AI Image Model Leaderboard May 2026: Imagen 4 vs GPT-Image-1 vs Flux 2 vs Recraft V4 vs Ideogram 3

The May 2026 Image-Model Lineup

If you read an AI-image comparison written before late 2025, throw it out.

The top of the field is no longer Midjourney, DALL-E 3, and Stable Diffusion. In May 2026 the five models that serious teams are shipping with are:

• Imagen 4 from Google DeepMind — current photorealism leader
• GPT-Image-1 from OpenAI — current prompt-adherence leader
• Flux 2 from Black Forest Labs — current cost and speed leader
• Recraft V4 — current design-system and vector leader
• Ideogram 3 — current typography and poster-layout leader

All five are generally available via paid APIs, all five have published commercial-use terms, and all five have shipped major updates between January and April 2026. This guide answers one question: which one do you actually pick for the job in front of you?

If you also need a non-image-model orientation, our Claude 4.7 vs GPT-5 vs Gemini 2.5 Deep Think comparison covers the text-model frontier on the same May 2026 timeline.

Methodology Box

Before the numbers, the rules:

• 25 prompts spanning portrait photography, product photography, infographic with text, complex multi-subject scene, brand illustration, poster typography, architectural rendering, and stylized art
• Each prompt run three times per model, median selection chosen via blind rating
• Three independent raters scored each output 1-5 on photorealism, text rendering, and aesthetic
• Prompt adherence measured as the percentage of explicit prompt requirements (e.g., "wearing a blue jacket", "two people", "sunset lighting") that the median output satisfied
• All five accessed via official APIs between May 2 and May 6, 2026
• Latency measured wall-clock from request to last byte received over the same gigabit connection
• All images generated at 1024x1024 with default sampler settings — no per-model prompt engineering

We are explicitly not including the open-weights Stable Diffusion 3.5 family. It is still useful for self-hosted workflows but is no longer competitive on quality with the five proprietary models above.

Headline Comparison Table

The headline tradeoffs are honest tradeoffs — there is no single winner across all six dimensions, which is why "best image model" is the wrong question.

Imagen 4 — Photorealism and the Camera Language

Imagen 4 is Google DeepMind's flagship image model, currently exposed through Vertex AI and the Gemini API. In our test set it landed the most natural-looking portraits and product shots of the five, with two specific advantages worth calling out.

First, camera-language adherence. Prompts like "85mm portrait, f/1.8, golden hour, shallow depth of field on Kodak Portra 400" produced outputs that genuinely looked like that combination of optics and film stock — bokeh fall-off shape, grain structure, and color science were all in the right neighborhood. The other four models broadly understood the words but composed flatter, more obviously synthetic frames.

Second, skin and hair detail. Across ten human-portrait prompts, Imagen 4 produced the fewest plastic-skin and uncanny-eye artifacts. GPT-Image-1 was a close second; Flux 2 and Ideogram 3 were noticeably behind on this axis.

Where Imagen 4 lost ground was stylized illustration. Prompts targeting Studio Ghibli-style or vintage-poster-style outputs consistently came back too photoreal, as if the model wanted to interpret every prompt as a photograph. For non-photo work, you will get better results from Flux 2 or Recraft V4.

The Vertex AI generative-AI documentation covers the full Imagen 4 parameter set, including the new aspect-ratio presets and editing endpoints that shipped in March 2026.

GPT-Image-1 — The Prompt-Adherence Champion

GPT-Image-1 is the model that powers ChatGPT image generation and the OpenAI Images API since DALL-E 3 was deprecated in late 2025. In our test it scored the highest raw prompt adherence at 95%, with a particular advantage on dense, multi-subject scenes.

The pattern was consistent: any prompt with three or more named subjects, explicit positional relationships, or counted objects ("five red apples on the left, three green pears on the right, a wooden table between them"), GPT-Image-1 nailed first try. The other four models needed retries to get the counts right.

Text rendering tied with Ideogram 3 at 5/5 — both are now genuinely reliable for short copy in posters, infographics, and signage. GPT-Image-1 went slightly further on layout-aware text (text that wraps around objects, follows curves, or sits inside callouts) which is meaningful for editorial illustration.

The downsides are real. Speed: ~12 seconds per image, three times slower than Flux 2. Cost: $0.07 per HD image, the most expensive of the five. Conservatism: GPT-Image-1's default style is the safest and least visually distinctive of the group, which is a feature for brand-safe work and a bug for art direction.

Flux 2 — Cost and Speed at Aesthetic Quality

Flux 2 from Black Forest Labs is the open-weights-friendly value play. Through the Flux 2 API directly, or through Replicate, fal.ai, and Together AI, you can generate at $0.025 per 1024x1024 image in roughly 4 seconds.

For aesthetic-driven work — blog hero images, social media, mood boards — Flux 2 was the favorite of two of our three blind raters. It produces more visually interesting compositions than Imagen 4 or GPT-Image-1, with stronger color, more dramatic lighting, and looser interpretation of prompts that benefits stylized work.

Where Flux 2 falls short:

• Long text strings still hallucinate characters; for typography-heavy work use Ideogram 3 or GPT-Image-1
• Strict prompt adherence on counted objects is meaningfully behind GPT-Image-1
• Hands and complex anatomy still produce artifacts at higher rates than Imagen 4

For most publications, the Flux 2 pricing and speed make it the right default for high-volume content workflows, with one of the other four models reserved for hero images that justify the extra cost.

Recraft V4 — The Design-System Specialist

Recraft V4 is unique in this group: it natively outputs editable SVG, supports persistent style references across generations, and produces vector-clean illustration that drops into a design system without a Photoshop pass.

For brand illustration, icon sets, editorial spot art, and infographic visuals that need to match an existing visual language, Recraft V4 is the right answer even when it loses on raw photorealism. The "Vector" mode generates SVG that opens cleanly in Figma or Illustrator. The "Recraft 20B" model handles photorealistic and illustrative outputs in a single endpoint.

Limitations are predictable: don't pick Recraft V4 for photo-heavy work. Pick it when you need an icon library that matches your brand, a series of editorial illustrations that read as one author's hand, or vector source files for further editing.

Ideogram 3 — Typography and Poster Layout

Ideogram 3 is the typography specialist. For posters, social-media-with-text, ad creative, and any layout where the text is the design, Ideogram 3 ties GPT-Image-1 on accuracy and beats it on layout sensibility — text scales, kerns, and aligns with image elements in ways that look art-directed rather than dropped on top.

At $0.03 per image and ~5-second latency, Ideogram 3 is also priced for production. The first-party API is straightforward; the Magic Prompt feature (auto-expansion of short prompts) helps for quick concept work.

Where Ideogram 3 loses: photorealism (3.5/5, the lowest in the group) and overall stylistic range. It is a specialist tool, not a generalist.

Verdict by Use Case

Pick by job, not by brand:

• Photorealistic marketing photography (people, products, lifestyle): Imagen 4. Camera-language adherence and skin/hair detail are unmatched.
• Text-heavy infographics, posters, or signage: GPT-Image-1 if you need maximum accuracy plus complex layouts; Ideogram 3 if you want better art direction at half the cost.
• High-volume aesthetic content (blog heroes, social posts, mood boards): Flux 2. The cost and speed math wins at scale.
• Brand-consistent illustration and design-system art: Recraft V4. The native SVG output is the differentiator.
• Complex multi-subject scenes with strict prompt requirements: GPT-Image-1. Counted objects, named subjects, explicit positions.
• Mixed editorial workflow with tight budget: Flux 2 as default, Imagen 4 or GPT-Image-1 reserved for hero shots.

For broader tooling context, our AI image generators compared 2026 overview covers UI-driven tools and platforms that wrap these models, and the best AI tools for developers 2026 roundup covers the broader stack these image APIs sit inside.

Where Every Model Still Fails

It would be dishonest to leave the impression that the May 2026 frontier is solved. Across all five models we still see:

• Hands in dense scenes — finger counts, finger lengths, and rings on hands fail at meaningful rates
• Complex reflections — mirrors, glass, and water surfaces produce inconsistent geometry
• More than six distinct subjects — adherence collapses past about six named entities in a single scene
• Real public figures — most providers refuse, and the ones that don't produce poor likeness
• Brand-mark accuracy — small logos and trademarks still get warped or mis-spelled

For hero images that face customers, humans-in-the-loop review is still required. The frontier is good enough to draft and good enough to ship 80% of editorial work, but the remaining 20% is where reputational damage happens.

API Access, Rate Limits, and Pricing in One Glance

Approximate May 2026 pricing for a standard 1024x1024 image, list price before volume discounts:

• Imagen 4 (Vertex AI / Gemini API): $0.04, ~6s, 60 req/min default
• GPT-Image-1 (OpenAI Images): $0.07, ~12s, 50 req/min default
• Flux 2 (Black Forest Labs API and partners): $0.025, ~4s, 100+ req/min via fal.ai or Replicate
• Recraft V4 (Recraft API): $0.04, ~5s, 60 req/min default
• Ideogram 3 (Ideogram API): $0.03, ~5s, 60 req/min default

All five support webhook-based async generation, batch endpoints, and seed-based determinism for reproducible outputs.

How Output Quality Changed Since 2024

The headline number: in May 2026 the median first-attempt output across these five models is roughly the quality of best-of-four cherry-picked outputs from late 2024. That is a 4x reduction in human curation effort for the same final quality, which materially changes the unit economics of editorial pipelines.

The dimension that improved most was prompt adherence on multi-subject scenes. In 2024, "two people, one in a red jacket and one in a blue dress, standing in front of a yellow car" required regenerations more often than not. In May 2026, GPT-Image-1 and Imagen 4 land that prompt first try in about nine cases out of ten.

Photorealism improved less dramatically — 2024 frontier output was already convincing for most use cases. The bigger 2024-to-2026 photorealism gain is in scene complexity rather than per-pixel realism: complex lighting, multiple light sources, accurate shadows, and physically plausible materials all improved meaningfully.

Bottom Line

In May 2026, "what's the best image model?" is a question that requires three follow-ups before it has a useful answer.

If you need photorealism, pick Imagen 4.

If you need prompt adherence on complex scenes, pick GPT-Image-1.

If you need cost and speed at acceptable quality, pick Flux 2.

If you need brand-consistent vector illustration, pick Recraft V4.

If you need typography or poster layout, pick Ideogram 3.

The teams shipping the most polished image work in May 2026 are using two or three of these five in the same pipeline — Flux 2 for volume, Imagen 4 or GPT-Image-1 for hero shots, Recraft V4 for brand assets, Ideogram 3 for any layout where text is the design. Single-model loyalty is the wrong frame.

For a broader survey of AI image tooling and platforms, see our [AI image generators compared 2026 pillar guide](/blog/ai-image-generators-compared-2026).