Midjourney v7 launched in alpha on April 3, 2025 — what CEO David Holz called “a totally different architecture,” not an incremental update. The most immediately useful addition is Draft Mode, which renders at ten times the speed of standard mode at half the GPU cost, and switches the interface to a conversational layout where you can say “make it nighttime” or “swap the cat for an owl.” That iteration loop is genuinely faster for early-stage concepting.

For pure visual quality, nothing else in 2026 touches it consistently. Curious Refuge’s 29-scenario benchmark scored it 8.62/10 overall, praising the “cinematic” and “painterly” default rendering. Skin textures, lighting fall-off, atmospheric depth — these details have a composure that makes other models’ output look like they’re trying. Midjourney also solved the long-running “hands problem” — anatomically correct hands are now the norm, not the exception.

But there are real limitations, and they’re worth being honest about. Text rendering is weak — roughly 30% accuracy on short phrases, which means any image requiring legible words is the wrong job for Midjourney. There’s no official API, which rules it out for developers building production workflows. And the ongoing copyright lawsuits from Disney, NBCUniversal, DreamWorks, and Warner Bros. (filed in 2025, unresolved as of April 2026) introduce genuine commercial risk for large-scale professional use.

The Niji 7 model (anime-focused, released January 2026) and the v8 Alpha (March 2026, adding 2K native resolution and 5x faster generation) show the pace of development remains high. Midjourney isn’t resting on its artistic reputation.

FLUX 2 Pro from Black Forest Labs is the model that makes professional photographers nervous, and that’s not an exaggeration. Built on a 32-billion-parameter Rectified Flow Transformer architecture, it generates photorealistic output up to 4 megapixels — sharp enough for large-format print — with sub-10-second generation times via API. The March 3, 2026 speed update doubled generation speed with no reported quality loss.

On the LM Arena leaderboard, FLUX 2 Pro v1.1 registers an ELO of 1,265 — essentially tied with GPT Image 1.5 at the top of the general-purpose rankings. In specialized photorealism testing, it consistently leads. Skin texture, lighting physics, fabric folds, material rendering — all of these behave like real photographs rather than generated images.

The Kontext Engine allows natural language image editing: describe the change you want, and the model applies it to an existing image while preserving the surrounding composition. Its multi-reference system lets you upload up to 10 source images for composite work. These are production features that matter on real projects.

Where FLUX 2 falls short is artistic flexibility. Its training skews heavily toward photographic realism, which means “watercolor illustration” or “impressionist painting” prompts produce competent but unremarkable results compared to Midjourney. It’s not a tool for creative exploration — it’s a tool for precise commercial execution. FLUX.1 Schnell (the lightweight open-weight variant) is also one of the few models fully trained on licensed data, giving it cleaner commercial credentials than the Pro variant.

GPT Image 2 (API name: gpt-image-2) launched April 21, 2026, introducing something genuinely new: a native reasoning loop that plans image layout mathematically before pixel generation begins. The result is measurably better handling of complex spatial instructions — “red ball on top of blue box, with green background, left side of frame” — compared to any diffusion model that jumps straight to generation.

Text accuracy is approximately 95% across Latin script, and the AVB benchmark documented it outperforming Midjourney v8.1 on typography while trailing Nano Banana Pro on photorealism specifically. On dense in-image text — infographics, UI mockups, posters with multiple text elements — it’s the most reliable option currently available.

The ChatGPT integration makes it the most accessible model for non-technical users. You describe what you want conversationally, refine it in plain English, and iterate without needing to learn prompt syntax. GPT-4o previously replaced DALL-E 3 as OpenAI’s flagship; GPT Image 2 has continued that evolution with the reasoning layer.

The tradeoff is aesthetic ceiling. Results trend toward clean, competent, and accurate — which is exactly what many commercial workflows need — but they rarely achieve Midjourney’s compositional drama or FLUX 2’s photographic depth. One practical note: Thinking Mode can take up to two minutes per generation, which matters for time-sensitive workflows. Disabling it for the Instant mode restores normal speed.

Ideogram was built by four former Google Brain researchers — Mohammad Norouzi, William Chan, Chitwan Saharia, and Jonathan Ho — specifically to solve the text-in-image problem. Version 3.0, released March 26, 2025, delivered on the founding premise: approximately 90–95% text rendering accuracy, compared to roughly 30% for Midjourney and comparable tools.

That gap is the entire story for a significant category of professional work. If you’ve ever tried generating a poster, logo, or marketing banner with Midjourney and ended up with garbled, unreadable text — you know what Ideogram fixes. A “Coffee Bar” logo actually says “Coffee Bar.” A Black Friday sale poster with pricing actually shows readable pricing.

Version 3.0 also introduced Style References: upload up to three images to guide the aesthetic of a generation. Describing visual style in words is notoriously unreliable; showing the model is a genuine quality-of-life improvement. The advanced color palette system lets designers specify exact brand colors, extracting palettes from reference images for consistency across assets.

The photorealism in v3.0 has improved substantially over v2.0 — faces, lighting, and textures are markedly better — though it still trails FLUX 2 on raw photographic output. As a specialist platform, Ideogram works best when text or typography is a core design requirement, not as a general-purpose replacement for Midjourney or FLUX.

In ELO testing, Ideogram V3 remains the artistic typography leader for t-shirts, neon signs, and embossed lettering, per the AVB benchmark battery. Google’s Nano Banana 2 is the first serious challenger in this specific area, but Ideogram still leads as of May 2026.

Adobe Firefly’s strategic story in 2026 is less about raw image quality and more about ecosystem position. It’s trained exclusively on Adobe Stock licensed imagery, openly licensed content, and public domain material. That’s not a footnote — it’s the reason enterprise legal teams approve Firefly assets without friction while sometimes rejecting Midjourney output outright. Adobe backs this with IP indemnification for enterprise customers.

The Generative Fill inside Photoshop remains one of the best AI-assisted editing features available — select an area, type what you want, and the blend is seamless in a way that standalone generators can’t match. The same applies to Generative Expand (extending canvas boundaries) and text-to-vector inside Illustrator.

Adobe’s December 2025 strategic pivot changed the platform’s nature substantially. Firefly is now an aggregator: through one subscription, users access Runway Gen-4.5 for video, Black Forest Labs FLUX.2 for photorealism, Google Nano Banana Pro for editing, ElevenLabs for audio, and Topaz Astra for 4K video upscaling. That partner model approach is unique in the market and gives Firefly a durable competitive position regardless of which individual model wins the quality race in any given month.

What Firefly doesn’t do well: standalone artistic generation. Its default outputs tend toward clean, stock-photo aesthetics — useful for commercial production, but without the compositional ambition of Midjourney’s best work. Text rendering in Firefly’s native model is still unreliable; the FLUX.2 integration inside Firefly is the better option for text-heavy designs. For Creative Cloud subscribers, however, Firefly costs essentially nothing additional — its value is already bundled into subscriptions most design professionals hold.

Stable Diffusion 3.5 is not a product — it’s a platform. The base model (available in three sizes, released late 2024) is the foundation for an ecosystem of hundreds of community fine-tunes, LoRAs, and specialized models for anime, architectural visualization, character consistency, product photography, and more. If you need a style that doesn’t exist elsewhere, or a workflow that can’t be built on closed APIs, SD 3.5 is your starting point.

The raw quality gap between SD 3.5 and frontier closed-source models (FLUX 2, Midjourney v7, GPT Image 2) is real and visible in direct comparisons. But that framing misses what SD 3.5 actually is: not a competitor to those models on quality alone, but a platform for teams with technical resources to build things those models can’t do. ControlNet (for enforcing poses or edge detection), custom LoRA training on proprietary datasets, and integration into automated enterprise pipelines — none of this is possible on closed platforms.

Minimum hardware requirement for local operation: NVIDIA GPU with at least 8–12GB VRAM (RTX 3060 minimum), 16GB RAM, and 50–100GB storage. For teams without that infrastructure, cloud-based access via platforms like Replicate or Automatic1111 is available. Expect a meaningful setup investment either way — this is not a tool for people who want results in 90 seconds.

Google’s Imagen 4 is the model that appears least in casual AI discussions and perhaps deserves more. On the LM Arena leaderboard, Gemini 3 Pro Image (which uses Imagen 4 Ultra) sits at ELO 1,252 — meaningfully close to the FLUX 2 / GPT Image 1.5 tie at the top. Its image editing and transformation capabilities are particularly strong: morphing two images together, changing perspectives, and object-level modifications work with a precision that’s ahead of most alternatives.

The API pricing is the most competitive among premium models. Imagen 4 Fast costs $0.02 per image — half the price of GPT Image 1.5 Standard at $0.04 and well below FLUX 2 Pro at $0.055. For high-volume workflows where quality is important but budget is constrained, Imagen 4 Fast via Google Cloud Vertex AI is the clearest value option in the market.

Nano Banana 2 (released early March 2026) is Google’s most recent challenger specifically for text rendering, drawing on Gemini’s knowledge base and real-time web search to generate research-informed visuals. For users inside the Google ecosystem — Google Workspace, Vertex AI pipelines, Gemini integrations — this combination is compelling. Outside that ecosystem, the consumer access story is less straightforward than competitors with cleaner consumer UIs.