Midjourney & Visual AI glossary from A to Z — Midjourney & Visual AI

Midjourney & Visual AI glossary — 124 definitions

Every term defined across the guide's 31 sections, in alphabetical order.
Tip: Ctrl + F to find a term on the page.

--ar (aspect ratio): Parameter setting the image format (16:9, 9:16, 1:1, 4:5…); choose it right from generation.
--chaos (--c): Diversity across the 4 proposals: 0 = very similar, high value = very different directions.
--cref (character reference): Reference image for keeping the same face/character from one generation to another.
--cw (character weight): Weight setting the resemblance intensity to the reference character.
--seed: Random seed; fixing it makes the result reproducible and isolates a change's effect.
--sref (style reference): Reference image carrying the visual signature (palette, texture, treatment).
--stylize (--s): Intensity of Midjourney's own style: low = faithful to the prompt, high = more aesthetic but freer.
--weird (--w): Dose of weirdness/originality to escape expected renderings; experimental at high value.
Visual A/B testing: Comparing several versions of a visual with real audiences to decide on data.
Adobe Firefly: Adobe's generative AI built into Photoshop (generative fill), geared toward pro production and commercial safety.
Photographic grounding: Medium and technique terms (photograph, 35mm film, film grain, film-stock brand) steering toward a photo rendering.
Camera angle: Camera position: eye level, low angle, high angle, top-down, dutch angle.
Midjourney web app: Official online interface (input, gallery, editing) more accessible than Discord for starting out.
Automatic1111 (A1111): Classic tab-and-form web interface; most accessible for starting locally.
Avatar / virtual presenter: Synthetic character presenting content; common teaching and internal use.
Model bias: Stereotypes reproduced by the model (gender, origin, body); actively correct in prompt and selection.
Bing Image Creator / Copilot: Free access to DALL·E 3 via Microsoft services, with quotas.
C2PA / Content Credentials: Signed-provenance metadata standard: creation and edit history of a piece of content.
Internal usage framework: Rules defining what is allowed, verified and disclosed when using AI images in business.
Certifying the authentic: The standard's logic: proving a true content's origin rather than trying to detect every fake.
AI visual guideline: Document setting how the brand generates its images (pillars, template prompts, parameters, references, prohibitions).
Series consistency: Getting several images that share the same style or character (see sections 3.5 and 5.1).
Temporal consistency: Stability of objects, characters and light from frame to frame; generative video's central difficulty.
ComfyUI: Node interface visually connecting steps; more demanding but highly reproducible and powerful.
Transferable skill: Know-how (prompt, light, framing) valid whatever tool you use.
Terms of Service (ToS): The tool's contract defining your rights over generated images; read them, they differ from one service to another.
Documented consent: A person's explicit, recorded agreement before any use of their face or voice. Non-negotiable.
ControlNet: Extension imposing structure on the image from a control image (pose, edges, depth, segmentation).
Trial cost: Cost of producing one variant; its drop lets you test instead of guessing.
Generative credits: Units consumed with each generation in some services (e.g. Adobe Firefly).
DALL·E 3: OpenAI model built into ChatGPT/Copilot; strong at natural-language understanding and text in images.
Denoising: Iterative noise-removal process that gradually reveals the image.
Adaptation: Adapting a visual into several formats, variants or messages (see section 5.4).
Multi-format adaptation: Adapting one concept to all formats and messages, generated directly at the right ratio.
Deepfake: Misleading synthetic depiction of a real person; avoid outside an explicit, consensual framework.
Discord: Messaging platform where Midjourney first ran: you talk to a bot via commands.
Pipeline documentation: Formalizing the working sequence, to save time, stay consistent and delegate.
Image rights: A person's right over depictions of their face; generating an identifiable person without consent is risky.
Copyright: Protection of original works; its application to generated images varies by country and keeps evolving.
Embedding: Mathematical representation of text that lets the model steer generation.
Rejected examples: Non-compliant visuals documented in the guideline; highly educational for framing production.
Exploration vs refinement: Two phases: explore (high chaos/weird, free seed) then converge (low chaos, fixed seed).
Quality/compliance filter: Pre-release check: flaws, bias, protected elements, AI-nature disclosure.
Flat design / vector: Graphic style with flat color areas, no realistic texture or shadow; highly reproducible in series.
Focal length (lens): 35mm (wide/natural), 50mm (human vision), 85mm (portrait), telephoto (isolates), wide-angle (exaggerates).
Probabilistic generation: Random nature of the result: the same prompt gives different images on each run.
Golden hour / blue hour: Warm grazing light at the start/end of the day; bluish dusk glow.
Grid of 4: The four proposals generated from one prompt, among which you choose.
Illustration vs documentation: Illustrating an idea (a use suited to AI) ≠ documenting a reality (real photo required).
Image-to-video: Animating an existing image; the most controllable approach since framing and style are already fixed.
img2vid (image animation): Animating a still portrait to give it natural micro-movements and bring it to life.
Inpainting: Regeneration of a SELECTED area of the image, blended with the rest; in Midjourney: Vary (Region).
IP-Adapter: Transfer of a reference image's visual content (face, style, mood) to new generations.
IP-Adapter / LoRA: Stable Diffusion-side equivalents: face/style transfer, or training a character or style (section 4.3).
Isometric: Projection view without perspective, used for diagrams, technical scenes and product illustrations.
Iteration: Working method: start simple, then add one element at a time to measure its effect.
Lip sync: Synchronizing mouth movements with an audio track.
LoRA: Small additional-training file teaching a model a precise concept (character, style, product).
Low key / high key: Mostly dark, dramatic image vs bright, low-contrast image.
Mockup / concept: Provisional visual used to validate an idea before actual production.
Medium: Named visual technique (watercolor, ink, gouache, 3D render, vector…) that determines the rendering.
Midjourney: Paid tool renowned for the aesthetic quality of its renderings; available via Discord or its web app.
Fast mode: Priority, fast generation, counted in hours included in the subscription.
Stealth mode: Higher-plan option keeping your generations from public view.
Relax mode: Unlimited but queued (hence slower) generation, available on higher plans.
Community model (checkpoint): Specialized model shared by the community; check its provenance AND usage license.
Diffusion model: Model that learns to remove noise step by step; it starts from random noise to make an image emerge.
Open / proprietary model: Open: weights downloadable and runnable at home. Proprietary: available only through the vendor's service.
Mood board: Ambience board gathering reference visuals to set an art direction.
Obsolete keywords: Terms like "8K, ultra realistic, high quality" inherited from old models, often useless or even harmful today.
Camera move: Movement indication (dolly, pan, slow zoom) requested in the video prompt.
Open source (Stable Diffusion): Freely downloadable and runnable model, with no subscription or technical usage limit.
OpenPose / Canny / Depth: Control types: pose skeleton, edge detection, depth map.
Finishing order: Fix flaws first, enlarge after, then apply color-grading adjustments.
Outpainting: Extending the image beyond the existing frame: the model invents what lies around.
Constrained palette: Limited number of colors imposed in the prompt; key to series and brand-guideline consistency.
Pan: Extending the image in a precise direction (left, right, up, down).
--no parameter: Midjourney's negation parameter for excluding an element (more reliable than writing "no…").
Frozen parameters: Locked model version, stylize and formats to avoid drift between productions.
Personalization: Varying visuals by audience segment, season or region.
Personalization (personal style): Mechanism by which the model learns your aesthetic preferences from your choices.
Composition phase: Assembling with logo, typography and brand elements in a design or layout tool.
Creation phase: Generating the visual base (composition, light, mood) with the aesthetically strongest tool.
Photorealism: Rendering indistinguishable from a photograph; achieved by adding realistic imperfections, not by asking for "ultra realistic".
Combined pipeline: Simultaneous use of several extensions (LoRA + ControlNet + IP-Adapter) in one professional workflow.
Word weighting / order: Terms placed at the start of the prompt generally weigh more on the result.
Depth of field: Limited sharpness zone (shallow depth of field); strong marker of photographic realism.
Prompt: Text description sent to the model; it guides image generation.
Descriptive prompt: Description of what must be SEEN (concrete visual elements) rather than an abstract emotion.
Template prompt: Standardized formula with variables, reusable by all to guarantee consistency.
Free quota: Limited number of generations offered per period (day/month) in subscription-free offerings.
Re-roll: Full rerun of the prompt to get four new proposals unrelated to the previous ones.
Brand reference: Style image, style identifier or home-trained LoRA embodying the visual identity.
Character reference (--cref): Reference image used to keep the same character from one image to another.
Style reference (--sref): Reference image supplied to carry a visual signature from one generation to another.
Composition rules: Rule of thirds, centered composition, leading lines, framing, negative space (free space for text).
Generative fill: Photoshop feature (via Firefly) for regenerating or removing a selected area.
3D render (3D render, low poly, clay): Families of three-dimensional renderings, from photorealistic to stylized (low poly, clay).
Manual retouching: Correction with classic tools (clone stamp, healer, density) where AI does not blend correctly.
Small-touch retouching: Method of chaining several targeted local fixes rather than one big regeneration.
Rim light / backlight: Bright edge outlining the subject; lighting coming from behind (backlit).
Runway / Pika / Sora / Veo: Main generative-video services; fast-moving ecosystem, check regularly.
Soft light / hard light: Soft light with gradual shadows (flattering) vs hard light with sharp shadows (contrasty).
Stable Diffusion: Open model, installable locally; free and highly customizable (ControlNet, LoRA), but more technical.
Local Stable Diffusion: Running the model on your machine: free, unlimited, private and fully tunable.
Prompt structure: Template: subject + action + context + style + light + framing + parameters; each block reduces randomness.
Voice synthesis: Generating a voice from text (e.g. ElevenLabs), with control of tone, rhythm and language.
Shot size: Apparent distance to the subject: close-up, medium shot, full body, wide shot / establishing shot.
Text-to-video: Generating a sequence from a text description; flexible but hard to control.
Skin texture: Explicit request for pores and irregularities; breaks the "plastic" look of generated faces.
Internal traceability: Keeping your prompts, versions and source files as evidence of your approach.
Non-destructive work: Keeping originals, layers and versions to fix and adapt later.
Upscale (U1-U4): High-definition enlargement of the chosen image; produces the usable final version.
Upscaler: Enlarging tool that rebuilds detail (Topaz Gigapixel, Magnific, Stable Diffusion upscalers).
Creative upscaler: Upscaler that invents detail: spectacular result but can alter the image — watch out.
Variations (V1-V4): New proposals inspired by a selected image: same spirit, different details.
Vary subtle / strong: Variation intensity: slight (final adjustment) or strong (exploring the same direction).
Model version (--v): Midjourney model generation used; each version substantially changes renderings.
Volumetric light (god rays): Light rays made visible by haze or dust.
VRAM: Graphics-card memory; the main limiting factor for local generation.
Invisible watermark: Imperceptible signature embedded in pixels, detectable by dedicated tools; fragile to edits.
Visible watermark: Apparent disclosure flagging a generated image; readable but degrading and easy to crop.
Combined workflow: Tool pipeline where each acts on what it does best: create, fix, compose, finalize.
Zoom Out: Zoom-out that widens the field; turns a close-up into a broader scene and recovers margin.