expensive, so the frame is first compressed into a compact form, generated there, and converted back into pixels only at the end. This is the origin of the phrase “latent diffusion.” ### Variational autoencoder (VAE) The part of a model that converts an image into a latent representation and back again. Its quality affects fine detail and colour in the final frame: the same generation rendered through a different decoder will look slightly different in saturation and microcontrast. ### Checkpoint A file containing a model’s trained weights: a specific version of a generator. Different checkpoints based on the same architecture can produce markedly different styles: one may favour photorealism, while another favours illustration. When people say they “changed the model,” they usually mean they changed the checkpoint. ### LoRA fine-tuning A small model add-on that introduces a particular style, object, or character. A LoRA is hundreds of times smaller than a full checkpoint because it does not alter the whole model, only a thin layer on top of it. Several such add-ons can be enabled at once and blended at different strengths. ### Sampling steps The number of passes in which a model turns noise into an image. With too few steps, the frame remains unfinished and blurry; adding too many increases processing time without a visible quality gain. Every model-and-sampler combination has a point beyond which more steps serve no purpose. ### Sampler The algorithm a model follows at each denoising step. Different samplers produce images with a different character even at identical settings: some converge quickly and smoothly, while others take longer but preserve more detail. This is not a quality upgrade but a choice of trajectory. ### Seed A number that sets the initial noise for a generation. The same seed, prompt, and settings reproduce the same frame. That is why creators lock the seed when they want to change one prompt detail and see only the effect of that change. ### CFG scale A setting that determines how strictly a model follows the prompt text. A low value produces a freer but more lively image; a high value follows words literally at the cost of burnt-out colours and harsh edges. Both extremes are undesirable, and the useful range is usually somewhere in the middle. ### Training data The collection of images and captions on which a model learned. It explains both a generator’s strengths and weaknesses: things that appeared frequently in the dataset are easy to create, while rare subjects are rendered poorly and with errors. Dataset composition is also the main subject of legal disputes surrounding generative models. ### Inference A single run of a trained model using finished weights: the generation itself. Training is expensive and happens once, while inference is inexpensive and happens millions of times. When people discuss the “cost of generation,” they almost always mean inference. ## Prompts and Frame Control This section explains what an author controls during generation and lists ways to specify a frame more precisely than with a single sentence. ### Prompt The textual description of a frame from which a model builds an image. A good prompt names the subject, setting, lighting, camera angle, and style, roughly in that order of importance. The structure of a prompt is explained in the [article on becoming a Picwin author](/en/blog/how-to-become-a-picwin-author). ### Negative prompt A list of things that should not appear in the frame. It works not as a prohibition but as a shift away from the listed concepts. Typical flaws go here: extra limbs, watermarks, captions, and borders. ### Token weight A multiplier that strengthens or weakens an individual word in a prompt. It lets you say “more of this, less of that” without rewriting the entire phrase. Syntax varies by interface, but the meaning is the same everywhere. ### Prompt engineering The practice of choosing wording that produces predictable results. In essence, it applies the one-variable method: lock the seed, change one part of the prompt, and compare the frames. Without a fixed seed, there is nothing to compare because everything changes at once. ### Reference image An existing image supplied to a model as an example. A reference specifies qualities that are hard to describe in words: a particular angle, distribution of light, or line quality. A textual description is generally still needed as well. ### Image-to-image generation (img2img) A mode that starts from an existing image rather than noise. The model redraws the picture at a chosen strength: a light edit preserves the composition and changes the texture, while a strong edit retains only the original’s broad masses. ### Inpainting Replacing a selected area of an image with new content. It can remove an unwanted object, repair a hand, or change the background without touching the rest of the frame. The model completes the area’s edges itself, so the mask is drawn with some margin. ### Outpainting Extending an image beyond the boundaries of its original frame. It can turn a vertical frame into a horizontal one or add open space for text. The model continues the scene rather than stretching pixels. ### ControlNet An additional model input that defines frame geometry separately from text. A contour, depth map, or human pose is supplied as input, and the model must fit within it. This is the most reliable way to obtain a required composition without relying on wording. ### Style transfer A technique that takes content from one source and the manner of execution from another. In generative models, it is achieved through fine-tuning, a style reference, or a separate input. Legally, this is the most sensitive area: reproducing the distinctive manner of an individual living artist is best avoided. ### Upscaling Increasing the resolution of a finished image. Basic upscaling stretches pixels, while model-based upscaling invents details that were absent from the original. The latter looks better but introduces errors of its own, so the result should be inspected at 100% zoom. ### Variations Several frames generated from one prompt using different seeds. The standard selection method is to generate variations first, then refine the successful one. Picwin has no variations: one run produces one frame, and the choice is made before generation when selecting an assignment. ## Lighting, Composition, and Optics This section explains which words describe a frame so that the model understands its staging. It is the vocabulary of photographers, and generators understand it well. ### Composition The arrangement of objects within a frame and their relationships to one another. It is the first thing that makes a frame read as intentional or accidental. In a prompt, it is specified through words describing shot size, angle, and the main subject’s position. ### Rule of thirds A technique that places the main subject at one-third of the frame rather than in the centre. It gives the image movement and room for the eye to travel. It does not always apply: symmetrical staging deliberately breaks the rule and often benefits from doing so. ### Negative space The empty part of a frame surrounding the main subject. For stock imagery, this is valuable rather than vacant: a headline, logo, or price can occupy the negative space. Frames with ample breathing room are purchased more often. ### Depth of field The range of distances over which an image remains sharp. A shallow depth separates the subject from the background, while a deep one shows the whole scene. In a prompt, it can be specified through aperture or direct wording about a blurred background. ### Bokeh The visual character of blur in out-of-focus areas. The word describes not the amount of blur but its appearance: the shape of highlights and softness of edges. It is almost indispensable in product and food imagery. ### Focal length A lens parameter that determines angle of view and perspective. A short focal length gives broad coverage and exaggerated perspective; a long one gives compressed planes and calm proportions. Models interpret specifications such as “85 mm” literally. ### Backlight Light directed at the subject from behind, opposite the camera. It creates a glowing outline and separates the subject from the background. It is one of the most reliable ways to make a generation resemble a photograph rather than a render. ### Soft light Lighting with a gradual transition from highlights to shadows. It comes from a large source or a diffuser. It is standard for portraits, food, and cosmetics because the shadows do not break up the form. ### Colour temperature The hue of white light in a frame, measured in kelvins. A low temperature produces warm yellow light, while a high temperature produces cool blue light. In a prompt, this is often more effective than naming a colour directly because temperature affects the entire scene at once. ### Flat lay An overhead shot with objects arranged on a flat surface. It is a staple format for catalogues, recipes, and product arrangements. Models render it well, and it sells well too: browse the [food category](/en/category/food). ### Camera angle The viewpoint from which a scene is captured. An eye-level shot is neutral, a low angle makes a subject imposing, and a high angle makes it vulnerable or provides an overview. Models respond more precisely to an explicit angle than to a description of emotion. ### Macro Photography of small objects at close to life-size reproduction. In generation, it creates texture: droplets, fibres, grain, and pores. The drawback is that artefacts are especially visible in macro imagery, so these frames must be checked at full magnification. ## Files, Formats, and Dimensions This section explains what technical file specifications mean and which ones genuinely affect an image’s suitability. ### Raster An image made up of pixels. Everything produced by a generator is raster. When enlarged, a raster image loses sharpness because the file contains no additional detail. ### Vector An image described by shapes and curves rather than pixels. It scales without quality loss and is suitable for logos and icons. Generative models do not draw vectors directly; they are created by tracing a raster image. ### JPEG A raster format with lossy compression and no transparency. It is the standard for photographs and stock previews, offering a small file size at acceptable quality. It handles repeated saving and sharp boundaries poorly. ### PNG A raster format with lossless compression and transparency support. It is the right choice for graphics with sharp edges, screenshots, and anything placed over another image. It is unnecessarily heavy for photography. ### WebP A modern format that supports both lossy compression and transparency. At the same visual quality, it is noticeably smaller than JPEG, so it is the default format on the web. All current browsers support it. ### Resolution An image’s width and height measured in pixels. This is the only size a file actually has. Its suitability for print or the web is determined by pixel dimensions, not by a “300 DPI” label. ### DPI and PPI The number of dots or pixels per inch, a measurement describing output rather than a file. In a file, DPI is simply a tag that can be changed without altering a single pixel. What matters is the resolution and the desired physical dimensions of the print. ### Aspect ratio The ratio of a frame’s width to its height. It determines where an image can fit at all: a cover, story, banner, and product card each require different proportions. Changing it after generation means either cropping or outpainting. ### Alpha channel An additional file channel that stores transparency. PNG and WebP have it; JPEG does not. An image “on a transparent background” is impossible without an alpha channel, no matter how white its background appears. ### sRGB colour space The standard range of colours used by screens and the web. A stock image is delivered in sRGB because any other profile may look dull or garish on another device. Wider colour spaces are needed only in print production. ### Lossy compression Reducing file size by irreversibly discarding some information. Its effects are visible along edges and in smooth gradients as halos and banding. Every new save adds further loss on top of the previous damage. ### Metadata Administrative information inside a file, such as dimensions, date, camera settings, and authorship. For generated images, metadata often includes the prompt and settings. Metadata is usually stripped during publication, removing traces of the file’s origin along with it. ## Stock Libraries and Their Rules This section explains exactly what buyers purchase from a stock library and the terms used to describe limits on use. The legal part of the vocabulary is the most underestimated. ### Stock library A catalogue of images whose usage rights are sold repeatedly. The point of the model is accessibility, not exclusivity: the same image can legally appear in hundreds of projects at once. ### Microstock A stock library that sells a high volume of images at low prices. It is the opposite of traditional agencies offering costly exclusives. Both subscriptions and royalties per download originated with microstock. ### Licence Permission to use an image under specified terms. When buying from a stock library, the buyer acquires a licence, not the image itself or its rights. The licence terms are the only document defining permitted use; Picwin’s showcase terms are collected on the [Terms page](/en/terms). ### Royalty-free A type of licence paid for once and used repeatedly. Here, “free” means freedom from recurring payments, not no charge. Restrictions still apply, including limits on print runs, resale, and sensitive subjects. ### Royalty A payment to an author for each use or download of their work. This is the traditional microstock model. Picwin has no royalties: the platform buys out the rights once; see the [buyout mechanism](/en/blog/what-happens-to-a-work-after-publishing). ### Extended licence A licence that removes some restrictions of a standard licence. It is usually required for merchandise, large print runs, and reselling an image as part of a template. It costs more and is purchased separately. ### Editorial use The right to use an image only to illustrate news and editorial material, not advertising. This label applies to frames featuring recognisable people, brands, and events for which releases have not been obtained. ### Commercial use The right to use an image in advertising, packaging, and promotion. It requires the image to contain no recognisable third-party elements without permission, hence releases and restrictions on brands appearing in the frame. ### Model release A person’s written consent to the use of their likeness. Without one, a frame with a recognisable face remains editorial. A generated face has no model release and cannot have one, so it is important that the face does not match a real person. ### Property release The owner’s consent to commercial use of an object’s image. It is required for private buildings, interiors, works of art, and recognisable designed objects. ### Exclusivity A condition under which an image is sold through only one platform or to only one buyer. It raises the price and narrows the pool of buyers. The Picwin showcase does not sell exclusive rights: a work remains in the catalogue after purchase. ### Keywording Selecting the keywords through which an image is found in a catalogue. On traditional stock sites, this is the author’s job and a major factor in sales. In Picwin, titles and descriptions are assembled from the catalogue assignment, so authors do not need to handle keywording. ## How Picwin Works This section explains terms used in the mini app and on the showcase. It describes a specific product rather than general theory. ### Mini app An application that runs inside Telegram without installation from an app store. That is how Picwin operates: the assignment catalogue, generation, and work gallery all open inside the messenger. ### Catalogue assignment A ready-made catalogue topic from which an author creates a work. An assignment has a category, difficulty, and prompt hint. There are forty categories, ranging from [interiors](/en/category/interior) to [technology](/en/category/technology). ### Brief A task specification consisting of an assignment topic and randomly selected conditions. It is assembled before a run and determines how the work will be judged. A detailed explanation appears in [How to Become a Picwin Author](/en/blog/how-to-become-a-picwin-author). ### Brief conditions A randomly selected purpose and set of restrictions for the topic. The purpose says what the frame is intended for; the restrictions identify what it must not contain. The draw is repeated each time, so the same topic is interpreted differently. ### Assignment difficulty An assignment’s level of complexity: easy, medium, or hard. The level determines the strictness of the quality check and positions the work higher in the showcase’s pricing structure. It is neither a payout amount nor the rarity of a buyer. ### Run A single generation attempt for an assignment with the selected budget. It requires either a stake or an available free generation. The run’s result is determined by the server and does not depend on how quickly the author opens the image. ### Quality check An assessment of a finished work against the selected assignment’s difficulty. A work that passes the check is bought out at a higher valuation and enters the showcase; one that fails is still bought out, but cheaply, and remains in the author’s gallery. ### Buyout The platform’s purchase of image rights from the author. It happens immediately after generation, with no external buyer waiting. The fee is paid once: later sales of copies bring the author nothing. ### Showcase A public catalogue of works available to every site visitor. A work appears there automatically if it passes the check. The showcase never displays the author’s name: the card identifies only the work, category, and price. ### Hidden work A work removed from the showcase by a moderator but retained by its author and buyers. A completed sale is not cancelled: the buyer keeps the licence. The stricter action is withdrawal from sale, in which case buyers receive refunds. ## Quality and Common Flaws This section explains what to look for in a finished frame before considering it successful. Almost everything listed here is visible only at full magnification. ### Artefact Any distortion caused by the generation or compression process itself. It is an umbrella term covering extra fingers, halos around edges, and broken fragments of patterns. It is the first thing to look for when checking a frame. ### Garbled text Marks that resemble letters but do not form words. It is a classic sign of generation, appearing on signs, labels, and book covers. Removing text from the frame entirely is easier than making it correct. ### Anatomy error An incorrect number or structure of body parts. Hands, teeth, ears, and limb joints are the most vulnerable areas. Such a work is unsuitable for stock use because the first attentive viewer will notice. ### Oversharpening Excessive sharpening that produces bright halos along edges. It often appears after upscaling. It looks like artificial sharpness and is particularly conspicuous where an object meets the sky. ### Plastic skin Portrait skin that is overly smooth and lacks pores and microtexture. It is a sign of overly aggressive settings or excessive smoothing. It can be corrected by adding texture and lowering the CFG scale. ### Banding Visible steps in place of a smooth colour transition. It is easy to see in skies, gradient backgrounds, and shadows. It results from insufficient bit depth or aggressive compression. ### Noise Image grain unrelated to the content of the frame. In moderation it adds realism; in excess it looks defective. After strong upscaling, noise is distributed unevenly, which makes it noticeable. ### Duplication An extra copy of an object or detail that appears on its own. Examples include a second set of buttons, an extra chair leg, or a window drawn twice. It should be corrected with inpainting rather than a new run. ## Image Rights and Provenance This section explains the legal boundaries surrounding an AI image. There are fewer definitive answers here than in the other sections, and that uncertainty is an honest part of the picture. ### Copyright in an AI image A disputed area in which some countries deny protection to images made without human creative input. Practice differs between jurisdictions and continues to change. That is why stock libraries define relationships through licences rather than relying on copyright. ### Deepfake A synthetic image or video depicting a real person in a situation that never occurred. It is expressly prohibited by the terms of most platforms. For stock imagery, this is not a grey area but a red line. ### Likeness The recognisability of a specific living person in a generated frame. Even an accidental match creates risk: personality rights protect a person regardless of whether they were photographed or drawn. ### Trademark Someone else’s logo, recognisable packaging shape, or element of brand identity appearing in a frame. Such a frame is unsuitable for commercial use. That is why assignment conditions often expressly prohibit brands. ### Attribution A requirement to identify an image’s author or source when using it. Stock licences usually do not require it, while open licences often do. The absence of an attribution requirement does not mean there are no other restrictions. ### Provenance A traceable history of a file’s origin: who created it, which tools they used, and what was done to it afterward. Content-signing standards such as C2PA exist to preserve this history. For now, these signatures are applied voluntarily and are easily lost when a file is resaved. ### AI disclosure A notice that an image was generated rather than photographed. Some countries and platforms already require it for advertising and news. The practical conclusion is simple: record the origin immediately rather than trying to remember it a year later. **Key takeaways** - Seed, sampling steps, and CFG scale are the three controls worth learning first. - “Royalty-free” means “without recurring payments,” not “free of charge and unrestricted.” - DPI in a file means nothing by itself: pixel resolution is what matters. - Model and property releases separate commercial use from editorial use. - Picwin has no royalties: the platform buys out the rights once, immediately. - Look for artefacts at 100% zoom, not in a preview. See the terms in action The assignment catalogue, briefs, and generation all live in the Picwin mini app, where half of this glossary appears directly on screen. [Open the mini app](/en/partners) ### How is “royalty-free” different from “free of charge”? Royalty-free is a paid licence without recurring royalties: you pay once and use the image repeatedly within the terms. It has nothing to do with being free of charge, and restrictions on print runs, resale, and sensitive subjects still apply. ### Which matters more for quality: the number of steps or CFG scale? CFG scale has a stronger effect and can ruin a frame more quickly. Steps improve the result only up to a certain point; beyond it, they add time rather than detail. ### Why lock the seed if the image already turned out well? To make comparisons. With a fixed seed, you can see exactly what the edited part of the prompt changed; without one, every run changes everything at once, leaving no basis for a conclusion. ### Does a generated face need a model release? There is nobody from whom to obtain one because the person does not exist. What matters instead is ensuring that the face does not resemble a specific living person, which would raise a likeness-rights issue. ### Is 300 DPI really required for printing? No. DPI in a file is just a tag. Pixel resolution combined with the physical print size is what matters: the same image may be suitable for a postcard but not for a poster. ### Why are there no royalties in Picwin? Because the platform buys the image rights from the author immediately after generation and assumes all risk of subsequent sales. Learn more in [What Happens to a Work After Publishing](/en/blog/what-happens-to-a-work-after-publishing). ### Where should I start if almost all these terms are unfamiliar? Start with the section on prompts and the section on Picwin: they describe what you encounter on your very first run. The rest will become familiar as you read briefs and licence terms.