Image models for slides in 2026: Nano Banana 2, GPT Image 2.5, FLUX 2 Pro, Seedream 5 and Recraft, and which to use for what

A year ago we wrote about Nano Banana, Google's image model that finally kept a subject consistent from one image to the next. That post is still one of the most read on this blog, because consistency is the thing a deck needs: the same product, the same character, the same style on slide three and slide thirty.

Since then the field has changed again. This is the September 2026 guide: what each current model is good at when the output is going on a slide, which ones our products offer, and how to choose without reading a benchmark.

What matters on a slide

Image models are usually judged on a single beautiful picture. A presentation asks different questions:

• Consistency. Can it produce a series that looks like one series?
• Text. Can it render a label, a number or a headline inside the image without gibberish?
• Brand fit. Can it hold a palette and a style you specify, across many images?
• Editing. Can it change one thing in an existing image and leave the rest alone?
• Speed and cost. A deck may need twenty images, not one.

The models, and what they are for

Nano Banana 2 and Nano Banana Pro (Google). The successors to the model that made consistency practical. Still the reference point for keeping a subject stable across a series and for edits that respect what is already in the picture. Nano Banana 2 is the everyday choice; Pro is for the hero image where detail matters.

GPT Image 2.5 (OpenAI). The strongest of the current models at putting readable text inside an image, which makes it the one to reach for when the image itself carries a label, a diagram caption or a number. It also follows long, specific prompts well, so a detailed brief about layout tends to come back as asked.

FLUX 2 Pro (Black Forest Labs). Photographic realism and lighting. If the slide needs something that reads as a photograph rather than an illustration, this is the model. Black Forest Labs shipped FLUX 3 in July; FLUX 2 Pro remains the version most tools run today.

Seedream 5 Pro (ByteDance). Strong on stylised and illustrative work and on aesthetic composition. A good choice for cover slides and section dividers where mood matters more than accuracy.

Recraft 4.1 (Recraft). Built for design work: vector-like output, clean shapes, strict adherence to a palette. When a deck needs icons, flat illustrations or anything that has to sit next to a brand kit without clashing, Recraft is the one that behaves.

Qwen Image 2.0 (Alibaba). Fast, inexpensive and good enough for most in-body illustrations. It is the model behind image generation on the NextDocs free plan.

Which models NextDocs and Shyne offer

NextDocs generates and edits images from the media panel. The free plan uses Qwen Image 2.0. Pro uses Nano Banana, and Pro+ and Ultra use Nano Banana 2. The panel handles both generation from a prompt and editing an image you already have, and the results drop straight into the document.

Shyne offers the wider set: GPT Image 2.5 is the default, with Nano Banana 2, Nano Banana Pro, FLUX 2 Pro, Seedream 5 Pro and Recraft 4.1 available on paid plans, each with an editing counterpart, and image-to-video with Veo 3.1 or Kling 3 when a still needs to move. In Shyne you ask the agent for an image rather than opening a panel: "make me a hero image for slide one in the brand palette, photographic, with the product on a desk", and it appears on the canvas where you can see it, then gets refined in the same conversation.

A simple rule for picking

The slide needs	Reach for
The same subject across many slides	Nano Banana 2
Text inside the image	GPT Image 2.5
Something that looks like a photograph	FLUX 2 Pro
Mood, atmosphere, a cover	Seedream 5 Pro
Icons, flat illustration, strict brand colours	Recraft 4.1
Twenty quick in-body illustrations	Qwen Image 2.0 or Nano Banana 2

Two habits that matter more than the model

Describe the series, not the image. "A set of six illustrations in the same flat style, muted violet palette, each showing one step of onboarding" gets you a consistent set from any of the models above. Six separate prompts get you six styles.

Edit instead of regenerating. All of these models can now change one element in an existing image. If slide four's image is right except for the colour of the chair, ask for the chair, not a new image. It is faster, cheaper and it keeps the rest of the deck consistent.

Image models will keep changing every few months. The habits do not.

Generate images in NextDocs · Ask the Shyne agent for one

---

The NextDocs Team