A kombucha brand wants fourteen lifestyle images for a spring campaign by Friday. There is no photographer booked, no studio, and no budget line for stock licensing. Two years ago that brief would have ended with a polite email about what’s possible. Now it ends with a Firefly window open and a moodboard on the second monitor.
Adobe Firefly has grown well past the “type a sentence, get a picture” stage that the first wave of Firefly coverage focused on. It is now a scattered set of tools across Photoshop, Illustrator, Express, and the browser app, and knowing which one to reach for at which moment is the difference between a 40-minute job and a four-hour fight with a slot machine. Here is the actual sequence I’d run for that kombucha brief.
Pick the tool before you type a prompt
Most disappointing Firefly results come from using the wrong module. Text to Image is for building a scene from nothing. Generative Fill is for fixing a scene that already mostly works. They behave very differently and cost credits differently.
- Text to Image (browser, Photoshop, Express): base compositions, moodboards, concepting.
- Generative Fill and Expand (Photoshop): removing clutter, swapping props, changing aspect ratios without re-generating the whole shot.
- Vector Recolour and Generate Vector (Illustrator): pushing a set of icons or line illustrations into a brand palette in one pass.
- Text Effects: short headline treatments, ideally one to three words, rendered as editable vector rather than flattened pixels.
- Custom models: trained on your own approved photos so output looks like your product, not a generic one.
Under the hood, Firefly draws on Adobe’s own models plus licensed partner models. Output carries Content Credentials, and the training data situation makes it one of the safer options for client work. Firefly also hosts third-party models now, part of the same realignment that reshaped OpenAI and Microsoft’s partnership terms.
Step 1: write prompts like a creative brief
“Beautiful kombucha bottle, high quality, 4k” gets you mush. A prompt that works reads like a shot list handed to a photographer. Five slots, filled in order:
- Subject: what it is, including material and finish.
- Setting: surface, background, time of day.
- Composition: angle and focal length.
- Light: source, direction, hardness.
- Palette and mood: two or three colours, one feeling.
Put together, that looks like: “matte glass bottle with an unlabelled amber liquid on a wet slate ledge, low three-quarter angle, 85mm, fog at dawn, soft rim light from the left, muted teal and warm amber palette.” Notice there is no “masterpiece” padding. Adjectives like that shift the model very little and eat prompt space you could spend on light direction.
Stop retyping style, start referencing it
Two reference features do more work than any word choice. Structure Reference holds a composition: upload a rough iPhone shot of where the product will sit and let Firefly relight and restyle it. Style Reference copies a look from an image, so a three-image moodboard of clay textures and cool daylight can be applied consistently across all fourteen assets. Build the moodboard once, reuse it all week.
Step 2: generate wide, then edit in place
Generate more than you need and resist the urge to perfect any single frame before you’ve seen the batch. Forty quick drafts at 1:1 tells you more than six careful ones at high resolution.
Generative Expand for the format juggling
The hero image needs 16:9, the story cut needs 9:16, and the paid social set needs 4:5. Instead of regenerating, open the keeper in Photoshop, expand the canvas, and let Generative Expand fill the new space. It invents believable edges, fog, tabletop, foliage, that blend with the original.
Generative Fill for the small embarrassments
A stray bottle cap in the corner. A mug that is the wrong colour for the brand. A reflection that reads as a smudge. Lasso the area, type what should be there instead (“matte black ceramic mug, same lighting”), and run it. Three or four attempts is normal. If the fifth still looks wrong, change the selection shape rather than the words. Selection geometry matters more than prompt phrasing in Fill work.
Step 3: lock the brand with vectors and text effects
Photographic assets get you the campaign. Vector assets get you the system around it. In Illustrator, Vector Recolour takes an existing icon set and maps it onto your brand palette in a single pass, which is genuinely faster than doing it swatch by swatch. Generate Vector handles the blank-page problem when you need a simple line illustration and have no idea where to start.
For headline treatments, Text Effects works best with short phrases. Long sentences tend to come back with wobbly letterforms, and you’ll spend more time cleaning kerning than you saved. Use it for three words, then rebuild anything longer as live type in your brand font.
A realistic 40-minute build for that kombucha brief
- 0-6 min: assemble a three-image style reference board and one structure reference.
- 6-16 min: batch generate roughly 40 drafts, cull hard to 12 keepers.
- 16-28 min: Generative Expand each keeper into 16:9, 4:5, and 9:16 versions.
- 28-36 min: Generative Fill to fix props, reflections, and edges. Export at full size.
- 36-40 min: name files properly, upload to a shared review link, and flag the two frames that need a human opinion.
That last step matters more than it sounds. Approvals move fastest when you keep assets in the same thread-based flow teams already use to run code reviews in Slack. It’s the same principle pushing studios toward scaling creative output without adding headcount: the bottleneck was never the idea, it was the handoff.
When Firefly fights back
The prompt gets blocked
Firefly refuses a fair amount: named people, brand logos, some celebrity-adjacent wording. Rewrite the specific term as a description. “A 1990s aerobics instructor with a high ponytail and leg warmers” gets through where a name does not, and it usually gives you more control anyway.
Skin looks like plastic and hands look wrong
Plastic skin is usually a light problem. Add a specific imperfect light source and a texture word: “window light with soft shadow falloff, visible pores, slight sheen.” For hands, generate them last with Generative Fill inside an existing frame rather than hoping the full composition gets them right.
Every result looks the same
You’re probably reusing one seed. Change the seed, change the focal length, and change the time of day. Three variables produce far more visual range than thirty new adjectives.
Credits, rights, and the practical housekeeping
Every generation and every fill spends generative credits. A free tier gives you a modest monthly allowance, and most paid Creative Cloud plans include a larger pool, but a heavy batch day can eat through it. Cull early and work at draft resolution until you’ve picked a keeper.
Output is generally safe for commercial use because of how the models were trained, and Content Credentials travel with exported files to show what was AI-generated. Two habits worth keeping: never present an AI-generated image as a real photo of a real customer, and keep a folder of source references so you can prove where a look came from if a client asks in six months.
If someone on the account doesn’t have a Creative Cloud seat, they can still finish retouching in free AI retouching tools like Photopea without breaking the versions you exported.
Build the prompt library once and stop rewriting it
The compounding win isn’t any single image. It’s saving the prompt, style reference, and settings for every asset that worked into a shared document, so the next campaign starts from twelve proven recipes instead of a blank box. Even better, take your ten best approved shots and train a custom model on them. From that point on, “product shot on slate ledge, dawn fog” returns results that already look like your brand rather than a stranger’s.
Do that and the interesting part of the work shifts back where it belongs: deciding which fourteen images the campaign actually needs, and why.