Midjourney vs ChatGPT Images vs Firefly vs Gemini: AI Image Generator Comparison 2026

Quick answer

Midjourney, ChatGPT Images, Adobe Firefly, and Google Gemini are four AI image tools that are often evaluated for the same production tasks. Between April and September 2026, all four changed in ways that affect tool selection: Midjourney moved from V7 to V8.2 as its default version, OpenAI released ChatGPT Images 2.5 on September 8, Google moved Nano Banana 2 and Nano Banana Pro to general availability, and Adobe expanded Firefly with partner models. DALL·E was retired over the same period: the API models shut down in May and the official DALL·E GPT was removed in August.

This guide was checked against vendor documentation, pricing pages, terms, and changelogs through 2026-09-20. It compares verifiable facts such as models, pricing, APIs, plans, commercial terms, and watermarking. It does not include a standardized output benchmark, so treat the table below as a shortlist for what to test first, not as an image-quality ranking. Use your own assets to evaluate actual visual quality, text accuracy, and character or product consistency.

Commercial permission and copyright protection are separate questions. A platform may permit commercial use without making a raw AI-generated image automatically copyrightable. Client work also needs separate checks for brand assets, real people, likeness rights, and platform policy.

What changed from April to September 2026

Side-by-side comparison

Tool details

Midjourney

Midjourney is available through midjourney.com and Discord. Midjourney says it does not provide an API, and automated access violates its terms. The web app also includes Draft mode for low-cost drafts and Conversational mode for writing prompts in natural language. There is no free trial on Discord or the main website; only the niji・journey mobile app offers a limited trial. The version documentation says V8.2 has been the default since 2026-07-24. V8.1 launched on 2026-04-14 and was the default from 2026-06-10 through 2026-07-23. The early V8.0 alpha is no longer available, while other versions can still be selected in settings.

Reference-image workflows depend on the version:

• V8.1 / V8.2: use the Edit Model. It accepts up to 4 reference images, supports instruction-based editing, and works with Editor tools for inpainting and expansion. It replaces Omni Reference and Character Reference for V8, does not support --tile, and its results cannot be directly remixed.
• V7: use Omni Reference with --oref; Midjourney documents it as V7-only.
• V6 / Niji 6: Character Reference uses --cref and --cw; Midjourney now treats this as a legacy workflow.

Common controls:

• --ar 16:9 or --ar 4:5: set the aspect ratio before iterating.
• --seed 1234: keep composition changes closer across variations.
• --sref: style reference for V6 and later; --sw ranges from 0 to 1000 and defaults to 100. It is not compatible with Moodboards.
• --hd / --sd: V8.1/V8.2 support HD, while V6 and V7 are SD only. At 1:1, HD is 2048x2048 and SD is 1024x1024. HD supports aspect ratios up to 4:1, versus 14:1 for non-HD output.
• Draft mode: V8.1/V8.2 generate 24 drafts at 512 px for 0.4 minutes of Fast time; V7 Draft generates 4 images at half the cost of a standard V7 prompt.

Pricing and commercial-use boundaries: Basic is $10/mo, Standard $30, Pro $60, and Mega $120, with roughly 20% off annual billing and additional Fast time at $4/hour (official plans). GPU cost is measured per prompt, not per image: Midjourney lists about 0.8 minutes for an SD prompt and 1.3 minutes for an HD prompt. A single prompt returns multiple images, so Basic's 3.3 Fast hours should be budgeted by prompt count rather than image count. The Terms of Service say a company, or an employee of a company, making over $1,000,000/year must use Pro or Mega to own generated assets. Content is public and remixable by default; confidentiality requires Pro or Mega Stealth mode. The current terms are effective 2026-05-27.

Inpainting or expanding an HD image returns the result to SD. Upscale again if the final deliverable needs HD output.

If the image must contain exact text, UI labels, legal copy, or strict product geometry, do not treat the generated image as final layout. Put those items on the acceptance checklist and finish them in a precision editing tool.

ChatGPT Images

For most users, what used to be called DALL·E now means image generation inside ChatGPT. OpenAI's release note says ChatGPT Images 2.5 began rolling out on 2026-09-08 across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. OpenAI also says generation latency can be up to 50% lower and that multi-turn editing consistency and transparent-background output improved. The creation tools are not identical on every surface: the current ChatGPT Release Notes say Templates are not yet available in Work mode, while Sketch and image edit/comment flows are documented on mobile.

Put ChatGPT Images on the shortlist when you need:

• Templates, Sketch references, or on-image annotations to show where a change should happen; Templates are not yet available in Work mode, and OpenAI currently documents Sketch plus image edit/comment flows on mobile.
• Natural-language multi-turn iteration for non-designers.
• Localized edits to an existing image; the API supports inpainting and reference-image editing.
• OpenAI's Images API or Responses API.

If text accuracy is a hard requirement, do not infer an accuracy rate from those product features. Put your real copy into the test set and verify every character.

OpenAI's image generation guide currently lists gpt-image-2.5-sunburst, positioned for precision, and gpt-image-2.5-flare, positioned for speed. Both use 2026-09-08 snapshots. dall-e-2 and dall-e-3 shut down on 2026-05-12. gpt-image-1 is scheduled to shut down on 2026-10-23, while gpt-image-1.5, gpt-image-1-mini, and chatgpt-image-latest are scheduled for 2026-12-01. OpenAI points users to the gpt-image-2 family as the replacement (deprecations). The official DALL·E GPT was removed from ChatGPT on 2026-08-30.

The 2.5 image models are priced by token rather than by a fixed image price: image output is $30/1M tokens, image input $8/1M, and text input $5/1M. As a legacy per-image reference, gpt-image-1.5 costs about $0.009 for a low-quality 1024x1024 image and about $0.133 at high quality, but that model shuts down on 2026-12-01, so it is not a durable budget baseline.

Adobe Firefly

Adobe Firefly exposes both Adobe-built Firefly models and partner models. Adobe publishes training-source information for its own models; partner models keep their own model, data, and licensing conditions. Legal or brand review should not treat those two categories as interchangeable.

Firefly belongs on the shortlist when:

• Final work goes through Photoshop, Illustrator, Adobe Express, or Creative Cloud.
• You need Generative Fill, Generative Expand, vector-oriented workflows, or drafts that stay inside an Adobe review process.
• A client asks how the model was trained or whether contractual IP protection is available.

Pricing is based on generative credits. As of 2026-09-20, Adobe's U.S. individual-plan page lists monthly standard prices of $9.99 for Standard with 2,000 credits, $19.99 for Pro with 4,000, $49.99 for Pro Plus with 10,000, and $199.99 for Premium with 50,000. Standard image features are unlimited. The plan page may also show temporary promotions, so use the actual checkout price for budgeting. Adobe's generative credits FAQ says credits do not roll over; users can wait for the next monthly reset or buy add-on credits. The free tier provides a limited daily allowance for selected models.

Adobe-built and partner models use different rules:

• Adobe says its Firefly models are trained on licensed Adobe Stock content and public-domain material whose copyright has expired. Outputs include Content Credentials, and Adobe says customer content is not used to train its models.
• Partner models are not developed by Adobe. Adobe says users are responsible for deciding whether a model is appropriate for their project. Some models from Chinese providers send prompts and reference files to the corresponding companies in China, and these models are not available in enterprise Creative Cloud.
• Firefly Output Indemnification for Adobe models is not automatically included with an individual paid plan. It only applies when the customer agreement includes the relevant indemnity and the output comes from an eligible Firefly feature, entry point, and export event. Adobe's Generative AI Product Specific Terms dated 2026-04-23 list copyright, trademark, right of publicity, and privacy among the covered third-party rights, subject to conditions and liability limits. Teams and enterprise customers should rely on their actual contract and entitlement.
• Partner models do not automatically inherit Firefly Output Indemnification. Adobe's current Creative Partner Model Supplemental Coverage explicitly lists eligible generally available Google and OpenAI image/video models and requires the relevant customer agreement or qualifying paid license to include that coverage. Beta, trial, preview, and early-access features are excluded. Do not assume the same coverage applies to other partner models.

Firefly AI Assistant entered public beta on 2026-04-27 and expanded on 2026-06-18 to Photoshop, Illustrator, InDesign, Premiere, and Frame.io. It is available to Creative Cloud Pro, Firefly Pro, Pro Plus, and Premium users and has a separate beta allowance.

Gemini Nano Banana

Gemini supports image generation and editing inside the same conversation as text and image context. Google's current image documentation lists three generally available models: Nano Banana 2 (gemini-3.1-flash-image) as the general-purpose option, Nano Banana Pro (gemini-3-pro-image) for more complex visual tasks, and Nano Banana 2 Lite (gemini-3.1-flash-lite-image) as the lowest-priced option. The older Nano Banana model, gemini-2.5-flash-image, is deprecated, with an earliest retirement date of 2026-10-02. Imagen 4 has been discontinued.

As of October 4, 2026, the official deprecation schedule describes its dates as the earliest possible retirement dates; the date alone does not establish that every project’s old requests have failed. If your application still calls the old API model, use the Gemini 2.5 Flash Image migration guide to check stable model IDs, choose Nano Banana 2 or Lite, and replay existing tasks. A preview replacement named in the retirement table may itself be outdated.

Current reference-image and multi-turn limits:

• Nano Banana 2 supports up to 10 object reference images and 4 character reference images.
• Nano Banana Pro supports up to 6 object, 5 character, and 3 style references. Google says the combined cap across those categories is 14, and Pro can output 1K, 2K, or 4K images.
• Nano Banana 2 Lite supports up to 14 object references but does not support character or style references. Google also says it is not intended for multi-reference or long multi-turn editing workflows.
• Google recommends multi-turn iteration and supports continuing edits through previous_interaction_id.
• Gemini 3.1 Flash Image can also turn video into images using a public YouTube URL or an uploaded file.
• Google says the Gemini 3 image family supports relatively complex text rendering for infographics, menus, diagrams, and marketing assets. Legal copy and labels still need character-by-character review.

Pricing is published by resolution (official pricing, updated 2026-09-16): Nano Banana 2 is about $0.045 for 0.5K, $0.067 for 1K, $0.101 for 2K, and $0.151 for 4K. Nano Banana 2 Lite is about $0.0336 for 1K. Nano Banana Pro is about $0.134 for 1K-2K and $0.24 for 4K. Batch pricing is about half of standard pricing. API image output has no free tier. The consumer Gemini app includes Nano Banana 2, while paid plans can regenerate with Nano Banana Pro. Google AI plans download at 2K, while non-Google-AI plans download at 1K. Google does not publish a fixed daily image count.

Gemini API images include an invisible SynthID watermark. In Gemini Apps, users can toggle the visible watermark, but the setting does not remove SynthID or the Content Credentials (C2PA) written by Gemini Apps. The setting only applies to Gemini Apps. In India, South Korea, and Vietnam, the toggle is currently limited to AI Ultra users; other users automatically receive the visible watermark. Google Flow has its own visible-watermark control, and the Flow help page says Flow output retains SynthID.

If you want to generate personalized images from a connected Google Photos library, see How to Use Gemini with Google Photos for Personalized AI Images.

A practical prompt skeleton

All four tools can start from the same information structure:

[subject] + [scene] + [style] + [lighting] + [camera/composition] + [constraints] + [output format]

Example for an article hero image:

Create a 16:9 editorial hero image for an article about cloud cost audits.
Subject: a small engineering team reviewing a glowing dashboard at night.
Style: cinematic but realistic, not cartoon, no corporate stock-photo smiles.
Lighting: cool monitor light with one warm desk lamp.
Composition: leave empty space on the right third for a headline.
Constraints: no logos, no readable text, no distorted hands.
Output: high-detail web hero image.

If the image must contain text, keep the copy short and quote it exactly:

Create a clean product announcement card. Put only these two text strings on the card:
"SPRING LAUNCH" and "30% OFF". Use large sans-serif type, centered layout,
plain cream background, and no other words.

Midjourney parameters can be appended to the visual description:

editorial portrait of a robotics designer in a small Tokyo studio, warm task light,
35mm documentary photography, natural skin texture, realistic workspace --ar 4:5 --seed 1872

With a reference image, spell out what must stay and what may change. The same principle works for ChatGPT Images, Gemini, and Midjourney's Edit Model:

Use the uploaded product photo as the exact object reference. Keep the shape,
material, and color. Replace only the background with a soft morning kitchen scene.
Do not add labels, logos, extra buttons, or new packaging text.

Test with your own assets

When you actually need to choose a tool, a small test using your own work is more useful than a broad synthetic benchmark.

1. Fix the test set first. Pull a few task types from real work: short text, the same character or product across multiple scenes, localized edits, and the everyday hero images or assets you actually produce. Use the same inputs across all candidates instead of rewriting prompts to favor one tool.

2. Record observable outcomes. Track retry count, which attempt became usable, manual cleanup time, whether text is correct, whether the same character or product stays consistent, and whether a localized edit damages untouched areas. You can record aesthetic preference separately, but do not collapse it into the same score as acceptance criteria.

3. Estimate cost at one production scale. For example, estimate 30 finished images. Midjourney is budgeted by prompt count because one prompt returns multiple images; using the official SD estimate of about 0.8 Fast minutes per prompt, Basic's 3.3 hours (198 minutes) is roughly 250 prompts before inpainting, upscaling, or HD costs. Convert OpenAI's token pricing into your expected image cost, multiply Gemini's per-resolution price, and check how quickly Firefly credits are consumed and when standard-image unlimited use applies.

4. Eliminate by hard constraints first. If you need an API, Midjourney is out. If the client requires published training-source information or a contractual indemnity review, start with Adobe's own Firefly models. If every character of image text must be correct, focus your test on ChatGPT Images and Gemini. Compare the remaining candidates only after those constraints are satisfied.

Stop testing a candidate when it repeatedly misses a hard requirement in your fixed test set. That only means the current model, settings, assets, and task do not meet your requirement; it does not establish a universal limit for the product.

Pre-delivery checklist

• Text: verify every required word, especially small UI copy, multilingual text, prices, dates, and legal language. Do not deliver generated text that has not been checked.
• Exact layout: if spacing, dimensions, grids, or pixel positions are hard requirements, use the generated image as a draft and finish the layout in Photoshop, Illustrator, Figma, or another precision tool.
• People and object constraints: if headcount, hands, props, poses, or object positions matter, include them in the fixed test set and check them item by item rather than relying on overall visual impression.
• Brands and characters: for logos, protected characters, real people, or client assets, verify permission, platform policy, and intended use. The fact that a model can generate something does not mean the output is cleared for commercial use.
• Final file: client-ready images should pass QA for dimensions, text, brand elements, and export format. Generation is one step in the production workflow.

Commercial use and copyright

As of 2026-09-20, the practical rule is to check commercial-use terms first, then evaluate copyright separately.

The U.S. Copyright Office continues to focus on human authorship. Prompting alone is generally not enough to make raw AI output automatically copyrightable. Selection, arrangement, localized repainting, compositing, and other human expression can matter. For client work, keep prompts, source files, edit history, and screenshots of applicable license terms.

FAQ

Is DALL-E still the right name?

Not for the current product path. dall-e-2 and dall-e-3 shut down on 2026-05-12, and the official DALL·E GPT was removed from ChatGPT on 2026-08-30. People still search for "DALL-E," but the current user-facing workflow is ChatGPT Images 2.5, and API integrations should use the gpt-image-2 family.

Which AI image generator should I check first for commercial client work?

If the client specifically requires published training-source information and a contractual indemnity review, Adobe publishes training-source information for its own Firefly models and offers conditional output-indemnity mechanisms. Actual eligibility depends on whether the customer agreement includes the relevant terms and whether the feature, entry point, and export event are eligible. Partner models use a separate Supplemental Coverage framework.

Which tool is best for text inside images?

This guide does not include a standardized text-accuracy test, so it does not name a winner. ChatGPT Images provides on-image annotations, templates, and multi-turn editing; Google describes relatively complex text rendering as a capability of the Gemini 3 image family. If text accuracy is a hard requirement, test both with your real copy and verify every character before delivery. Legal copy, nutrition labels, and complex UI screenshots should never be accepted without manual review.

Which API is better for image generation?

Gemini publishes image prices by resolution and supports multiple reference images plus video-to-image input. OpenAI's current Images and Responses APIs use token pricing and support reference-image editing and conversational workflows. Choose based on your input types, latency, expected per-image cost, and fixed test-set results rather than assuming one API has better real-world quality or instruction following.

Can AI-generated images be copyrighted?

In the U.S., raw AI output is not automatically copyrightable simply because it looks like a finished work. Human authorship is the key issue. Substantial editing, arrangement, compositing, or repainting may make the human-created portions protectable, but the analysis is fact-specific.

When per-image cost has to fit a monthly budget, the AI coding pricing comparison has usage-based monthly estimates. If the output is video rather than stills, the AI video generation comparison splits tool choice by source material.