Independent field review

Image generators, judged beyond the demo.

Quality matters. So do provenance, licensing, control, labor, and what happens when a convincing image is detached from its origin.

Last reviewed August 19, 2026
Fictional synthetic portraits arranged with evaluation marks
How to read this page: Prices and model names move quickly. We link to official plan and policy pages, date every review, and separate vendor claims from what the product actually makes easy. “Commercial use” is not the same as guaranteed copyright, exclusivity, or freedom from third-party claims.

At a glance

Six strong tools. Six different compromises.

ToolBest forStarting priceProvenanceControlEditorial score
Adobe FireflyCommercial creative workflowsFree; Standard $9.99/moContent Credentials on eligible outputsHigh8.8 / 10
ChatGPT Images 2.0Prompt fidelity, editing, textPlan dependent; API usage basedC2PA + SynthID on supported imagesHigh through conversation8.7 / 10
Midjourney V8.2Aesthetic exploration$10/moLimited user-facing provenance emphasisHigh, but idiosyncratic8.4 / 10
Google Imagen / GeminiGrounded iteration and enterprise usePlan or usage basedSynthID by default on supported modelsHigh8.5 / 10
Ideogram 4.0Typography and designed compositionsFree plans; API from $0.03/imagePolicy-dependent; verify export contextHigh for layout8.3 / 10
Leonardo.AiAsset pipelines and creator presetsFree + credit plansLess legible than the leadersHigh7.8 / 10

Scores reflect our editorial weighting: 30% output and instruction fidelity, 20% creator control, 20% safety/provenance, 15% terms and commercial clarity, 10% accessibility, 5% price stability. They are not universal truth wearing a decimal point.

Public review notes

The honest short version.

Best responsible default

Adobe Firefly

Reviewed 19 Aug 2026
8.8

The clearest fit for teams that value production controls and provenance as much as visual quality.

Quality
Very good
Ethics
Stronger training story
Watermarking
Content Credentials
Commercial
Allowed for non-beta features

Firefly’s own models lean on licensed and public-domain material, and Adobe places provenance closer to the product. The tradeoff is a credit system that becomes harder to compare as image, video, audio, and partner models share the same interface.

Best instruction fidelity

ChatGPT Images 2.0

Reviewed 19 Aug 2026
8.7

Remarkably capable at edits, text, and following a complex brief—and therefore unusually capable of making persuasive falsehoods.

Quality
Excellent
Ethics
Strong safeguards, real residual risk
Watermarking
C2PA + SynthID
Commercial
Subject to OpenAI terms

Conversation is the control surface: useful for multi-step correction, reference-aware edits, and dense text. The system card is blunt about the same capability raising deepfake risk. Provenance is now layered, but no signal should be treated as indestructible.

Best aesthetic engine

Midjourney V8.2

Reviewed 19 Aug 2026
8.4

Still exceptional at visually coherent taste—sometimes so eager to make things beautiful that it edits your intent.

Quality
Excellent
Ethics
Mixed transparency
Watermarking
No standard visible mark
Commercial
Plan restrictions apply

V8.2 improves aesthetic personalization and remains a superb ideation partner. Public-by-default workflows and commercial-use conditions deserve attention; companies above $1 million gross revenue need Pro or Mega for covered commercial use.

Best provenance signal

Google Imagen / Gemini

Reviewed 19 Aug 2026
8.5

A strong enterprise and iterative option with SynthID woven into the product story rather than bolted on afterward.

Quality
Excellent
Ethics
Strong documented safety layer
Watermarking
SynthID
Commercial
Service terms apply

Google’s current image stack spans consumer Gemini, developer Gemini image models, and Imagen on Vertex AI. That flexibility is useful, but it makes a single “price” misleading. Treat model, interface, region, and contract as separate choices.

Best typography

Ideogram 4.0

Reviewed 19 Aug 2026
8.3

The designed-composition specialist: text, layout, graphic assets, and an increasingly practical production path.

Quality
Very good
Ethics
Policy clear; provenance less central
Watermarking
Context dependent
Commercial
Broad use under terms

Ideogram remains unusually good when the image must contain readable words or behave like a designed object. The 4.0 API pricing is refreshingly legible. Free generations and public visibility still require deliberate privacy decisions.

Best creator toolkit

Leonardo.Ai

Reviewed 19 Aug 2026
7.8

An approachable production toolbox with excellent iteration features, paired with ownership terms that deserve actual reading.

Quality
Very good
Ethics
Moderate
Watermarking
Less prominent
Commercial
Free-user caveats

Leonardo is useful for asset families, character work, canvas edits, and creators who want controls without living in a prompt box. Its own guidance says free users receive commercial rights but their public creations are owned by Leonardo as between the parties—an important distinction.

Field notes — public

Failures, watermark reality, and the fine print that bites.

This is the useful depth we keep outside the paywall. Members get the repeatable prompt pack, deletion-test logs, plan-picker calculator, and the versioned decision matrix — but you shouldn't need to pay $1 to avoid a predictable mistake.

Prompt collapses we actually hit

Literal intent loses to house style

Midjourney: “a person in a plain white room, no decoration” → added tasteful texture and window anyway. Ideogram: dense layout briefs drift on spatial relationships (“logo top-left” slips center). ChatGPT Images: complex multi-constraint edits compound error — verify each atomic edit before chaining.

Mitigation: seed once, vary one constraint at a time, keep the brief under 3 competing instructions. Lock a reference composition before pushing fidelity.

Provenance in the wild

Metadata rarely survives distribution

Content Credentials (C2PA), SynthID, and similar signals are stripped by screenshots, recompression, crops, and many messaging apps. A “clean” export can look identical to human-made media downstream. Treat provenance as one layer, not proof.

What survives better: visible disclosure near the image, plus a durable caption or alt-text that persists when metadata doesn't. Log the source capture alongside distribution.

Watermark stripping — honest test

Light edits break invisible marks often

In our quick strip-test (resize to 1024px, JPEG q=82, mild crop, screenshot), at least one invisible scheme became unverifiable on every tool after 2 steps. No tool we checked guarantees persistence through adversarial distribution.

Takeaway: don't promise compliance on watermark grounds alone. Use contractual consent + disclosure rather than “the tech will catch it.”

Commercial-use gotchas

“Allowed” ≠ indemnified, exclusive, or yours alone

Leonardo free-tier outputs carry platform ownership language. Adobe Firefly “commercially safe” applies to eligible generations, not partner models — check the model badge. Midjourney plan tier affects visibility/stealth and qualifying commercial status. ChatGPT subject to OpenAI terms, not yours.

Before client delivery: confirm tier, model source, stock/brand/IP exposure, and whether client expects exclusivity your provider does not grant.

How we tested this: one-line prompt on each tool, then one constrained edit, one resize+recompress, one screenshot round-trip. Sources and dates on every claim; corrections feed is public and flag mail is editor@persongenerated.com. Deeper repetition logs and matrices in member reviews.

How scores travel

What changes our ranking.

Data batch

Privacy & deletion

We re-check data retention, training opt-outs, and deletion-workflow reality on a rolling cadence. Dated.

Price batch

Plans & credits

Generative-credit math, overage, and plan-tier fences. Linked to official plan pages at review time.

Integrity batch

Provenance

Which exports retain C2PA/SynthID after real distribution — and what we tell readers instead.

Explorer membership

The full reviews cost one dollar.

Unlock tested prompt behavior, privacy and deletion notes, failure modes, tier-by-tier recommendations, change history, and our complete decision matrix. Analyst members also receive the downloadable comparison PDF.

Member reviews