AI for UX · Step 2 · 8 min read
AI for visual design: images and video
How image and video generators work in plain words, the main tools designers use in late 2026, how to brief them for consistent results, what you may and may not do with the output (licences, copyright, the EU labelling rule, Content Credentials), and how to keep generated imagery accessible and fair.
Generating an image now takes seconds; knowing whether you may use it takes longer. Image and video models are excellent for moodboards, concepts, placeholders and, with care, final assets. This note explains how they work, which tools designers use, how to get consistent results, and the rights, labelling and accessibility questions to settle before anything ships.
What you'll be able to do after reading:
- Explain in plain words how a diffusion model makes an image
- Choose a tool for a task: exploration, brand-safe assets, video, or local generation
- Brief an image model for consistent, on-brand results
- Know the basics of licences, copyright and the EU rule on labelling AI content
- Keep generated images accessible and representative
Tool details are as of November 2026. Licences and terms change: read them for each tool and model before commercial use.
1. How image models work, in plain words
Most image and video generators are diffusion models. During training, the model sees millions of images with noise added step by step, and learns to remove the noise. To generate, it starts from pure random noise and refines it over a number of steps, guided by your text, until an image appears (Hugging Face Diffusion Course; the original method: Ho et al., 2020).
- 1Pure noise
A random field of pixels, fixed by the seed.
- 2Rough shapes
Early steps find large forms that match the prompt.
- 3Details
Later steps add edges, texture and light.
- 4Image
After a few dozen steps, a finished picture.
What follows from this:
- Every generation is different. The starting noise is random; a "seed" number fixes it so you can reproduce a result.
- The model knows patterns, not facts. It can draw "a hand", but may give it six fingers; it can draw a sign, but may misspell it (newer models are much better at text).
- It reflects its training images, including their gaps and stereotypes.
2. The tools, by job
| Job | Tools (Nov 2026) | Notes |
|---|---|---|
| Images inside your design tool | Figma Weave and Weave tools | Many models in one place, paid with Figma AI credits; check Weave's own terms (see Figma AI) |
| Brand-safe stock-style assets | Adobe Firefly | Adobe says Firefly models are trained on Adobe Stock and public-domain content and are "safe for commercial use"; partner models offered inside Firefly are not covered by that claim |
| Art direction and illustration | Midjourney | You own your outputs; companies earning more than US$1M a year need a Pro or Mega plan for commercial use |
| General images and editing | OpenAI image models, Google Gemini Image ("Nano Banana") | Strong at following instructions and text in images; outputs carry provenance marks |
| Video | Google Veo, video models in Weave and Firefly | Veo 3.1 adds sound and up to 4K |
| Local, on your own computer | Open-weight models such as FLUX 3 Image Klein (Black Forest Labs) | No upload of your material; each model has its own licence, some non-commercial |
3. Briefing an image model
A good image prompt reads like an art-direction brief:
| Part | Example |
|---|---|
| Subject | "A member in their sixties stretching after a game of padel" |
| Setting and moment | "Indoor court, early evening, warm light through high windows" |
| Composition | "Medium shot, subject on the left third, space for a headline on the right" |
| Style | "Documentary photography, natural colour, shallow depth of field" |
| Brand cues | "Muted greens and off-white, like our brand palette; no logos" |
| Avoid | "No text, no watermarks, no exaggerated smiles" |
For consistency across a set:
- Keep a written style block and reuse it word for word.
- Use reference images when the tool supports them (several tools accept your own images as style or subject references).
- Fix the seed while you adjust the prompt, so you can see what each change does.
- Describe people fully. If you only give a style reference, the model may copy the person in the reference.
4. What you may do with the output
The terms of the tool and of the model decide commercial use.
Check per tool and per modelIn the US, prompts alone do not make you the author.
Keep a record of your human editsSince Aug 2026, realistic deepfakes must be disclosed.
AI Act, Article 50C2PA records how an image was made; platforms may strip it.
Content CredentialsLicences and terms
Each tool's terms decide commercial use, ownership and what you may upload. Read them per tool and per model: the same app can offer models with different terms (Firefly's partner models are the clearest example).
Copyright
In the US, the Copyright Office's January 2025 report says AI outputs are protected "only where a human author has determined sufficient expressive elements", and not through "the mere provision of prompts". Human selection, arrangement or modification of AI material can be protected (US Copyright Office). Other countries differ: if ownership matters (a logo, a key brand asset), ask a lawyer and keep records of your human contribution.
Labelling in the EU
Article 50 of the EU AI Act has applied since 2 August 2026. Providers of generators must mark outputs in a machine-readable way; those who publish deepfakes (realistic images, audio or video of real people, places or events) must disclose that they're AI-generated, with lighter rules for clearly artistic or satirical work (Article 50; EU Code of Practice on marking and labelling, Jun 2026). See AI governance and data protection.
Content Credentials (C2PA)
Content Credentials are a standard way to attach a record of how an image was made, backed by Adobe, Microsoft, Google, OpenAI, Meta, the BBC and others (contentcredentials.org; C2PA spec). Firefly and OpenAI outputs carry them. Platforms can strip this metadata, so don't rely on it alone; some tools add invisible watermarks too.
5. Accessibility and fairness
- Alt text still applies. Every meaningful image needs a text alternative that describes its information or function; decorative images get
alt=""; charts need a full text equivalent (W3C WAI images tutorial; WCAG 1.1.1, Level A). AI can draft alt text; a person checks it. - Representation matters more than source. In an NN/g study (Aug 2026), 77 participants rated AI-generated hero images no worse than stock photos; what mattered more was who was represented, but suspecting an image was AI lowered ratings (NN/g).
- Watch for stereotypes. Ask for specific, varied people (age, body, skin tone, disability) instead of "a professional" or "a happy family", and review the set as a whole. See Algorithmic bias.
- Never fake evidence. Don't use generated images of "real" customers, testimonials, products or places where people would assume they're real.
6. Vocabulary
| Term | Meaning |
|---|---|
| Diffusion model | A model that creates images by removing noise step by step |
| Seed | The number that fixes the starting noise, so a result can be reproduced |
| Reference image | Your image given to the model as a guide for style or subject |
| Open weights | A model you can download and run yourself, under its licence |
| Deepfake | Realistic AI content showing real people, places or events that didn't happen |
| Content Credentials (C2PA) | Attached records of how a piece of media was made |
7. Try it (no code needed)
- Write a style block for a brand you know (palette, light, composition, what to avoid).
- Generate four images for the same subject with an approved tool, reusing the block word for word.
- Review the set: consistent? on brand? who is represented, and who is missing?
- Write alt text for one image, then ask the tool for alt text and compare.
- Find the tool's terms: may you use these images commercially? Write the answer and the link down.
Reflection questions: Where in your work would a generated image be fine, and where would only a real photo be honest? Which of your images would need a label under the EU rule?
8. If you lead a team
- Keep a short list of approved image tools and models, with their commercial terms.
- Agree when generated imagery is allowed (concepts, placeholders, illustration) and when it isn't (real people, testimonials, evidence).
- Keep records of prompts and human edits for key brand assets.
- Add alt text and representation to your design review.
Checklist
Learn more
- How it works: Hugging Face Diffusion Course · Denoising Diffusion Probabilistic Models (Ho et al., 2020)
- Rights and rules: US Copyright Office, AI report part 2 · EU AI Act Article 50 · EU Code of Practice on AI-generated content · Content Credentials
- Accessibility and research: W3C WAI images tutorial · NN/g on AI-generated images
Books, courses and videos to go deeper
Related
AI for content design · Diffusion model · Figma AI · Algorithmic bias · AI governance and data protection · AI design review process · Prototyping with AI