Best AI image generator, and how it works inside a chat

"Best AI image generator" is nearly always answered with a ranking, and a ranking is the least useful part of the answer for anyone who types the request into a chat window instead of a dedicated art tool. Several parts of the process affect the result: which model reads your words, who writes the prompt that actually reaches the image model, which dials are exposed to you, and how many attempts a turn buys you. Get that wrong and the best model in the world produces something you did not ask for.

So this is a description of the machinery. MultiChats runs seven image models from three makers, and the way they are wired into a conversation explains most of the surprises people hit on their first afternoon of trying.

What happens between your sentence and the picture

Type your request and send it. There is nothing to switch on first, because Intent Detection is on by default for every account and reads each message to decide whether to search the web, generate an image, or adjust reasoning. Your message goes to whichever chat model you have selected. That model reads the request, writes an image prompt in its own words, and calls a tool. The tool passes that prompt, along with any earlier images it has been asked to work from, to one of the seven image models, which returns the image bytes. We store the result in your account and place it in the thread.

Three consequences fall out of that sequence.

First, the chat model is acting as a prompt writer. What you type is an instruction to it, and the string that arrives at the image model is its reading of what you meant. If you write two words, it invents the other forty. That is usually fine and occasionally the whole problem, so the most effective habit is asking the model to show you the prompt it intends to send.

Second, the image model receives that prompt and, when you ask for a change, the images it should work from. It never sees the rest of the conversation. Since 12 October 2026 the tool that carries the request has two fields: the prompt, and a list of images already in the thread, which can be a photo you attached or a picture generated earlier. Grok Imagine takes one such image per request; the other six take up to three.

Third, and this is the part that catches people: once an image lands, the conversation carries a short written description of it in place of the image itself. The chat model is instructed to answer as though it were looking at the picture, and it is working from that description. Ask it to "make the coat darker" and it can hand the earlier picture back to the image model with a prompt describing the change, so the edit starts from that exact image. The chat model decides when to do that, so if a follow-up comes back as a different picture altogether, ask it plainly to edit the previous image.

Because a chat model stands between your sentence and the image model, you are briefing a writer who will fill in every gap you leave, and who has no way to ask you a question first. Specify the details the model would otherwise have to invent: the subject and what it is doing, the framing and distance, the light, the medium you want it to imitate, and the things that must not appear. Vague adjectives get replaced with the model's defaults, and the defaults are dull by construction.

When a result misses, a reply like "same but with a red door" now has the picture itself to start from. The prompt behind the edit is still written from the chat model's own words, though, so if details you liked start to drift, restate the whole picture with the one change folded in.

One more property of a turn is worth knowing before you start. A request produces a single image. There is no count to raise anywhere, because the tool carries a prompt and a list of images to work from, and no number. Ask for four variations and you get one picture, so four variations means four turns.

Three makers sit behind the picker

Seven models: two from OpenAI, three from Google and two from xAI. The controls are not the same on each, because the makers do not expose the same parameters, and that difference is more likely to shape your output than any quality argument between them.

The OpenAI pair is the only place a quality control appears, and OpenAI is also the only maker that asks for output shape in literal pixel dimensions where the other two ask for a ratio. If you need something to land at an exact pixel size, that is where to look.

The Google names come from Google. Its own image generation documentation says that "Nano Banana is the name for Gemini's native image generation capabilities" and, when we checked on 19 October 2026, listed five models under that banner, of which we carry three. Read that page next to our picker and one difference stands out: Google documents at least ten aspect ratios for every model in the family, and we expose five. The narrower set is a deliberate choice, so that one picker reads the same across three makers whose parameter vocabularies do not line up.

On xAI, quality comes with the model. Grok Imagine and Grok Imagine 2.0 are separate entries in the picker, and choosing between them is how you choose the fidelity. Anyone hunting for a quality dropdown there will hunt for a while, because the decision was already made one step earlier.

Shape is the setting that trips up almost everyone, whichever maker you are on. Aspect ratio lives in the image configuration control, where it stays until you move it. On the mobile apps that control sits in the composer. On web it arrives with the image button, which is there once you switch Intent Detection off in Settings, then Tool Preferences, and switching it off is also how you decide per message instead of letting the app decide. Writing "make it wide" leaves the ratio exactly where it was and merely tells the prompt writer you would like a wide-feeling composition inside the same frame, which is how you end up with a letterboxed subject floating in a square. Set 16:9 for a slide or 9:16 for a phone screen, then spend your sentences on the picture.

Quotas, plainly stated

On the free plan you get two image generations every 30 days, on a fixed window that does not reset with the calendar month, and they come from Google Nano Banana. The picker's aspect ratio and quality choices do not apply to that path. Signed-out visitors get none at all: image generation is off for anonymous requests, with no exception. And on the mobile apps, image generation sits behind a paid plan at the button, so the free allowance is a web-browser affair.

The counter is more forgiving than it looks, and this is the kind of detail nobody puts on a pricing page. The check that runs before your message is a peek and does not spend anything. The counter only moves after an image has actually been produced and saved. A refusal, a provider error, or a request you cancel halfway through leaves your two intact.

Paid plans open the full picker: any of the seven models, the aspect ratios each maker supports, and the quality dial where OpenAI offers one. You can set a preferred image model once in Settings and it follows you from chat model to chat model, or override it per message from that same configuration control. There is no separate monthly image cap on the paid plans, though every image draws on your plan's usage allowance, the same one that bounds how many turns you take. Pricing sits on the plans section of the homepage.

Two of the chat models in the catalogue, Llama 4 Maverick and Nemotron 3 Super, cannot call tools at all. Image generation runs as a tool, so it is unavailable while either of those is your selected model. Everything you generate is stored in your account and downloads as a PNG. Images made inside a temporary chat are written to a storage area that expires with the chat, so they leave when the conversation does.

Does the image carry an invisible marker?

This question arrives more often than it used to, and it deserves a straight answer. Whether a generated image carries provenance metadata, an invisible watermark, or any other embedded signal is decided by the provider that generated it, and it can vary by model and change without notice. We are not the party making that decision. What we can tell you is what we do with the bytes: the image comes back from the provider, gets decoded, and is written to storage as it arrived. There is no re-encoding step, nothing stripped out of the file, and no marking of our own added on top.

We will not tell you your image is unmarked, because we do not know that and neither does anyone else who is guessing. If provenance matters for what you are doing, check the current documentation of the maker whose model you selected.

A related question with a different answer is whether a provider trains on what you send it. We do not train on your conversations. What the provider behind a given chat model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker. Image models sit outside that arrangement entirely: the image registry has no field for such a notice, so no image model carries one and there is no card to go and read. For image prompts the pointer is the provider's own terms and our privacy policy, which says plainly that data handling practices, including whether your data may be used for model training, vary between providers. Two limits apply everywhere regardless: no sexually explicit content, and no synthetic media depicting a real, identifiable person without their consent. Both sit in the terms and are enforced in the prompt the models run under.

So which one is best

Best depends on which constraint you are working against. Need an exact pixel size, or a quality setting you can turn down? The OpenAI pair is the only place either of those exists. Working to a 9:16 phone screen or a 16:9 slide? Ratios are the native vocabulary on Google and xAI, so the fit needs no arithmetic. On the free plan the choice is already made for you: your two attempts run on Google Nano Banana whatever the picker is showing.

What a ranking cannot capture is the thing that makes any of this worth doing in a chat window at all. The picture arrives inside the conversation that produced it, next to the brief you wrote and the argument you had with yourself about what the thing should look like, and you can keep going from there. The limits worth planning around are how many turns you want to spend circling in on what you meant, and how much of the brief survives from one turn to the next.