Image Generation: 5 Frontier Models, One Dashboard

There is no "best" image model. There is only the best image model for this prompt, this mood, this use case, and this moment in time.

Nova ships with five frontier image generators under one roof: Gemini 3.1, FLUX 2, GPT-5, Riverflow, and Seedream. They each have a personality. Here's how we use them — and when you'd want each one.

Gemini 3.1 — the text-in-images champion

Strengths: Rendering legible text inside images, preserving brand-perfect colors, multi-subject compositions that actually respect your prompt.

Gemini 3.1 (Google's Nano Banana-class model) is what you reach for when you need typography to work. Posters with actual readable text. Mockups with UI labels. Infographics. Diagrams with callouts. For years, AI image models made text look like hieroglyphics — Gemini 3.1 is the generation that finally fixed it.

Great for: social posts with captions, product mockups, UI concepts, logos with wordmarks, anything meme-adjacent.

FLUX 2 — the prompt-follower

Strengths: Photorealistic fidelity, precise prompt adherence, lighting that behaves like real lighting.

FLUX 2 from Black Forest Labs is the "does exactly what I asked" model. Give it a seven-adjective compound sentence with specific camera angles and it will return an image where most of those things are actually true. No other open-weight model comes close for prompt fidelity.

Great for: photography references, realistic product shots, scene composition, concept art, anything where "close enough" isn't close enough.

GPT-5 — the creative collaborator

Strengths: Stylized illustrations, editorial compositions, "vibes" images where the feeling matters more than the photographic accuracy.

OpenAI's image generator (gpt-image-1.5 under the hood) is the one you use when you don't quite know what you want but you know the mood. It's opinionated about style in a way FLUX isn't — you give it "cozy Sunday morning" and it delivers a cohesive, editorial-looking result.

Great for: blog hero images, marketing illustrations, kids' book art, stylized portraits, anything in a recognizable "illustration" idiom.

Riverflow — the artistic one

Strengths: Painterly textures, dreamlike compositions, striking color palettes.

Riverflow is the newcomer and the art-director's favorite. It leans into painterly and fine-art aesthetics in a way the others don't. If GPT-5 is the designer and FLUX is the photographer, Riverflow is the illustrator who went to art school and has opinions about brush strokes.

Great for: album covers, book covers, editorial illustration, anything with a strong artistic point of view.

Seedream — the fast, affordable workhorse

Strengths: Speed, price, "good enough for 90% of cases" quality.

ByteDance's Seedream 4 is the model you run in a loop when you're iterating fast. It's not trying to beat FLUX on fidelity or Gemini on text rendering. It's trying to be cheap and fast, and it succeeds. You can burn through 20 variations in the time it takes a single premium generation to finish.

Great for: bulk generation, A/B testing concepts, first-pass exploration, social content where speed matters more than a specific look.

How to actually pick

Here's a simple decision tree we use internally:

1. Does it have text in it? → Gemini 3.1.
2. Is it supposed to look photographic? → FLUX 2.
3. Is it an illustration with a specific mood? → GPT-5 (stylized) or Riverflow (painterly).
4. Are you generating 10+ variations and deciding later? → Seedream.

When you genuinely can't decide, a trick we've seen power users love: generate the same prompt across two or three models and compare. This is exactly the kind of thing pay-as-you-go enables and a single-model subscription doesn't.

The cost comparison

Here's what this mix would cost you à la carte, roughly, with separate subscriptions or credit packs:

• DALL·E / ChatGPT Plus for GPT-5-class generation: $20/month.
• Midjourney Basic for stylized work: $10/month.
• Google AI Pro for Gemini 3.1: $20/month.
• FLUX via paid APIs: usage-based, effectively $10–30/month at steady use.
• Seedream / third-party aggregators: $10–30/month.

Add it up: $70–100/month to have all five available, spread across five dashboards, five billing cycles, and five sets of UI quirks.

Nova gives you all five behind one image picker, pay-per-generation, billed against the same token balance you use for chat. The cost per image varies by model — Seedream is cheapest, FLUX and Gemini are mid-range, GPT-5 premium-quality renders cost more — but you only pay when you hit generate.

One dashboard, five personalities

The deeper win isn't just price. It's that you stop picking a model once and start picking it per image.

A full campaign might look like:

• Concept exploration → Seedream (fast, cheap, 30 variations).
• Refine the winner → FLUX 2 (photorealism, precise prompt).
• Hero image with brand text → Gemini 3.1 (readable typography).
• Social illustrations → GPT-5 or Riverflow (stylized, cohesive).

You can't do that fluidly in a subscription world. You can in Nova.

Give it a try — a few generations cost pocket change, and you'll immediately feel why locking yourself to one image model was always the wrong move.