One Tool Per Image Job

AI image generation is not one tool

A marketer pays for Midjourney, a ChatGPT Plus plan that includes DALL-E, and an Adobe Creative Cloud subscription that includes Firefly. All three live one tab away from each other. The campaign brief gets prompted into whichever window is open. By the end of the quarter the bill is $80 a month and the assets do not look like one brand.

A different marketer generates a logo in Midjourney for a Series A launch, gets a beautiful mark, and ships it. Three weeks later the mark drifts at small sizes, fails in monochrome, has no clean vector, and reads as someone else's brand. The first agency call costs more than the year of Midjourney.

A third marketer asks DALL-E for a product hero shot. The output is photoreal until the second glance: the bottle has six fingers wrapped around it, the label spells the product name wrong, and the reflection on the desk runs the wrong direction. The retoucher rebuilds the shot from scratch in Photoshop.

Three marketers, three failures, one category mistake. AI image generation has not become one tool; it has become a stack of single-job tools that share a marketing label.

The same tool that nails an expressive editorial illustration is the wrong source for a trademarkable logo. The same model that ships a clean text-heavy ad variant produces six-fingered hands on a product mockup. The vector-first generator that produces beautiful icons cannot retouch a real product photograph. The mistake is not the tool. The mistake is the absence of a per-job routing map.

The right question is not which AI image generator is best. The model leaderboard answers that question with a different model every quarter, and the answer keeps changing. The right question is sharper: what image job am I doing today, what tool earns that job, and where does the designer or photographer have to step back in regardless of which generator produced the file?

Seven image jobs matter for a working marketer or designer: brand mark and logo concept, editorial illustration, hero and web imagery, campaign and ad creative, product mockup in context, icon and pictogram, and photo edit or retouching. Each has a different tool that earns the slot today, a different ceiling (commercial-use safety, brand fidelity, photoreal accuracy, typography reliability), and a different return-point for the human designer or photographer.

The seven image jobs

The image job decides the tool. Below, each job gets the same treatment: the tool that earns the slot today, what ships clean, where the ceiling sits, and the action plan for week one. The action plans assume you are producing that job from scratch this quarter; if a workflow is already in place and working, the answer is to keep it and skip the section.

Brand mark and logo concept

Logo direction passes, symbol exploration, lockup variations, wordmark sketches. The slot belongs to Recraft. The product is explicitly built around vector generation, editable vector graphics, design-ready typography, and logo and icon use cases. Recraft free credits produce public, non-commercial output owned by Recraft; paid Basic at $12 per month or Pro at $20 unlocks commercial ownership and private generation. The work product is not "the logo."

What ships clean: symbol forms, lockup variations, pictogram options, and rough wordmarks for a designer to refine into the final mark. The ceiling appears at the exact point where a real logo begins: distinctiveness, reproducibility, legal clearance, vector precision, and system behavior across sizes. Midjourney can make a beautiful fake brand mark, but its terms disclaim guarantees of title and non-infringement. Ideogram can generate strong lettering, but its docs say it does not support SVG or vector formats. Canva's licensing guide is the bluntest: Canva library content should not be used in trademarks except fonts, basic shapes, and lines. Treat any generator output as direction, not a final mark.

If you start this week, produce twelve to twenty Recraft directions for one brand challenge and hand them to a designer with a written brief. The designer owns reduction, geometry, grid, vector rebuild, lockups, color, monochrome behavior, small-size testing, trademark search coordination, and usage rules. Ship the brief and the directions to the designer by Friday. The finished logo comes later, after the designer redraws, simplifies, and clears trademark risk.

Editorial illustration

Article heroes, in-line illustration, magazine-style spreads, expressive editorial visuals. The slot belongs to Midjourney. The job is not only prompt accuracy; it is taste, composition, lighting, atmosphere, and the ability to create a visual world that does not feel like a template. Midjourney's current controls cover aspect ratio, style reference, omni reference, personalization, raw mode, stylize, seed, and versioning, which is the surface area that editorial illustration actually needs.

What ships clean: an expressive editorial image with art direction, lighting, and atmosphere. The ceiling appears at brand persistence and likeness. Editorial illustration fails when a publication needs legally defensible likeness, exact symbolic meaning, or a repeatable house style over months. Midjourney can hold mood. It does not become an art director. OpenAI Images is the strong adjacent fit when the illustration must include readable labels, infographic elements, or iterative non-designer collaboration. Ideogram is adjacent when typography is the hero, though its own docs warn that spelling errors still occur.

If you start this week, generate one editorial illustration set for a real upcoming article: hero, two supporting visuals, one decorative element. Crop each into the publication's actual layout and check for caption legibility, focal-point readability, and brand fit. Ship the set by Friday with art-director review notes attached. The illustration set is approved or sent back for revision before publication, not after.

Hero and web imagery

Landing-page heroes, section banners, large-format web imagery, campaign hero shots. The slot belongs to Midjourney for expressive heroes and Adobe Photoshop with Firefly for production finishing. The reliable route is two-step: generate the hero world in Midjourney, then finish in Photoshop or the site design system. Midjourney's default V7 1:1 upscaled output is 2,048 by 2,048 pixels, enough for screens but not automatically enough for large print without further upscaling.

What ships clean: a screen-ready hero finished through the site design system. The ceiling appears when teams treat the generated frame as a brand system. A one-off hero does not solve recurring art direction across pages, products, and campaigns. It also fails on product UI, customer scenes, and faces needing consent or representation review. OpenAI Images is the better choice when the hero has to follow an explicit product narrative or include readable copy. Adobe is the better finishing lane because Photoshop bundles Generative Fill, Harmonize, Generative Expand, Generative Upscale, and distraction and reflection removal.

If you start this week, generate one hero in Midjourney, drop it into Photoshop, and finish it through the site design system: crop behavior, responsive art direction, accessibility contrast, compression, alt text. Ship the finished hero into the actual page layout by Friday. The hero is reviewed in the layout, not in the generator.

Campaign and ad creative

Paid social units, ad variants, campaign visuals, lead-magnet covers, promotional graphics. The slot belongs to OpenAI Images through ChatGPT or the API. Campaign work is iterative, copy-sensitive, and version-heavy. GPT Image 2's own guide names it as the recommended default for production workflows and calls out text-heavy images, photorealism, compositing, identity-sensitive edits, and fewer retries. The API distinguishes single-image generation from conversational, multi-step image flows through the Responses API, which is how a marketer actually works when trying variants.

What ships clean: text-heavy ad variants, social units across sizes, and rapid iteration when copy and layout shift between rounds. The ceiling appears at claims, regulated categories, brand safety, demographic representation, and logos. OpenAI's terms place responsibility for content on the user, and organizational users indemnify OpenAI against claims arising from use. Canva's AI terms similarly put responsibility for inputs and outputs on the user. Adobe is the alternative when campaign variants need Adobe workflow, Firefly Boards, Express, or qualifying indemnity. Ideogram is the typographic specialist; Recraft is the graphic-system specialist; Midjourney is the visual-territory specialist.

If you start this week, generate three ad directions in OpenAI Images, each with three social-size variants and one text-heavy version. Assemble them into Canva or Adobe Express, route through the actual paid-media review, and ship the campaign kit by Friday. The unit shipped is a campaign kit, not a generated image.

Product mockup in context

Packaging mockups, product-in-use scenes, ecommerce hero shots, marketing-page product visuals. The slot belongs to Adobe Photoshop with Firefly. Product mockup is compositing an existing object, preserving label geometry, matching shadows, matching lens behavior, and producing a file that can be corrected. Photoshop is the production home because Adobe documents the exact features for the job: Generative Fill with model selection and reference images, Harmonize for blending objects into backgrounds, Generate Background, Generative Expand, and Generative Upscale.

What ships clean: a real product composited into a generated scene with matched shadows and reflections, layered for retoucher review. The ceiling appears at exact SKUs and labels. A generator can invent a plausible bottle. It cannot be trusted to preserve the real cap, legal copy, nutrition facts, barcode, logo spacing, or reflections unless a human verifies and retouches. Recraft is adjacent for early packaging exploration or stylized ecommerce scenes. Krea is adjacent for ecommerce enhancement and upscaling. OpenAI Images is useful for concepting the scene, but product truth remains a retouching problem.

If you start this week, take one real product and produce two mockups: one stylized brand-world scene and one realistic in-use shot. Use Firefly to generate the background and Harmonize to blend the product, then have a retoucher verify label fidelity and reflection direction. Ship the verified mockups by Friday. AI generates the scene; the retoucher signs off on the truth.

Icon and pictogram

UI icon sets, navigation pictograms, marketing icons, system icons, slide-deck pictograms. The slot belongs to Recraft. Icon work is vector work, and Recraft foregrounds vector generation, complex editable vector graphics, consistent styles, and icons explicitly. Paid Recraft plans give commercial ownership and private generation; the free plan output is public, owned by Recraft, and non-commercial. Adobe Illustrator is the finishing environment for production icon systems, especially when icons need grid logic, stroke consistency, naming, exports, and design-system integration.

What ships clean: a family of editable vector icons in a consistent style, exported as SVG for finishing in Illustrator. The ceiling appears at optical consistency and system rules. AI can make a set of similar-looking pictograms. It does not guarantee stroke behavior, corner radius, semantic coverage, or accessibility. The more icons in the set, the more visible the drift becomes. Ideogram can help when an icon or pictogram includes lettering, but the lack of SVG output makes it an exploration tool rather than the production source.

If you start this week, generate one icon family in Recraft with twelve to twenty pictograms covering a real UI surface: navigation, system actions, status, settings. Export to SVG, drop into Illustrator, and hand to a designer for grid normalization, stroke alignment, and naming. Ship the production-ready icon family by Friday. The icon system is approved against the design system, not against the generator's preview.

Photo edit or retouching

Background removal, object removal, scene expansion, color correction, compositing, upscaling. The slot belongs to Adobe Photoshop with Firefly. This is the easiest routing call. Adobe's Photoshop AI page lists the exact task map: Generative Fill, Reference Image, Harmonize, Generative Expand, Generative Upscale, Distraction Removal, Reflection Removal, Generate Background, and object compositing workflows.

What ships clean: a retouched photo with wires removed, distractions cleaned, background extended, and resolution upscaled, layered for review. The ceiling appears at client standards. Skin, jewelry, food, apparel texture, architecture, and product surfaces are unforgiving. AI can remove a wire. It can also hallucinate a wrong seam, wrong skin texture, wrong label, or wrong reflection. Krea is adjacent for upscaling and restoration, particularly when a marketer has low-resolution assets, compressed ecommerce images, or social files that need to survive a larger placement. OpenAI Images and Google Gemini are adjacent for conversational edits where the edit is conceptual rather than precision retouching.

If you start this week, take one real photo with a known retouching problem and run it through Photoshop's AI task map: remove the distraction, expand the canvas, harmonize a generated background, upscale if needed. Have a retoucher review every change for label fidelity, texture accuracy, and reflection logic. Ship the retouched photo by Friday with the change log attached. The retouch passes only when the human signed off, not when the tool finished.

Where the designer or photographer steps back in regardless of tool

The boundary is not "when the output looks bad." The boundary is when the asset carries brand, legal, or trust risk. AI image generators produce default-plausible visuals across every job. Default-plausible is fine for concept exploration and rough campaign variants. It fails at the surfaces where brand fidelity, commercial-use posture, photographic truth, or legal exposure is the differentiator.

The designer owns brand-system fidelity. Logo finals, brand mark systems, typography pairings, color palette decisions, and asset families that have to look like one brand across six surfaces and three regions. A generator can produce a hundred plausible directions; the designer chooses, refines, and codifies the rules that make the system durable.

Trademark, likeness, and indemnification stay with legal and rights review. Adobe has the most explicit commercial-use posture, but its own product description ties indemnification to qualifying plans, eligible Firefly features, eligible surfaces, and contract terms; partner-model supplemental coverage excludes beta, preview, and pre-GA features. "Open weights" is not commercial safety; Stability AI's Community License has a $1 million annual revenue threshold and Black Forest Labs licenses Flux variants differently by model and use case.

The photographer or retoucher owns photographic truth. Product hero shots, beauty, fashion, food, architecture, medical, finance, legal, or any brand asset reused for months. AI can remove a wire; a retoucher verifies it did not hallucinate the wrong reflection, wrong texture, wrong label, or wrong skin tone. Generators do not produce the final truth layer for packaging or brand campaigns.

The marketer owns claims and audience. AI generates variants; it does not approve the campaign. Demographic representation, regulated-industry claims, audience targeting, and channel fit are review decisions, not generation decisions.

Cost calculus and coexistence

Free tiers are useful for evaluation, not for running the team's image work. Commercial rights, premium credits, private generation, vector outputs, and team controls sit on paid plans across the stack. The first paid stack that earns its keep for a solo marketer is small and usually already paid for: ChatGPT Plus at $20 per month plus whatever Adobe plan you already have, often Creative Cloud Pro at around $70 per month including Photoshop, Illustrator, Firefly, and 4,000 premium credits. That is two subscriptions totaling around $90 per month, and it covers campaign creative, hero finishing, product mockup, and photo retouching without adding a third image generator.

The mid-paid stack sits between $100 and $200 per month: ChatGPT Business, Creative Cloud Pro, Recraft Pro at $20 for logos and icons, and Midjourney Standard at $30 for editorial illustration. This is already enough for most small marketing teams unless image volume is high or specialty work is constant. The mid stack earns its keep when vector work is recurring and editorial illustration is published weekly, not when leadership likes the model leaderboard.

The team tier above $200 per month adds Canva Business at $20 per seat for campaign assembly, Recraft Teams at $30 per seat, and Midjourney Pro at $60 only where privacy, commercial revenue threshold, or visual volume justifies it.

The enterprise tier above $500 per month is the one to slow down on: Krea Business at $200 per month, dedicated API contracts, Adobe enterprise contracts for indemnification, and legal review of every model's training, indemnity, and data-use terms. At that spend, compare the stack against actual designer or retoucher hours saved.

Adobe is not a separate image-generation subscription decision if you already pay for Creative Cloud. Firefly is bundled; the line item is the Adobe plan. Buy a standalone Firefly plan only when premium credits run out and the team does not need other Adobe apps. Conversely, ChatGPT Plus is rarely a "buy this for image generation" decision because most teams already have it for other reasons. The default move is to use what is paid for first, and only add specialist subscriptions when one job is recurring enough to justify the line.

Pitfalls and anti-patterns

Paying for three generators that overlap on the same image

Midjourney plus DALL-E (via ChatGPT) plus Firefly (via Adobe) is usually three tools competing for the same campaign visual. Unless each is producing a different job (illustration, campaign, retouching), the second and third subscriptions are paying for variants of the same prompt in a different keybinding. Map by job; cancel by overlap.

Treating Midjourney as a logo generator

The first mark looks plausible. The system around it fails: dark mode, monochrome, small-size legibility, vector precision, trademark distinctiveness. Midjourney's own terms disclaim non-infringement guarantees, and its raster output is not a vector file. Use Midjourney to explore visual territory; use Recraft for vector concepts; hand both to a designer for the final mark.

Calling open weights commercially safe

"It is open source" is not a commercial-use posture. Stability AI's Community License has a $1 million annual revenue threshold, and Black Forest Labs licenses Flux variants differently by model and use case. FLUX.1 dev model use is licensed for non-commercial and non-production purposes even though outputs may sometimes be used commercially under restrictions. Read the license per model before shipping the asset.

Using AI image generation to bypass brand discipline

Faster image generation creates faster brand drift when no one owns the system. A team can produce twenty variants in an hour; without a brand owner, twenty variants ship and the brand fragments. Name an owner before scaling generation.

What to validate before paying for the stack

• Map the image jobs the team actually produces this quarter. If logo work is not recurring, do not pay for Recraft; if editorial illustration is not weekly, do not pay for Midjourney.
• Confirm the commercial-use posture per tool against the team's actual revenue and use case. Adobe's indemnification is plan- and feature-specific; Midjourney requires Pro or Mega above $1 million revenue; open-weights licenses vary by model and distribution path.
• Confirm vector versus raster output for the jobs that need it. Recraft and Illustrator produce real vectors; Midjourney and DALL-E produce raster files that must be redrawn for logos and icons.
• Confirm training-data provenance for any asset that will face public review. Adobe Firefly's first model was trained on licensed Adobe Stock and public-domain content; partner models, OpenAI, Gemini, Flux variants, and Stable Diffusion derivatives have different provenance stories that should be checked per model.
• Confirm someone on the team owns the brand decision. AI image generation only matters if a human designer or marketer signs off on the asset, routes it into the brand system, and replaces it when the system changes.

One tool per image job

One tool per image job. The subscription that earns its keep is the one that owns a row in the matrix.

A team that buys by the model leaderboard buys overlap. A team that buys by image job buys leverage. Vendors will rename and reprice every quarter; the image jobs will not.