Skip to content

AI Image Generation and Design · How AI image models work · lesson 2 of 18 · 8 min

Choosing the right image tool for the job

There is no single best tool

The AI image landscape changes quickly. New models launch, older ones improve, pricing and license terms change, and features like reference images or text rendering move from rare to standard. Instead of memorizing a leaderboard, evaluate tools against the needs of each brief.

As of September 2026, widely used options fall into five groups (treat any list as a snapshot and check current names and plans):

| Group | Examples | Typical strengths | |---|---|---| | Standalone generators | Midjourney, Ideogram, Recraft | Aesthetics (Midjourney), text and layout (Ideogram), vector and brand styles (Recraft) | | Assistant-built models | OpenAI GPT Image models (ChatGPT, API), Google Gemini image models (Gemini app, AI Studio, Vertex AI) | Following complex instructions, conversational editing, multi-image references | | Design-suite features | Adobe Firefly and Photoshop Generative Fill, Canva AI, Figma's AI image tools | Editing inside your layout, brand kits, team workflows | | Open-weight models | FLUX and Stable Diffusion families, run locally or via hosted services | Control, privacy, custom fine-tuning | | Specialists | Upscalers, background removers, product-shot tools | One job done very well |

Note that the lines blur: Adobe's Firefly app and Photoshop Generative Fill now let you pick partner models from other labs alongside Adobe's own Firefly models, so "which tool" and "which model" are separate decisions.

Evaluation criteria

| Criterion | Questions to ask | |---|---| | Output quality for your style | Does it do photorealism, illustration, 3D, or graphic styles well? | | Prompt adherence | Does it follow complex instructions (counts, positions, colors)? | | Text rendering | Can it produce legible words, or will you add text later? | | References and consistency | Style references, character/subject references, brand kits? | | Editing | Inpainting, outpainting, conversational edits, background removal? | | Resolution and upscaling | Native resolution, upscaler quality, aspect ratios supported | | Commercial terms | Can outputs be used commercially on your plan? Any indemnity? | | Data and privacy | Are your uploads and prompts used for training? Enterprise options? | | Safety and provenance | Content filters, content credentials (C2PA) or watermarks | | Workflow fit | Integrates with your design tools, API access, team features | | Cost | Subscription, credits per image, API pricing |

Commercial terms deserve special attention

Read the current terms of service for any tool used in client or commercial work. Key points to check:

  • Ownership/usage of outputs: many providers say you may use outputs commercially, but terms vary by plan (free tiers may differ).
  • Restrictions: prohibited uses (for example, political content, impersonation, certain sensitive categories).
  • Input rights: you must have rights to any images you upload as references.
  • Indemnity: some providers offer intellectual-property indemnities for certain paid or enterprise plans, typically with conditions. This can matter for brand clients.
  • Training on your data: whether prompts and uploads may be used to improve models, and how to opt out.
  • Visibility: on some platforms, generations are public by default unless you use a private mode.

Privacy and confidentiality

Uploading unreleased product photos, client logos or customers' faces to a tool can expose confidential information or personal data. Use tools and settings appropriate to the sensitivity: private modes, enterprise plans with data controls, or local/open-weight models for highly confidential work. Personal data laws (such as the GDPR in the UK and EU, and data protection laws in the UAE, Saudi Arabia and elsewhere) may apply to images of identifiable people.

Building a small tool stack

Most professionals use a combination rather than one tool:

Example tool stack (roles, not brands)
Ideation & moodboards:   fast generator with strong aesthetics
Brand/product work:      tool with reference images + commercial terms suited to clients
Text-heavy concepts:     model with reliable text rendering (or add text in design tool)
Editing:                 design suite with generative fill/expand + manual retouching
Upscaling:               dedicated upscaler or built-in high-res mode
Layout & typography:     your normal design tool (Figma, Canva, Adobe, Affinity…)

A tool-test protocol

Before adopting a tool for client work, run a standard test:

  1. Use the same five prompts across candidate tools: a product shot, a portrait, an illustration, a scene with text, and a complex composition.
  2. Score each output for quality, adherence, text and consistency.
  3. Test editing on one image.
  4. Read the terms and data policy; note plan requirements.
  5. Record costs per usable image (not per generation — many generations are discarded).

Worked example: a small agency choosing tools

An agency in Dubai produces social ads for restaurants and beauty brands.

  • Requirements: realistic food and product scenes, Arabic and English text (added in design tool), brand consistency, commercial safety, client confidentiality.
  • Decision: a design-suite generator with commercial-use terms for client-facing assets; a fast standalone generator for internal moodboards only; a dedicated upscaler; all typography done in their design tool; no client product photos uploaded to tools that train on user data.

Common mistakes

  • Choosing tools based on viral examples rather than your actual briefs.
  • Ignoring terms of service and plan differences.
  • Uploading confidential client material to public or training-enabled tools.
  • Measuring cost per generation instead of cost per usable image.

Hands-on: a tool scorecard you can reuse

Copy this scorecard into a spreadsheet and run it for each candidate tool, using the five-prompt protocol above. Score 1–5.

TOOL SCORECARD — brief: ______________________  date: ________
Tool / model / plan: ___________________________________________
                                  score  notes
1. Style quality for this brief    [ ]   ______________________
2. Prompt adherence (counts, positions, colors) [ ]  _________
3. Text rendering (or will we add text in layout?) [ ]  ______
4. References: style / character / product     [ ]  _________
5. Editing: inpaint, expand, instruction edits [ ]  _________
6. Resolution + upscaling for largest output   [ ]  _________
7. Commercial terms on OUR plan (quote clause) [ ]  _________
8. Data use: training on uploads? opt-out?     [ ]  _________
9. Provenance: Content Credentials / watermark [ ]  _________
10. Workflow fit: API, Figma/Adobe/Canva, team [ ]  _________
Cost per USABLE image = total spend / images approved: ______
Decision: adopt / adopt for internal only / reject — why: _____

Worked example: cost per usable image

A freelance designer in Manchester tests two tools on the same brief (illustrative numbers). Tool A costs less per generation, but only 1 in 12 outputs passes review because it mangles the product. Tool B costs more per generation, but 1 in 4 passes because it accepts the real product photo as a reference. Once review time is included, Tool B is cheaper per usable image and much faster. This is why the scorecard asks for approved images, not generations.

Before/after: the same brief in two tool types

| Brief: "Eid sale banner, 1200×628, Arabic and English headline" | Before | After | |---|---|---| | Approach | Ask a standalone generator to render the whole banner, text included | Generate the background scene only; set both headlines in the design tool with licensed Arabic and Latin fonts | | Result | Arabic letterforms broken, English headline misspelled | Crisp, correct, editable text; background reusable across sizes |

Even models that render text well should not be trusted with a second script, legal copy or prices. Put words in the layout, where you can edit and check them.

Summary

Evaluate tools against the brief using quality, adherence, text, references, editing, resolution, terms, privacy, provenance, workflow and cost. Build a small stack, run a standard test before adoption, and re-check terms regularly because they change.

Video lecture: Choosing the right image tool for the job

Lecture coming soon · 10 chapters · about 9 minutes. Read the full transcript below.

  1. Choosing the right image tool
  2. Why tool choice matters
  3. Five groups of tools
  4. Three questions, ten criteria
  5. Worked example 1: recipe card backgrounds
  6. Worked example 2: cost per usable image
  7. Watch me do it: a one-hour tool test
  8. Terms and privacy
  9. Common mistakes
  10. Recap and try this now

Lecture transcript

Choosing the right image tool

Let me ask you a question. Which AI image tool is the best? If you answered with a brand name, I have bad news. The honest answer is, best for what? A tool that makes gorgeous moody fashion images might be useless for a product shot that has to match the real bottle exactly. And a tool that nails text might have terms that don't suit your client work. In this lecture, you'll learn how to choose image tools like a professional. You'll get a set of evaluation criteria, a repeatable test protocol, and a scorecard, and you'll learn to measure the one number that actually matters: cost per usable image. By the end, you'll be able to defend your tool choices to a client in two minutes.

Why tool choice matters

Why does this matter so much? Because the market moves every few months. Models get new versions, features jump between tools, and pricing and license terms change. If you pick a tool because of a viral post, you're making a decision based on someone else's brief. Think of it like hiring. You wouldn't hire a photographer because they have a nice Instagram. You'd look at work that matches your job, check their contract, and ask about deadlines. Tools deserve the same process. And there's a business reason too. When a client asks why you used a particular tool, you want an answer that mentions their needs, their data and their rights, not just, it looked cool.

Five groups of tools

Here's the landscape as five groups, and remember, names change. First, standalone generators like Midjourney for aesthetics, Ideogram for text and layout, and Recraft for vectors and brand styles. Second, models built into assistants, like OpenAI's GPT Image models and Google's Gemini image models, which are strong at following complex instructions and conversational editing. Third, features inside design suites, like Adobe Firefly and Photoshop Generative Fill, Canva's AI tools, and Figma's image features, where the big advantage is editing right inside your layout. Fourth, open-weight model families like FLUX and Stable Diffusion, which give you control and privacy. And fifth, specialists, like upscalers and background removers. One twist. Adobe now lets you choose partner models inside Firefly and Photoshop, so which app and which model are separate decisions.

Three questions, ten criteria

Now the criteria. I group them into three questions. Can it make the image? That's style quality, prompt adherence, text rendering, references for consistency, editing features and resolution. Can I legally and safely use it? That's commercial terms on your specific plan, whether your uploads are used for training, privacy options, and provenance, meaning whether it attaches Content Credentials or a watermark. And does it fit how we work? That's API access, integrations with Figma, Adobe or Canva, team features and cost. Here's the key idea. A tool has to pass all three questions. Brilliant images with the wrong terms fail. Safe terms with images that need an hour of retouching also fail.

Worked example 1: recipe card backgrounds

Worked example one, simple. A food creator wants a flat-lay background for recipe cards. She tries two tools with the same prompt. Tool A gives beautiful, moody images, but every one has fake text on the spice jars. Tool B is a little less dramatic, but clean, and it lets her generate at exactly the aspect ratio of her card. Which wins? For this brief, Tool B. She's adding real recipe text in her design tool anyway, so clean surfaces matter more than drama. And notice what she didn't do. She didn't ask the model to render the recipe text. Even models that handle text well shouldn't be trusted with ingredients, prices or legal copy. Put words in the layout, where you can edit and check them.

Worked example 2: cost per usable image

Worked example two, a business case with illustrative numbers. Hamza is a freelance designer in Manchester, pitching product imagery for a skincare brand. Tool A costs less per generation, but only about one in twelve images passes review, because it keeps changing the bottle shape. Tool B costs more per generation, but about one in four passes, because it accepts the real product photo as a reference. Now add review time. Every rejected image costs Hamza a minute or two of checking. When he divides total spend and time by the images the client actually approves, Tool B is cheaper and much faster. That's the metric. Not cost per generation. Cost per usable image. He puts that number in his proposal, and the client understands immediately.

Watch me do it: a one-hour tool test

Watch me do it. Here's how I run a tool test in about an hour. I open a spreadsheet with the scorecard from the lesson. I write five standard prompts: a product shot, a portrait, an illustration, a scene with a short headline, and a complex composition with counts and positions. I run each prompt in each candidate tool on the plan I'd actually pay for. Then I score one to five for style, adherence, text, references, editing and resolution. Next, I try one edit on my best image, say, change the background but keep the product. Then I open the terms page and paste the exact clauses about commercial use, training on uploads and indemnity into the notes column. Finally, I calculate cost per usable image and write a one-line decision: adopt, internal only, or reject.

Terms and privacy

Let's talk about the legal and privacy side, because this is where tool choice gets serious. Read the current terms for the plan you use, because free tiers often differ from paid ones. Check whether you can use outputs commercially, whether your prompts and uploads can be used to train models, whether generations are public by default, and whether the provider offers an intellectual property indemnity, which usually comes with conditions. Then think about data. Uploading an unreleased product, a client's logo or a customer's face to the wrong tool can leak confidential information or personal data, and privacy laws like the GDPR in the UK and EU, and data protection laws in the UAE, Saudi Arabia and elsewhere, may apply. Use private modes, enterprise plans or local models for sensitive work.

Common mistakes

Common mistakes. Choosing a tool because of viral examples instead of your briefs. Assuming the terms on a free trial apply to your paid client work, or the other way around. Uploading confidential client material to tools that train on user data. Measuring cost per generation instead of cost per usable image. And trying to make one tool do everything. Most professionals run a small stack. Something fast for moodboards. Something with strong references and suitable terms for client assets. A design suite for editing and layout. And an upscaler. Finally, don't treat your decision as permanent. Put a reminder in your calendar to re-run the test every quarter, because the tool that won in spring may not win in the fall.

Recap and try this now

Let's recap. There's no single best tool, only the best tool for a brief. Evaluate with three questions: can it make the image, can we use it legally and safely, and does it fit how we work. Run the same five prompts across candidates, read the terms for your actual plan, and decide on cost per usable image. Keep words, prices and second scripts in your layout, not in the generated pixels. And build a small stack instead of betting on one tool. Here's your try this now. Pick two tools you already have access to. Run the five-prompt protocol, fill in the scorecard, and write a one-paragraph recommendation you could send to a client. You'll reuse that scorecard for years.

Key takeaways

  • Evaluate tools against each brief: can it make the image, can you use it legally and safely, and does it fit your workflow.
  • Design suites such as Firefly and Photoshop now offer partner models, so the app and the model are separate choices.
  • Read the terms for the plan you actually use — commercial rights, training on uploads, visibility and indemnities vary.
  • Measure cost per usable (approved) image, including review time, not cost per generation.
  • Keep text, prices and second scripts in the layout, and re-test your stack regularly.

Try it

Run the five-prompt tool-test protocol on two tools you have access to. Score results, read each tool's current commercial and data terms, and write a one-paragraph recommendation.