How does Adobe Firefly work?
Adobe Firefly turns a text prompt into images, video, or vectors using generative AI models trained on licensed Adobe Stock and public-domain content. You type a description, Firefly generates several options in seconds, and you refine with settings or reference images. It runs at firefly.adobe.com and inside apps like Photoshop and Express.
Why — the first-principles explanation
Firefly is built on diffusion models. During training, the model is shown millions of licensed images paired with text descriptions, then taught to add noise to those images and reverse the process. It learns the statistical patterns that connect words like "golden retriever on a beach" to the shapes, colors, and textures that make up such a picture.
When you type a prompt, Firefly starts from a field of random noise and denoises it step by step, nudging the pixels toward something that matches your words. It isn't copying a stored photo; it's reconstructing an image from learned patterns. That's why you get several different results from the same prompt, and why small wording changes shift the output.
The part that sets Firefly apart is what it learned from. Adobe trained it on Adobe Stock (already licensed), openly licensed work, and public-domain content, deliberately avoiding random web scraping. This is a design choice about legal safety, not image quality, and it's why Adobe markets Firefly outputs as commercially usable.
Firefly also embeds Content Credentials (open C2PA metadata) into what it makes, acting like a tamper-evident label that says "AI-generated with Firefly." And because Adobe owns Photoshop, Illustrator, and Express, the same Firefly engine powers in-app tools like Generative Fill, so you rarely visit the Firefly website to use it.
An example that makes it click
Imagine an art student who studied 10,000 labeled paintings but is never allowed to trace any of them. Ask for "a foggy lighthouse at dawn," and they start with a blank, smudgy canvas and keep sharpening it, foggy shape, then a tower, then warm dawn light, until it matches your words. Ask three times and you get three different lighthouses, because they're painting from memory of patterns, not copying one picture.
Firefly is that student, sped up to a few seconds, and it only ever studied paintings it was allowed to see. It even signs the back of each canvas with a little label saying "made by AI" so anyone later knows how it was created.
How to do it
- Go to firefly.adobe.com and sign in with a free Adobe account, or open Firefly features inside Photoshop, Illustrator, or Express.
- Pick a module: Text to Image, Generative Fill, Text Effects, Generative Video, or others.
- Type a clear prompt describing subject, style, lighting, and mood.
- Let Firefly generate several variations in a few seconds.
- Refine using controls like aspect ratio, content type, style presets, or a reference image.
- Download your favorite; note that free-plan downloads include a watermark and all outputs carry Content Credentials.
Key facts
- Firefly uses diffusion-based generative models that build images from noise guided by your text prompt.
- Training data is licensed Adobe Stock, openly licensed, and public-domain content, not general web scrapes.
- Every Firefly output carries Content Credentials (C2PA metadata) marking it as AI-generated.
- Firefly powers in-app features like Photoshop's Generative Fill, not just the standalone website.
- Firefly is accessible via a free plan at firefly.adobe.com and through Creative Cloud apps.
Adobe's commercially-safe image generator, trained on licensed content.
Affiliate link — we may earn a commission at no cost to you.▶ The 60-second explainer (script)
Here's how Adobe Firefly actually works. Under the hood it uses something called a diffusion model. During training, it studied millions of licensed images paired with text descriptions and learned to connect words to shapes, colors, and textures. When you type a prompt, Firefly starts with a screen of random noise and cleans it up step by step, nudging the pixels until they match your description. It's not pulling up a saved photo; it's rebuilding an image from learned patterns, which is why the same prompt gives you several different results. What makes Firefly special is what it learned from: Adobe Stock, openly licensed, and public-domain content, not random web scraping. That's why Adobe calls its outputs commercially safe. Firefly also stamps every image with Content Credentials, an open metadata label that marks it as AI-generated. And because Adobe makes Photoshop and Illustrator, the same engine powers tools like Generative Fill right inside those apps. To try it, just visit firefly dot adobe dot com, type a description, and watch it generate.
What authoritative sources say
People also ask
Does Firefly copy existing images?
No. It generates new images from learned patterns by denoising random static toward your prompt, rather than retrieving or pasting a stored picture.
What can Firefly generate?
Images, vector art, text effects, and video, plus in-app edits like Generative Fill. You describe what you want in plain language.
Do I need Photoshop to use Firefly?
No. Firefly works on its own at firefly.adobe.com with a free account, and it's also built into Photoshop, Illustrator, and Express.
How does Firefly mark AI images?
It embeds Content Credentials, open C2PA metadata that acts like a nutrition label showing the image was AI-generated with Firefly.