AI Image Models 2026 — Midjourney, SDXL, Flux, Ideogram

Choosing an AI image model in 2026 means choosing a license, a price structure, and a pace of change at least as much as an image style. Midjourney is on its eighth major version, Black Forest Labs shipped an open-weight FLUX 3 model this month, OpenAI has retired DALL-E 3, and Google's Nano Banana 2 sits inside Gemini and Adobe Firefly. This comparison, current as of October 22, 2026, explains what each family is built for, how to run or buy it, and why any "best model" ranking ages within months.

The 30-second answer

The table matches common jobs to the families that were built for them. It describes how each vendor positions its model and what its own documentation says, not a head-to-head quality test, because no neutral test covers all of these tools on your prompts.

If your job is	Start with	Reason
Stylized, art-directed imagery	Midjourney	Closed hosted service whose latest listed version is V8.2, released July 24, 2026
Running a model on your own GPU	FLUX 3 Image Klein or Stable Diffusion 3.5	Both ship open weights; Klein's INT8 build runs on a GPU with as little as 8 GB of VRAM
Posters, logos, and images containing text	Ideogram, Nano Banana 2, or GPT Image models	All three vendors document text rendering as a headline feature
Editable vector graphics (SVG)	Recraft V4 Vector	Its documentation says it outputs SVG with scalable geometry
Client work where training-data origin matters	Adobe Firefly's own models	Adobe states they are trained on licensed Adobe Stock and public-domain material
Conversational edits inside a chatbot	GPT Image models in ChatGPT, or Nano Banana 2 in Gemini	Both are built into assistants

Most teams that produce images every week use two or three of these tools and pick per task. Our AI tools for designers guide shows how image generators fit into a wider design stack.

How image models work in one page

Nearly every model in this article starts from the same idea. Training teaches a neural network to remove noise from an image. Run the process in reverse, starting from pure random static, and a prompt steers the cleanup toward a picture that matches the words. It resembles a photographic print developing in a tray: a blur sharpens step by step until a scene appears, except that your text decides which scene.

Black Forest Labs describes FLUX 3 as a "multimodal flow matching model," a close relative of diffusion that learns a smooth path from noise to an image, and it trains on images, video, and audio together (Black Forest Labs, FLUX 3). The mathematics differs between families, but the practical experience is the same: you write a prompt, the model refines noise into pixels, and you iterate. For a full walk-through, see our explainers on how diffusion models work and how image generators understand prompts.

Two consequences matter for choosing a tool.

Text is hard because letters are fine detail. A model has to place thin, exact shapes in the right order, and a single wrong stroke turns a word into gibberish. That is why vendors advertise text rendering as a feature. Ideogram, Recraft, Google, and Black Forest Labs all do so in their own documentation, and it is the clearest axis on which newer models differ from older ones.

Open weights versus hosted service is the biggest practical fork. Open-weight models are files you download and run yourself, which gives control, privacy, and fine-tuning, in exchange for hardware and setup. Hosted models give convenience and usually higher peak quality, in exchange for per-image costs and someone else's rules.

Why "best model" lists expire in months

The timeline below mixes vendor announcements, official documentation, and Wikipedia's dated version table for Midjourney. It shows how quickly the field turns over.

Date	Event	Source
October 22, 2024	Stable Diffusion 3.5 announced (Medium added October 29)	Stability AI
April 4, 2025	Midjourney V7 alpha	Wikipedia
November 14, 2025	OpenAI notifies developers that DALL-E 2 and 3 will leave the API	OpenAI deprecations page
February 2026	Recraft V4 released	Recraft documentation
February 26, 2026	Google announces Nano Banana 2 (Gemini 3.1 Flash Image)	Google
March 17, 2026	Midjourney V8 alpha	Wikipedia
May 12, 2026	DALL-E 2 and DALL-E 3 removed from the OpenAI API	OpenAI deprecations page
July 23, 2026	Black Forest Labs announces FLUX 3	Black Forest Labs
July 24, 2026	Midjourney V8.2	Wikipedia
October 15, 2026	FLUX 3 Image Klein, open weights	Black Forest Labs
October 23, 2026	Scheduled removal of gpt-image-1 from the OpenAI API	OpenAI deprecations page

Two lessons follow. First, a model name is a snapshot. OpenAI's deprecations page already lists the successors to its own successors: gpt-image-1 is scheduled to shut down on October 23, 2026, and gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest on December 1, 2026, with gpt-image-2.5-sunburst and gpt-image-2.5-flare recommended as replacements (OpenAI deprecations). A workflow hard-coded to one model name has a shelf life of months. Second, an older article that calls Midjourney V7 an "early 2026" release, or lists DALL-E 3 as a current option, is out of date. The version history above corrects both.

Midjourney

Midjourney is a hosted service that began on Discord. According to Wikipedia's version table, V7 arrived in alpha on April 4, 2025, V8 on March 17, 2026, V8.1 on April 14, 2026, and V8.2 on July 24, 2026 (Wikipedia, Midjourney). Midjourney's own updates page, in October 2026, shows rolling alpha changelogs and a post titled "Testing Thinking Mode" (Midjourney updates).

What it is good at. Midjourney's reputation rests on defaults: images that look art-directed without much prompt engineering. It remains a popular first stop for concept art, mood boards, and brand-style exploration. Our Midjourney beginner walkthrough and budget guide cover the practical side.

Limits. It is a closed service with no weights to download, and it cannot be self-hosted. It is sold by subscription, and plan prices change, so check Midjourney's site before budgeting.

Legal exposure. Midjourney is a defendant in several copyright cases. Wikipedia lists the 2023 suit by artists against Stability AI, Midjourney, and DeviantArt; a June 11, 2025 suit by Disney and Universal; a September 4, 2025 suit by Warner Bros. Discovery; and a November 2025 consolidation of the three studio suits. None of these has established a final rule on whether training on copyrighted work is lawful, so the question stays open. See AI art, copyright, and ethics for the arguments on each side.

FLUX from Black Forest Labs

Black Forest Labs (BFL) builds the FLUX family. Its site now lists FLUX 3 Max, Video, Image, Image Klein, and Action, along with FLUX Tools. FLUX 3, announced on July 23, 2026, is a multimodal model trained jointly on images, video, and audio. Video clips run up to 20 seconds with native audio, and an "Action" variant predicts robot control actions. At launch, Video and Action were in gated Early Access, with no downloadable weights (VentureBeat on FLUX 3).

The open-weight piece. On October 15, 2026, BFL published FLUX 3 Image Klein, a 5.6 billion-parameter open-weight model for text-to-image generation, editing, and multi-reference generation. Its blog post lists support for up to 10 reference images, native 4K output, and bounding boxes that place elements or confine edits to a region. The generator needs about 11.2 GB of memory in BF16 and 5.6 GB in FP8, and an INT8 build runs on NVIDIA RTX GPUs with as little as 8 GB of VRAM. BFL reports about 4 seconds per 1-megapixel image on an RTX 5090 with the distilled model and INT8 in ComfyUI (Black Forest Labs, FLUX 3 Image Klein). Quantization, the technique behind those smaller builds, is explained in LLM quantization explained, which applies to image models as well.

The license. Klein's weights fall under the FLUX Kommunity License: free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use, with self-hosted licensing or the API for larger organizations. The distilled model is also available through BFL's API as flux-3-image-klein.

What to keep in mind. BFL's own pages describe preliminary image gains in complex prompts and multilingual text rendering. Those are vendor claims. Test them on your own prompts before moving a workflow.

Stable Diffusion

Stability AI's models were the open-source starting point for the field. Stable Diffusion 3.5, announced on October 22, 2024, came in three variants: Large (8.1 billion parameters, aimed at professional use at 1 megapixel), Large Turbo (a distilled version that generates images in 4 steps), and Medium (2.5 billion parameters, generating from 0.25 to 2 megapixels). Medium needs 9.9 GB of VRAM, excluding text encoders (Stability AI, Introducing Stable Diffusion 3.5).

The license. The Stability AI Community License is free for non-commercial use, including research, and for commercial use by organizations under $1 million in annual revenue. Larger organizations need an Enterprise License, and users keep rights to the media they generate.

Why people still use it. The ecosystem around open models matters as much as the models. Tools such as ComfyUI, ControlNet-style conditioning, LoRA fine-tunes, and community checkpoints let a team lock a brand look or a character design in ways a hosted service does not expose. The cost is setup and tuning. The explainer What Is Stable Diffusion? covers the basics.

A caution on dates. SD 3.5 is two years old as of this writing, and the open-weight bar has moved. A buyer comparing it with FLUX 3 Image Klein should compare license terms, VRAM needs, and results on real prompts, not release-year reputation. The older SDXL model that gives this article's title its acronym remains widely supported by community tools, which is its main remaining advantage.

Ideogram

Ideogram built its reputation on text inside images. Its documentation lists the models available on every plan: Ideogram 4.5, which accepts up to five reference images; Ideogram 4.0, which offers very wide aspect ratios suited to banners; and Ideogram 3.0, described as the only model with style presets, style reference images, negative prompts, and color palettes. Transparent-background variants of 3.0 and 4.0 produce images for stickers and icons. Prompts can be plain text or JSON, up to 10,000 characters on 4.5 (Ideogram documentation).

A model hub as well as a model. Ideogram's model list also includes partner models from other vendors, among them GPT Image 2.0, GPT Image 2.5 Flare and Sunburst, and Nano Banana 2 and Pro, marked as premium. That is a useful reminder that the line between "a model" and "a platform that routes to several models" is blurring. Adobe Firefly does the same, as described below.

Pricing. Ideogram's documentation, last verified October 15, 2026, lists Free (30 credits a week with Google, Apple, or Microsoft sign-in), Plus at $20 per month or $15 per month billed yearly (2,400 credits per month), Pro at $60 per month or $42 per month billed yearly (7,500 credits), and Team at $30 per user per month or $20 billed yearly. The Basic plan is no longer sold. Credit costs per image depend on model, resolution, and quality, so the plan price does not translate directly into a per-image price (Ideogram plans). For commercial-use rights, read Ideogram's Terms of Service.

OpenAI's GPT Image models

DALL-E 3 is gone. OpenAI's deprecations page records that DALL-E 2 and DALL-E 3 were announced for removal on November 14, 2025 and shut down in the API on May 12, 2026, with gpt-image-2, gpt-image-1, and gpt-image-1-mini named as replacements. Our article DALL-E 3 for Graphic Design: What It Was and What Replaced It covers the consumer-side switch to GPT Image 2 in April 2026.

The bigger lesson is the speed of replacement. By October 2026, two generations of DALL-E 3's successors are already marked for removal, and the current recommendation is the GPT Image 2.5 pair, Sunburst and Flare. The deprecations page does not describe the differences between those two, so confirm model names and capabilities in OpenAI's API docs before building on them.

Where it fits. These models live inside ChatGPT and the API, so the appeal is conversational editing: describe a change, get a revised image. They also appear as partner models in Ideogram and Adobe Firefly.

Google's Nano Banana

Google's image models live inside Gemini, Search, Flow, Google Ads, and Adobe Firefly. Nano Banana 2, also called Gemini 3.1 Flash Image, was announced on February 26, 2026. Google's announcement lists output from 512 pixels to 4K, readable text for mockups and cards, translation and localization of text inside an image, and availability in the Gemini app, AI Mode and Lens in Search, Google AI Studio, the Gemini API, Vertex AI, and Flow (Google, Nano Banana 2).

Google's model page for the follow-up, titled Nano Banana 2.1, describes keeping up to five characters and fourteen objects consistent across a workflow, mask-based editing, and use of web search for real-time information in infographics. The page uses several names for what appears to be the same model line, so treat version labels with care and confirm the model ID you are calling in the documentation (Google DeepMind, Gemini Image).

Provenance. Google says images carry SynthID watermarks and C2PA Content Credentials. That matters for anyone who needs to show where an image came from, a topic covered in the provenance section below.

Recraft and Adobe Firefly

Recraft targets designers who need graphics rather than pictures. Recraft V4, released in February 2026, comes in raster and vector variants. Its documentation lists V4 at $0.04 per image (about 10 seconds), V4 Pro at $0.25 (2048 by 2048 pixels, about 30 seconds), V4 Vector at $0.08, and V4 Pro Vector at $0.30, and describes V4 as "the only model capable of generating editable, production-quality SVG graphics." Export formats include SVG, PNG, JPG, PDF, TIFF, and Lottie. The same page lists current limitations: style creation, prompt-based editing, image sets, and artistic level control are not yet supported (Recraft V4 documentation). An SVG is a set of editable shapes, not a grid of pixels, so a logo or icon can be resized without blur and recolored in a design tool.

Adobe Firefly is both a model family and an aggregator. Adobe states that its own Firefly models are "commercially safe," trained on licensed Adobe Stock material and public-domain works, and not trained on users' personal or generated content. Adobe's page also lists partner models from companies such as Google and OpenAI, and it does not say how those were trained. Outputs carry Content Credentials. Paid plans come in Standard, Pro, and Pro Plus tiers with prices that vary by region (Adobe Firefly). For client work where training-data origin is part of the contract, that distinction between Adobe-trained and partner models is the one to read carefully.

Side-by-side comparison

The table collects the verifiable differences. It deliberately leaves out quality rankings, which depend on the prompt, the day, and the evaluator.

Family	Weights	Main way to access	Documented license or cost note	Documented text or edit feature
Midjourney V8.2	Closed	Midjourney service	Subscription; check vendor	"Thinking Mode" listed as in testing
FLUX 3 Image Klein	Open	Download or BFL API	Kommunity License: free commercial under $5M revenue	Up to 10 references, native 4K, bounding-box edits
FLUX 3 Video and Action	Closed (early access)	Gated Early Access	No public API at launch	Video to 20 seconds with audio
Stable Diffusion 3.5	Open	Download or hosts	Community License: free commercial under $1M revenue	Medium needs 9.9 GB VRAM
Ideogram 4.5 and 4.0	Closed	Ideogram web app	Plus $20 monthly; Pro $60 monthly	Text or JSON prompts; up to 5 references on 4.5
GPT Image 2.5	Closed	ChatGPT and OpenAI API	Priced through API usage	Conversational editing
Nano Banana 2	Closed	Gemini, AI Studio, Vertex AI, Firefly	Varies by product	512 px to 4K; text translation in image; SynthID and C2PA
Recraft V4	Closed	Recraft app and API	$0.04 to $0.30 per image	Native SVG vector output
Adobe Firefly models	Closed	Adobe apps and web	Tiered plans; check regional pricing	Stated commercially safe; Content Credentials

Running models locally: what it takes

The open-weight route is realistic on consumer hardware but not free of effort.

Memory is the first constraint. SD 3.5 Medium needs 9.9 GB of VRAM before text encoders, per Stability. FLUX 3 Image Klein needs about 11.2 GB in BF16 or 5.6 GB in FP8 for the generator, and BFL lists an INT8 build for RTX GPUs with as little as 8 GB. VRAM is the graphics card's own fast memory, and it limits which models load at all.

Speed is the second. BFL's figure of about 4 seconds per megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070 with the distilled INT8 build gives a sense of scale, though those numbers come from the vendor's setup in ComfyUI.

Licensing is the third. "Open weights" does not mean "no strings." Both Stability's Community License and BFL's Kommunity License set revenue thresholds, and larger companies must buy a different license or use an API. Read the license before building a product on a downloaded model.

Maintenance is the fourth. Running locally means you manage model versions, drivers, and node graphs in tools such as ComfyUI. Teams that value control and privacy accept that work. Teams that want an image by lunchtime usually do not. The article on the hidden cost of running a local LLM on your laptop describes the same trade-offs for language models, and most of it transfers.

Cost: what a month of image work looks like

Costs fall into three shapes: flat subscriptions, credit plans, and per-image API pricing. The figures below come from vendor pages read in October 2026 and will change.

Approach	Verified price point	How it is billed
Ideogram Plus	$20 monthly, $15 per month yearly	2,400 credits per month
Ideogram Pro	$60 monthly, $42 per month yearly	7,500 credits per month
Recraft V4 via app or API	$0.04 per raster image; $0.08 per vector image	Per image
Recraft V4 Pro	$0.25 raster; $0.30 vector	Per image
Adobe Firefly Standard, Pro, Pro Plus	Varies by region	Monthly subscription
FLUX 3 Image Klein (self-hosted)	Hardware and electricity only, within license terms	Local
Midjourney	See vendor	Subscription

For a rough sense of API pricing, 500 images a month at Recraft V4's $0.04 per raster image is $20, and 500 vector images at $0.08 is $40. A flat plan beats per-image billing when volume is high and predictable. Per-image billing wins for bursty work. Whatever the model, budget time for rejected generations: you pay for every attempt, not only the ones you keep.

Copyright, licensing, and provenance

This section covers general, well-documented points, not legal advice.

Training-data lawsuits are unresolved. Artists sued Stability AI, Midjourney, and DeviantArt in January 2023, and studios have sued Midjourney since 2025. These cases ask whether training on copyrighted images is lawful, and Wikipedia's account shows they remain in progress in 2026. Until courts settle the question, vendors differ in how they handle it. Adobe emphasizes licensed training data; others rely on legal arguments about fair use.

Output rights depend on the contract. Stability's Community License says users keep rights to the media they generate. Other vendors' terms differ, and free tiers often carry restrictions. The safe habit is to read the terms of the exact plan you use before shipping client work.

Provenance labels are spreading. C2PA is an open standard, also called Content Credentials, that records where a piece of content came from and how it was edited, likened by its backers to a nutrition label. Its steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok, and Truepic (C2PA). Google says Nano Banana 2 images carry both SynthID watermarks and C2PA credentials, and Adobe says Firefly outputs carry Content Credentials. A label proves origin only when it is present and intact; metadata can be stripped, so absence of a credential does not show that an image is human-made. Our guide to AI detection tools covers the limits of detecting unlabeled images.

Workflows that combine models

The patterns below come from how the tools' documented features fit together, not from a controlled study.

Exploration, then precision. Generate broad variations in a style-oriented service such as Midjourney, then move the chosen direction into a tool with stronger edit controls: masked editing, reference images, or bounding boxes. FLUX 3 Image Klein and Ideogram 4.5 both accept multiple reference images for that second step.

Brand consistency at volume. Train a LoRA on a small set of approved images, generate variants programmatically, and review by hand. Open-weight models support this pattern, and Ideogram's plans list custom models, with different license and privacy implications.

Text-heavy layouts. Generate the picture, and set the final typography in a layout tool such as Figma or Canva when exact wording matters. Even models that render text well can alter a letter, and a proofread costs less than a reprint.

Logos and icons. Use a vector-native model such as Recraft V4 Vector so the result stays editable, then clean up in a vector editor. For after-generation retouching, upscaling, and compositing, see our guide to AI image editing tools.

Marketing assets. Teams that compare tools for campaigns should read how to compare AI image editors for marketing and AI tools for marketers first.

What still goes wrong

Every vendor lists strengths. A few limitations show up in the vendors' own documentation, and others are reasons to test before you commit.

• Features are uneven even within one vendor. Recraft's V4 page says style creation and prompt-based editing are not yet supported, and BFL says its FLUX 3 image results are preliminary and expected to improve before release.
• Access is staged. FLUX 3 Video and Action started in gated Early Access, and Midjourney tests new modes in alpha. A feature announced is not a feature available.
• Names and versions shift. Google's Nano Banana model page uses 2 and 2.1 in the same document. OpenAI's deprecations page replaces model IDs every few months. Pin the exact model ID in any automated workflow.
• Provenance is partial. Watermarks and credentials help, but only for images that keep them.
• Specialist accuracy is unproven. Medical illustrations, engineering diagrams, and scientific figures need expert review, because a plausible-looking image can contain wrong content.

A practical test: write your three hardest prompts, with a block of text, a scene with several distinct subjects, and a product shot with a specific logo or color, and run them on each candidate. That reveals more than any leaderboard.

FAQ

Q: Which AI image model is best in 2026? A: It depends on the job. Midjourney is known for art-directed looks, Ideogram and Recraft emphasize text and design assets, FLUX 3 Image Klein and Stable Diffusion 3.5 offer open weights, and GPT Image and Nano Banana are built into popular assistants. Test your own prompts, because vendor claims and public rankings change faster than most buying cycles.

Q: Is DALL-E 3 still available? A: No. OpenAI's deprecations page lists DALL-E 2 and DALL-E 3 as removed from the API on May 12, 2026, and recommends GPT Image models as replacements. GPT Image 1 is scheduled to shut down on October 23, 2026, and older GPT Image variants on December 1, 2026.

Q: Can I use AI-generated images commercially? A: Often yes, but it depends on the vendor, the plan, and the model's license. Stability's Community License allows commercial use under $1 million in annual revenue, BFL's Kommunity License under $5 million, and Adobe describes its own Firefly models as "commercially safe." Always read the terms for your exact plan, and consult a lawyer for high-stakes work.

Q: Can I run a good image model on my own computer? A: Yes, within hardware limits. FLUX 3 Image Klein has an INT8 build for RTX GPUs with as little as 8 GB of VRAM, and SD 3.5 Medium needs 9.9 GB, excluding text encoders. Speed varies; BFL reports about 4 seconds per 1-megapixel image on an RTX 5090 with the distilled model.

Q: What is the cheapest way to start? A: Free tiers and open weights. Ideogram offers 30 credits a week on its free plan with Google, Apple, or Microsoft sign-in, and open-weight models cost nothing per image on your own hardware. Our roundup of free AI image generators lists more options, with current terms to verify.

Q: How can I tell whether an image was made by AI? A: Check for Content Credentials or a vendor watermark, such as Google's SynthID. If the label is missing, that does not prove the image is human-made, because metadata can be stripped and detection of unlabeled images is unreliable.

Q: Why do AI images sometimes garble text? A: Letters are thin, exact shapes, and one wrong stroke ruins a word. Newer models from Ideogram, Recraft, Google, and Black Forest Labs document better text rendering, but for exact wording, set the text in a layout tool afterward.

Related Articles

• AI Image Generators Explained: How DALL-E, Midjourney, and Stable Diffusion Work
• Text-to-Image AI: DALL-E, Midjourney, and Stable Diffusion Compared
• AI Chatbot Pricing 2026

The takeaway

The useful question about AI image models in October 2026 is less "which is best" than "which constraints matter for this job." Open weights, a revenue-capped license, a vector output format, a provenance label, and a monthly price are all easier to verify than a claim about beauty, and they decide whether a tool fits a workflow or breaks it.

The fast turnover is the other lesson. DALL-E 3 left the API in May 2026, its successors are already scheduled to follow, and Midjourney has released V8, V8.1, and V8.2 since March. Build workflows around capabilities you can test and exact model IDs you can pin, expect to revisit the choice every quarter, and treat any ranking, including this one, as dated the day it is published.

Last verified: October 22, 2026