What do you like best about Adobe Firefly?
Generative Fill in Photoshop is the part I touch most days, and it is the feature that moved Firefly from "interesting" to "in the workflow." I am already in a file retouching a client photo when I realize the background needs to run wider for a banner crop, or there is a stray cable or exit sign that has to go. I make a rough selection, type what I want, and it comes back on its own layer with a mask, so I can blend it, dial back the opacity, or throw it out without touching the original pixels. That last part matters more than the generation itself. It behaves like a proper Photoshop layer, not a magic button that rewrites my file, so it slots into the way I already work rather than asking me to change it.
The reason it is in paid client work at all, though, is the commercial-safety story, and that is not marketing to me, it is the deciding factor. Adobe's own image and video models are trained on licensed Adobe Stock and public domain material, so when I hand a deliverable to a client and they ask where an asset came from, I have an answer that holds up. Content Credentials get attached to fully AI-generated images as a kind of provenance label, which means the file carries its own record of how it was made. For a studio that invoices for this work, that is the difference between being able to use generative AI in a deliverable and having to keep it out of anything that leaves the building.
The standalone Firefly app has become where I do my thinking before I open Photoshop. Text to Image with the model picker, and Boards as a shared canvas, let me put a brief in front of the team as a wall of generated directions instead of a paragraph of description. We drop references and variations into one board, react to them, and settle on a direction together before anyone commits real hours. I will be upfront that I rarely ship straight out of the app. It is an ideation surface for me, and the finishing still happens in the desktop tools, but for getting from a vague client ask to an agreed look, it is faster than the moodboard process it replaced.
Generative Expand earns its place every time a campaign needs the same image in five shapes. A shot framed vertical for a story has to become a wide hero for the site and a square for the feed, and historically that meant re-cropping into compromises or asking for another shoot. Now I extend the canvas and let it fill the new space in a way that matches the existing scene, then clean up the seam by hand if it needs it. It is not flawless on busy backgrounds, but on skies, walls, and plain gradients it is quick and convincing.
Having the partner models inside the same app is a genuinely good decision and one I did not expect Adobe to make. From a single model dropdown I can pick Adobe's own model when the work has to be commercially clean, FLUX.2 when I need photoreal detail and better text, or one of the Google and OpenAI image models when I want a different aesthetic, without juggling five browser tabs and five separate subscriptions. The practical effect is that I stop thinking about which service to log into and start thinking about which model suits the job, which is the right way round.
Generative Recolor in Illustrator is a smaller thing I reach for during brand work. Feed it a vector pattern or a set of icons, describe a palette, and it gives me variations to react to in seconds. It does not replace deciding on the palette myself, but as a way to see a brand mark across ten color directions before a client meeting, it saves me the tedium of recoloring each one by hand.
For quick branded headline treatments I lean on Text Effects, when a layout needs a piece of display type with some texture or material to it. I would not build a final logo this way, but for a campaign headline or a social graphic that needs a bit of typographic character without commissioning a custom illustration, describing the look and getting usable variants back is faster than building the effect by hand.
Prompt to Edit is the feature I have warmed to more slowly. Describing a change in plain language instead of building a precise mask felt imprecise at first, and for fine work it still is. For the broad moves though, swap this background, take that object out, shift the time of day, it is quicker to type the instruction than to set up the selection, and it has been getting more reliable with each model update. Review collected by and hosted on G2.com.
What do you dislike about Adobe Firefly?
The consistency is the thing I bump into most. For a clean, specific brief, especially anything with people in it, I am regularly generating the same prompt two or three times before I get something usable, and faces and hands still come out wrong often enough that I do not trust it for a hero image where a person is the subject. My settled workaround is to keep Firefly on the jobs it is good at, backgrounds, objects, textures, expansions, and to use real photography or hand retouching whenever a recognizable human face is the point of the shot. It has improved on this over the past year, but it is not where I would call it dependable.
The credit system is the other recurring friction, and it is as much a clarity problem as a cost one. There is a distinction between standard and premium generations, video burns through credits far faster than images, and the plans and promotions keep shifting underneath me, so telling a client what a given piece of generative work actually "costs" in credits is harder than it should be. The way I manage it is to do all my exploration and dead ends on the cheaper standard generations, and only switch to the expensive premium model or a high-resolution render once the composition is locked. It works, but it means I am always keeping half an eye on a credit counter, which is not where I want my attention during creative work.
Text inside generated images is still not something I rely on. It has clearly got better, and the newer partner models render type far more cleanly than the early Firefly model did, but it is inconsistent enough that I would never let generated lettering go out in a client deliverable. Anything with real copy on it, I still set in InDesign or Illustrator over the top, and treat the generated image as a background plate only.
The experience is noticeably stronger inside the Adobe ecosystem than outside it, which is fine when the whole job lives in Creative Cloud and less fine when it does not. If a client works in a non-Adobe stack, or I need to get assets into a tool that is not Photoshop, Illustrator, or Express, the handoff is more manual and the tight in-app behavior I rely on falls away. There is an API for connecting it elsewhere, but that is a developer's answer, not something I am wiring up between design tasks.
Video is the part I am most cautious about recommending. The tooling has come a long way, the camera controls and the browser-based editor are real, and for previsualization and storyboards it is already useful. But the clip lengths are short, the output is good rather than finished, and for anything that has to stand next to professionally shot footage it is not there yet. I use the video model the way I use a sketch, to test a concept or block out a sequence, and then the actual production happens in proper video tools. Treated as a previz aid it is handy. Treated as a finishing tool it would disappoint. Review collected by and hosted on G2.com.
