Most Adobe Firefly walkthroughs show you a prompt box, a few sliders, and a gallery of pretty results. That’s the easy half. The tool starts earning its keep when you have a specific job to finish: a product photo with a stray charging cable in frame, a hero image that needs re-cropping for three social formats, a client who wants the vase terracotta instead of sage.
What follows is the order I actually work in, with the prompts, settings and hard limits that matter.
Set up the text-to-image panel before you type a word
Open the Firefly web app, pick Text to Image, and resist the urge to start writing. Four controls shape the output more than your adjectives do:
- Aspect ratio. 1:1 for product tiles, 16:9 for web heroes, 9:16 for Stories. Choosing this first stops you re-generating later just to fit a layout.
- Content type. Photo, Art or Graphic. Switching to Graphic nudges the model toward flat, vector-friendly compositions.
- Model version. Newer models handle hands and complex scenes noticeably better, and they cost more credits per generation. Don’t reach for the biggest one when you’re sketching.
- Reference image. Upload a shot you like and use it as a style reference, a structure reference, or both. This is the single fastest way to match an existing brand look.
Get these right and a mediocre prompt still lands close. Get them wrong and no amount of scene-setting rescues the result.
A prompt formula that removes the guesswork
Subject, setting, light, lens, mood. In roughly that order.
Say you need a hero image for a coffee subscription landing page. This is what I’d type:
“a matte ceramic mug of black coffee on a rumpled linen cloth, a few roasted beans scattered beside it, soft window light from the left, shallow depth of field, muted earth tones, editorial product photography”
Two details do most of the work there. “Black coffee” beats “coffee,” and “from the left” beats “good lighting.” Firefly’s web app has no negative prompt field, so spend your words describing what should be present rather than listing what you want gone. Once you’ve got a prompt that sings, save it. Rebuilding a good prompt from memory three days later is a genuine waste of an afternoon.
Generate in batches, then move one variable
Each run gives you four options. Pick the closest one and hit Generate Similar, then change exactly one thing. Better light. Tighter crop. A different background material. Change five things at once and you learn nothing about which one mattered.
Generative Fill is where the real time savings live
Select a region, describe what belongs there instead, generate. Three variations, usually one keeper.
A concrete example from last week: a client sent a desk photo with a black USB cable snaking across the bottom corner. Rather than reshooting, here’s the sequence.
- Open the file in Photoshop and lasso the cable with a few pixels of margin. Selections that hug the object too tightly produce visible seams.
- Prompt with the surface, not the absence: “empty light oak desk surface, soft shadow.”
- Generate, compare the three results at 100%, and check the wood grain runs in the same direction as the surrounding desk.
Generative Expand works the same way but outward. Crop past the edge of your canvas, leave the new area blank, and let Firefly invent what continues there. A 16:9 product shot becomes a 9:16 vertical in about forty seconds, and the invented edges hold up as long as you keep the extension under roughly a third of the original width. Push further and furniture starts melting. For a deeper look at what’s possible when you feed an existing photo in rather than starting blank, our guide to transforming an existing photo with an AI image generator covers the wider range of techniques.
Text Effects: brilliant for one word, useless for a sentence
Type a word, describe a material, get lettering. “Melted chocolate” on the word SWEET. “Brushed copper” on a brand name. “Moss and river stone” for an outdoors label. It’s a fast way to produce a hero graphic without opening Illustrator.
Keep it to one or two words. Kerning degrades quickly on longer strings, and the output is pixels, not editable type. You’re getting a picture of a wordmark, not a font file. If you need genuinely legible text sitting inside a generated scene, a tool designed around that problem performs better than Firefly does, which is exactly what we found when we looked at how Ideogram solved text rendering in AI images.
For Illustrator users: Generative Recolor and Text to Vector
Select existing vector artwork, open Generative Recolor, and type a direction. “Warm autumn palette.” “1980s Miami neon.” “Muted Scandinavian pastels.” Firefly returns several colourways as editable swatches, and it’s smart enough to distinguish background shapes from highlights instead of recolouring everything flat. I’ve used it to pitch four palette options to a client in under five minutes, which used to take an afternoon of manual swatch swapping.
Text to Vector generates editable paths from a description. “Minimal line icon of a mountain range” or “flat illustration of a bicycle.” Be honest about the output quality: you’ll get more anchor points than a human designer would place, and strokes often arrive as outlines rather than strokes. Budget ten minutes of path cleanup per icon. That still beats drawing from zero.
Credits, licensing and what you can legally sell
Every generation burns credits, and free accounts get a monthly allowance that disappears faster than you’d expect once you’re running four-up batches. Premium plans raise the ceiling. Enterprise plans add something more valuable: IP indemnification. Regardless of tier, use Generate Similar instead of fresh runs when you’re refining, and set your aspect ratio before your first generation.
On rights, Firefly’s image models were trained largely on Adobe Stock and licensed content, which is why Adobe can offer commercial use on outputs and back it with indemnification for enterprise customers. That’s a meaningful difference from tools trained on scraped web data, and it’s the main reason agencies adopt Firefly over cheaper alternatives. Our backgrounder on the Firefly model family and how each model is trained is worth reading before you commit to it for client work.
Downloads arrive with Content Credentials, a signed metadata tag noting the file was AI-generated. Leave it attached. Some editors strip it, some don’t, and a growing number of clients now ask the question either way.
Where Firefly loses, and what to use instead
Photorealistic faces are not its strength. Output tends toward a soft, slightly over-lit look, and skin can read as waxy under close inspection. For product, landscape and abstract work this rarely matters. For portrait-heavy campaigns, it does. Very long prompts also get muddled, and hands still go wrong occasionally.
None of that makes Firefly a bad choice, just a specific one. It’s the tool you reach for when the workflow matters as much as the raw image quality, because Photoshop, Illustrator and Express are all one click away. If you’re still shopping across the category, our roundup of AI image generators that actually deliver in 2025 breaks down where each tool wins. And if you’d rather not pay at all while you learn the fundamentals, there are free AI photo generators that produce genuinely professional results.
Five habits that quietly burn your credits
- Generating before setting the aspect ratio. You’ll just generate again later, at double the cost.
- Rewriting the whole prompt every time. Change one variable, keep the rest. Otherwise you can’t tell what improved.
- Making tiny Generative Fill selections. Two-pixel margins around an object give you a visible halo. Loose selections blend better.
- Ignoring Generate Similar. It’s the cheapest route from a decent result to a great one.
- Generating a whole new image when you needed a patch. If 80% of the frame already works, fill the other 20%.
The thread running through all five is the same. Firefly rewards tight iteration and punishes the shotgun approach, so treat it less like a vending machine and more like a conversation with someone who’s fast but needs clear direction.