Start with your situation, not the generate button
Image generation looks like a slot machine and works like a craft: the people who get good results treat the prompt as a brief and each generation as a draft. This guide teaches that workflow in Canva's generators (Magic Media inside the editor, and Dream Lab, Canva's higher-quality generation surface), with a worked example iterated three times, because the iteration loop, not the first prompt, is where quality comes from.
| Your situation | Your honest starting point | What this guide gives you |
|---|
| First time generating | Steps 1 to 3 with the worked example open | The prompt formula and the variation controls |
| Your results look generic or wrong | Step 3, the iteration loop | What to change between generations, in order |
| You generate for client or product work | The selling rules section first | What the AI terms allow, and what marketplaces demand |
| You are on the free plan | The allowance section before generating | How to learn the tool without burning the month's pool |
| You need the whole AI suite | The map article | Our Canva AI guide maps every Magic tool by job |
Generators live inside the editor: open a design, find Magic Media in the Apps or elements area (or open Dream Lab from Canva's AI entry points), describe the image, choose style and aspect ratio, and generate. The result lands in your design like any image, ready to crop, filter, and layer. Everything below is about making that result good and usable.
Step 1: The prompt formula
A working prompt has five parts, in this order: Subject (the main thing, specific: "a ceramic teapot with a crackle glaze", not "a teapot"), Setting (where it is: "on a wooden cafe table beside a window"), Style (the visual language: "editorial food photography", "flat vector illustration", "watercolor"), Light and mood ("morning window light, warm and calm"), and Format cues ("close-up", "wide shot", "copy space on the left" for designs that need text room). Write prompts like you are briefing a photographer who cannot ask questions. Vague prompts return the average of the internet; specific prompts return something with a point of view. Save prompts that work in a notes page; a personal prompt library is a real production asset, and the element codes guide teaches the same library thinking for stock elements.
Step 2: Style, aspect ratio, and the four variations
Choose the style preset to match the job before generating: photo styles for anything meant to look real, illustration styles for graphics, and consistency of style across a set matters more than any single image's brilliance. Set the aspect ratio to the destination (square for posts, wide for banners, the design's own shape when generating inside a frame) because regenerating at a new ratio crops your composition, not just your canvas. Generate and review the variation set as a set: you are looking for the best composition and the fewest defects, not perfection. Check hands, text, and edges at full zoom; AI defects hide at thumbnail size and announce themselves in print. Pick the winner and use the variation and edit tools from there, rather than regenerating the whole brief.
Step 3: The iteration loop, what to change between generations
Worked example: a bakery needs a hero image for a sourdough post. Generation one, prompt: "bread on a table, photo". Result: competent, generic, could be any bakery on earth. The loop now changes one thing per round, in this order of leverage:
- Specificity round. "A rustic sourdough loaf with a scored crust, sliced, on a linen cloth, bakery counter, editorial food photography, morning window light." Result: a real brief's image, warm and specific. Most of the quality jump lives in this round, and it costs one generation.
- Composition round. Add format cues: "wide shot, loaf on the right third, copy space on the left for text". Result: an image designed to be designed with, which is the actual job inside Canva.
- Polish round. Generate variations of the winner, pick the cleanest, then stop generating and start editing: the result goes through the normal photo routine (crop, adjust, filter at half strength) from the photo editing guide, inside the design, where text and brand colors finish the job.
Three generations, one brief. The discipline is changing one variable per round and knowing which variable: specificity first, composition second, selection third. Endless regenerating with the same prompt is not iteration; it is hoping.
Step 4: Edit the result inside the design
A generated image is raw material. Inside the design you can crop it into frames, apply filters for a consistent feed look, use Magic Edit to change a region (the wrong-colored prop, the sign with gibberish text), and layer your text and elements over it. Generated text inside images is unreliable: if the image needs words, keep them out of the prompt's image and add real text boxes, which you control. For a set of images (a campaign, a product line), generate them in one session with the same style words and the same seed prompt structure, then unify them with one filter at low intensity across all of them; consistency reads as professionalism.
The allowance, spent like a budget
Generation draws from the same shared monthly AI allowance as every Magic tool, detailed in the suite guide. Generation is the hungriest consumer in the suite, so budget it: draft prompts on paper, generate in review sets of four, and do exploratory styles early in the allowance month. Free accounts get a smaller pool, which is enough to learn the loop on a handful of images monthly; production volume is a paid-plan conversation. When the pool is empty, the stock library is not: search first for images that exist, and generate only the ones that do not, a rule that also happens to produce more believable designs.
The selling rules for generated images
Canva's AI terms, not the stock license, govern generated content. The working summary, with the caution that terms evolve and you should read the current AI terms before a commercial launch: generated content can generally be used commercially in your designs, you do not own the generation the way you own a photograph you shot (others may generate similar results), and you must not present generated content as unaltered human work where disclosure is required. Marketplaces add their own layer: several require AI disclosure on listings, and some categories restrict AI imagery. The money guide puts these rules in the full route context; the safe production pattern is generated imagery as an ingredient inside your composition and branding, honestly disclosed where a platform asks, never a raw generation sold as the product itself. Trademark use of generated logos carries the same caution as element-based logos: do not build a registrable brand mark on it.
When not to generate
Three cases. When the real thing exists and trust matters (your actual product, your actual team, your actual premises), photograph or use real library photography; audiences smell generated imagery in trust contexts. When the image is a person who will be recognized as a fake human endorsing something, which is a disclosure and ethics problem, not a style choice. And when a library photo is simply better and free of allowance cost, which is more often than the novelty suggests. Generation is a tool for the images that do not exist; the library is the tool for the ones that do.
Related guides
Frequently asked questions
What is the difference between Magic Media and Dream Lab?
Both generate images from text prompts inside Canva's ecosystem. Magic Media is the in-editor generator you use without leaving your design, with style presets and aspect ratios matched to the canvas. Dream Lab is Canva's higher-quality generation surface, promoted for richer, more detailed results. In practice: Magic Media for quick in-design needs, Dream Lab when the image is the centerpiece and quality justifies the trip.
How do I write a prompt that gets good results?
Use the five-part formula: specific subject, setting, style, light and mood, and format cues like shot type and copy space. Brief the generator like a photographer who cannot ask questions. Then iterate one variable at a time: specificity first, composition second, then choose among variations. Generic prompts return generic images; the formula is the fix.
Why do my generated images have weird hands and text?
Hands, text, and fine repeated details are known weak points of image models. Mitigate by choosing compositions that avoid them (hands out of frame or holding large simple objects), keeping all real words in Canva text boxes rather than in the image, and inspecting at full zoom before use. Magic Edit can repair a small defective region without regenerating the whole image.
Can I sell designs that include AI-generated images?
Generally yes under Canva's AI terms, with conditions: read the current terms before a commercial launch, disclose AI content where your marketplace requires it, and do not sell a raw generation as a standalone product or build a trademark on generated art. The durable pattern is generated imagery as one ingredient in your own composition, which is also what the money guide recommends.
How much does image generation cost on the free plan?
It draws from the free plan's shared monthly AI allowance, the same pool as Magic Write and the other Magic tools. The pool is small enough that free users should generate deliberately: paper-drafted prompts, sets of four, library first for images that already exist. When the allowance is spent, generation waits for the monthly reset or a paid plan.
Who owns an image I generate in Canva?
Ownership of AI output is limited by the nature of the technology and by Canva's terms: you receive rights to use the content under the AI terms, but you cannot claim the kind of exclusive ownership you hold in an original photograph, and Canva does not guarantee another user could not generate something similar. Treat generated images as licensed ingredients, not as owned assets, in any business planning.
Photo credits: Dmitry Mashkin (Pexels). Photos are served from this site; follow the credit to the library page for the original.