How to Use Photoshop AI: A Practical Walkthrough

Why Photoshop AI Feels Overwhelming (and Why It's Worth Learning)

Open Photoshop today and you'll find not one AI feature but a layered ecosystem: Generative Fill and Expand powered by Firefly Image 3, a conversational AI Assistant in public beta, precision tools like AI Markup and Harmonize, and an emerging agentic layer with pre-built "Creative Skills." Just keeping track of what shipped in the July 2025, October 2025, and 2026 updates is a job in itself.

The confusion is understandable — Adobe has been shipping AI features faster than most users can absorb them. The late-October 2025 update alone added Harmonize, improved Generative Fill, and AI-powered object selection, while the AI Assistant now lets you edit by describing changes in natural language, with voice input supported in the Photoshop app.

Here's why it's worth pushing through the learning curve: the adoption data is emphatic. According to an October 2025 Harris Poll of 16,000+ creators, 86% of global creators now use generative AI, and 94% of creative professionals say they produce content faster with it. Roughly 81% credit AI with enabling content that would otherwise be impossible.

That said, go in with clear eyes about the constraints:

• AI credits add up quickly — a hands-on reviewer warns that "Adobe credits aren't cheap," and the free render credit offer expired December 1, 2025
• The free AI Assistant tier includes only 20 free generations, though paid subscribers currently get unlimited generations
• Features are in flux — the AI Assistant and AI Markup are public betas, so availability and limits may shift between updates
• Vendor claims like "amazing results within seconds" deserve skepticism; independent testers offer more measured assessments

The payoff for learning the ecosystem is real, though. Practitioners who've integrated these tools report meaningful speed gains — one reviewer notes that Harmonize "has truly sped up my AI production workflow and produces much higher quality results." The consensus among educators and working designers is that AI tools are best understood as workflow accelerators, not replacements for craft.

And the skills trend is bigger than Photoshop. Adobe's new Firefly AI Assistant ships with customizable "Creative Skills" that orchestrate multi-step workflows across apps — a signal of where all AI tooling is headed. The same question applies everywhere: which skills actually help, and which are risky? Published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building a free directory that security-scans, benchmarks, and verifies agent skills before you rely on them. Learning Photoshop AI well means learning to stack the right capabilities — and to verify what you stack.

The Core Workflow: Generate in Firefly, Refine in Photoshop

If you want predictable, professional results from Photoshop's AI features, the pattern that works isn't a single magic tool — it's a two-app pipeline. Adobe's own documentation recommends generating images in Firefly and then manually refining them in Photoshop, and practitioners have turned that advice into a reliable production workflow.

The division of labor is straightforward. As one educational guide puts it, Firefly is ideal for creating — start with words and make a full image — while Photoshop is ideal for compositing, using text prompts and selections to build up an image in layers. Firefly runs on the Firefly Image 3 model, which powers Generative Fill, Generative Expand, and Generate Similar in Photoshop as of the July 2025 update.

For compositing work, video producer Jeff Foster's tested pipeline has become a go-to template:

• Extract your subject with Select Subject for a fast, accurate initial cutout.
• Refine the edges in Select & Mask so the subject sits naturally in any scene.
• Keep the subject on its own isolated layer so you can reuse the same character across different backgrounds and wardrobe changes.
• Composite the subject over an AI-generated background created in Firefly.
• Click Harmonize to match lighting and color between layers automatically.

Foster reports that Harmonize has genuinely sped up his AI production workflow while producing higher-quality results — a meaningful claim given that Firefly in Photoshop returns three candidate results per generation, so quality control still happens on your side of the screen. The same principle applies to prompting: Photoshop's AI blends local selections more effectively than trying to generate everything at once, so work in smaller, targeted passes rather than one giant prompt.

For business users, there's a compliance reason to prefer this stack. Firefly was trained on licensed Adobe Stock images and public domain content, and Adobe automatically attaches Content Credentials to Firefly-generated output — provenance you can show clients. As one design educator puts it, commercially safe AI is a requirement for professional client work, not just a feature. That framing matters: with 86% of creators now using generative AI, differentiation comes from workflow discipline, not tool access.

One caution before you build this into your pipeline: credits. Independent reviewers warn that Adobe credits aren't cheap, and render limits add up quickly when you're iterating. Budget your generations, lean on the refine-in-Photoshop steps (which don't consume credits the way re-rendering does), and treat AI output as raw material rather than finished work. Once the creative pipeline is running, agent skills can automate the surrounding steps — batching, file naming, asset delivery — and SkillzBench is building a free directory where you can browse and benchmark those skills before wiring them into your agents.

Step-by-Step: Using Generative Fill, Expand, and the AI Assistant

The fastest way to learn Photoshop AI isn't by clicking through menus — it's by building a scene from scratch, one selection at a time. An educational walkthrough from BMCC's media arts program shows how beginners can composite an entire image using nothing but selections and text prompts.

Start with a new document, then grab the Rectangular Marquee tool and select an area. When you release the selection, the Contextual Task Bar appears with the Generative Fill option. Type a prompt describing the background you want, and Firefly Image 3 — the model behind Generative Fill and Generative Expand as of the July 2025 update — returns three variations to choose from per generation.

The guide's key insight: build elements with smaller selections one at a time rather than trying to generate everything at once. As the BMCC guide puts it, "Photoshop's AI blends local selections more effectively than trying to generate everything at once." Structure each prompt as Subject → Details → Style/Mood → Composition, and remember that the less guidance you give the model, the more your image will look like the average of whatever nouns you put in the prompt.

For the newest layer of the workflow, the AI Assistant in public beta lets you edit conversationally — removing distractions, changing backgrounds, or adjusting color by describing the change in plain language. It works in two modes:

• Automatic mode applies your described edits directly to the image.
• Guided step-by-step mode walks you through each change, which is ideal while you're still learning the tools.
• AI Markup, also in beta on Photoshop web, lets you draw directly on the image and add a prompt, controlling exactly where changes happen.
• The free version includes 20 free generations, while paid Photoshop web and mobile subscribers get unlimited generations through April 9.

One practical caution for business users: AI credits add up quickly — a hands-on reviewer warns that "Adobe credits aren't cheap," so budget your generations deliberately. And because Firefly was trained on licensed Adobe Stock and public-domain content, with Content Credentials attached automatically, the output is suited to commercial client work.

This selection-and-prompt approach is itself a skill — and it's exactly the kind of repeatable workflow that agent skills are built to automate. SkillzBench is building a free directory where you can browse, security-scan, and benchmark agent skills, plus a free MCP connector to wire them into your agents; join the waitlist to be first in line when it launches.

Prompting Skills That Separate Good Results from Generic Ones

The difference between a striking composite and a forgettable one rarely comes down to the model — it comes down to the words you feed it. Photoshop's AI is remarkably capable, but as an educational prompting guide from BMCC puts it, "the less guidance you give the model, the more your image will look like the average of whatever nouns you put in the prompt." Generic prompts produce average images.

Adobe's own prompt-writing documentation lays out four principles: be specific, descriptive, original, and empathetic — meaning you should include emotion words, not just objects. Prompts should run at least three words, and you should skip meta-instructions like "generate" or "create." When results disappoint, reword and iterate rather than starting over.

A dependable structure keeps prompts from collapsing into noun soup. The BMCC guide recommends building prompts in layers:

• Subject — what the image is actually about
• Details — textures, colors, context, environment
• Style and medium — photographic, painted, illustrated, cinematic
• Lighting and mood — golden hour, moody, soft, dramatic
• Camera and composition — wide shot, close-up, low angle, depth of field

So instead of "a beach," you write "a windswept rocky beach at dusk, long exposure, moody blue tones, shot from a low angle with distant cliffs in soft focus." Same nouns, radically different output. That specificity is what separates work that looks intentional from work that looks assembled.

The second skill is less obvious: work small. The BMCC guide highlights that Photoshop's AI blends local selections more effectively than trying to generate everything at once. Multiple smaller prompts — one selection for the sky, another for the foreground, a third for a single object — consistently beat a single large generation. Each generation also returns three results to choose from, according to a hands-on practitioner review, so breaking work into pieces multiplies your options without inflating your prompt complexity.

There's a practical cost angle, too. The same reviewer warns that "Adobe credits aren't cheap," and Adobe's free render credit offer expired in December 2025. Disciplined prompting isn't just about quality — fewer wasted generations means fewer credits burned chasing vague results.

Prompting well is becoming a transferable craft, not a Photoshop-only trick. Adobe is moving toward agentic workflows with its Firefly AI Assistant and pre-built "Creative Skills" for multi-step tasks, and the same instruction-package pattern — SKILL.md files for coding and chat agents — is spreading across AI tooling. SkillzBench is building a free directory and MCP connector for exactly that: discovering, security-scanning, and benchmarking agent skills, since published research found roughly 26% carry at least one vulnerability. If you're prompting AI tools professionally, testing what you plug into your workflow matters as much as the prompts themselves.

Managing Costs, Credits, and the Agentic Future of Creative Work

Every Generative Fill click costs credits, and those credits compound faster than most creators expect. Hands-on reviewer Jeff Foster puts it bluntly: "Adobe credits aren't cheap," noting that AI render limits and credit purchases add up quickly, especially now that Adobe's free render credit offer has expired (ProVideo Coalition).

The current pricing landscape gives you several entry points. The Photography plan runs $19.99/month, a single-app subscription costs $55/month (or $49.99 prepaid annually), and the mobile/web plan sits at $7.99/month (SQ Magazine). The free AI Assistant tier includes 20 free generations, while paid Photoshop web and mobile subscribers get unlimited generations through April 9 (Adobe's blog). Budget accordingly before you start iterating on prompts.

Smart workflow habits stretch those credits further:

• Keep subjects on isolated layers so you can reuse the same character across different backgrounds and wardrobe changes without re-generating.
• Use multiple smaller selections with targeted prompts rather than one large generation — Photoshop's AI blends local selections more effectively.
• Refine your prompt before re-running it; rewording is free, re-generating is not.
• Extract subjects with Select Subject, refine with Select & Mask, then use Harmonize to match lighting — Foster reports this "has truly sped up my AI production workflow."

That last habit matters beyond cost. Firefly's Generate Similar and Generative Fill run on the Firefly Image 3 model, and Adobe attaches Content Credentials automatically to Firefly-generated content — a meaningful safeguard for client work where provenance questions increasingly come up.

The bigger shift on the horizon is agentic. Adobe's Firefly AI Assistant, now in public beta, orchestrates multi-step workflows across Creative Cloud apps from natural-language descriptions, complete with a library of pre-built "Creative Skills" — things like retouching portraits with consistent presets — that creators can customize or extend. Adobe is even bridging this into Anthropic's Claude, letting creators conceptualize there and execute in Firefly.

This "skills" pattern is spreading across the entire AI toolchain, not just Adobe's. Agent skills — instruction packages like SKILL.md files for Claude Code, Codex, Cursor, and Gemini CLI — can automate the rest of a creative workflow: batch processing, file organization, prompt iteration. The catch is quality control. Published research found roughly 26% of community-published skills carry at least one vulnerability, and nobody has time to audit them all manually.

That's the problem SkillzBench is building for. Its free, independent directory will continuously crawl and verify skills from GitHub, security-scan every version, and benchmark each one in a sandbox — publishing a Lift score showing exactly how much a skill improves agent performance. A free MCP connector will let you equip any agent with only the skills you choose. It's pre-launch, but the waitlist is live if you want the agentic layer of your creative workflow to be both powerful and safe.

Use the credits deliberately, keep your layers isolated, and let skills handle the repetitive parts.