AI Image Generator:
Choose Models, Tune Settings, Edit Results

Secret Chat includes a built-in AI image generator that creates a picture from a description and edits a picture you already have. You can describe a change, remove the background, or upscale the picture, and any generated result can be reworked with one button. You are not tied to one image model, and each generating model brings its own settings.

An empty image chat opens with a welcome message that holds the model picker and settings, in the same place where the first picture will appear. After that message is gone, the sliders button opens the same controls whenever you need them.

For the wider product context (multiple models and local browser storage) see our main private AI chat page.

GPT Image, Grok Imagine, Nano Banana, Qwen, Seedream and Venice are trademarks of their respective owners (OpenAI, xAI, Google, Alibaba, ByteDance and Venice AI). Secret Chat is an independent product and isn't affiliated with, endorsed by, or sponsored by any of these model creators. See our Disclaimers for details.

One tool, several image models

Different image models have different styles and strengths, so the Image Generator lets you choose the one that fits the job. You select a model from the Settings, and your choice is remembered for next time. The models available are:

• GPT Image Flare — OpenAI's fast image model and the default: strong at following detailed instructions and at text inside images, and quick enough to keep trying ideas.
• GPT Image Sunburst — the same family, tuned for precision instead of speed. Slower to draw, and the one to reach for when a picture has to come out exactly right — precise text, an exact layout, a product shot you intend to use.
• Nano Banana — a fast, reliable everyday choice.
• Nano Banana Pro — a higher-end version for more detailed results.
• Grok Imagine — xAI's image model: bold, creative visuals, and the fastest way to get several variants of one idea at once.
• Seedream — an option with fine-grained style controls.
• Qwen Image Pro — another flexible model with extra style controls.

Because each model is a little different, it's worth trying the same idea on more than one to see which look you prefer. Results vary by model and by how you word your description.

Different models, different settings

At a glance, here are the settings each group of models offers:

When you open the Settings, you will see a Model selector at the top and, below it, the options that apply to the model you picked. The settings change to match the model — simpler models show fewer options, while the more advanced ones offer more control. The possible settings are:

• Model — which image model to use. Changing it updates the rest of the settings to match.
• Variants — how many different images to create from the same description at once, so you can pick your favorite. Some models allow a handful, others more.
• Aspect Ratio — the shape of the image, such as square, portrait, or landscape. Choose it to fit where the picture will be used.
• Resolution — how much detail the picture is rendered with: 1K, 2K or 4K, depending on the model. Only the Nano Banana models go to 4K; the GPT Image models and Grok Imagine stop at 2K, deliberately, because at 4K they're slow and expensive out of proportion to the result. Most models start on 1K, the size they produce by default, the quickest and the cheapest in credits. Nano Banana Pro is the exception and starts on 2K, because on that model the larger size costs no extra credits, so opening it any smaller would simply give you less for the same price. Stepping up to 2K or 4K gives you a larger, more detailed picture, takes longer to generate, and costs more credits, so it's worth reserving for images you actually intend to print or use large. Not every model offers every step: Grok Imagine goes up to 2K, and the Qwen and Seedream models have no resolution setting at all — they each render at their own fixed size, which for Seedream is already a large one.
• Image Style — a preset look (offered by some models) that nudges the overall style of the result.
• Negative Prompt — a short note of what you do not want in the image, for example “clouds, rain, snow.” Available on the models that support it.
• Transparent background — asks a GPT Image model for a PNG with alpha instead of a JPEG.
• Preset — a ready job: Avatar and Poster add a brief and choose the matching aspect ratio; Remove background and Upscale ×2 / ×4 switch to the right model and work on the picture you attach.

For example, the Nano Banana models keep things simple with an aspect ratio and a resolution, while models like Qwen Image Pro and Seedream trade the resolution choice for variants, a style preset, and a negative prompt. You don't have to learn all of this up front — sensible defaults are already set (including the resolution, which starts at whatever each model gives you the most for, so nothing costs more until you ask it to), and a Reset option puts everything back if you experiment. The Upscaler is not in this table because it does not generate a new scene and has only one setting: the factor, ×2 or ×4.

Transparent background (PNG)

The Transparent background switch appears on GPT Image Flare and GPT Image Sunburst only. Turn it on and the picture comes back as a PNG with a transparent background instead of a JPEG. The chat shows the transparent area over a checkerboard, and Download keeps the PNG and its alpha channel. Other generating models cannot draw a transparent background and do not show the switch.

In a Smart Agentic chat on ChatGPT, the router makes this choice itself. Ask for a sticker with “no background” and it requests transparency; ask for a scene and it returns a JPEG.

Remove background and upscale with one click

Both live in the Preset field of the image settings (and in the welcome message of an empty image chat): pick Remove background or Upscale ×2 / ×4 and attach the picture with the paperclip — no words needed; a background removal goes out with Send, an upscale starts by itself as soon as the picture has loaded. The same two jobs sit under every generated picture as an eraser and a magnifier button, which pick that preset and run it at once. Remove background asks GPT Image Flare to redraw the picture on a transparent background from a fixed instruction. This is an AI edit rather than a pixel-exact matte, so edges and fine details may change. The result is a transparent PNG.

Upscale sends the picture to the Venice upscaler and returns the same image two or four times larger, with no creative change, always in a chat of its own so the large result is not re-sent with every later message of the chat you were working in. ×4 costs four times as much as ×2. The provider uses a smaller factor when a full ×4 result would exceed about 16.8 megapixels, and the input must be at least 256×256 pixels.

The Upscaler is also the last entry in the model list; picking it opens on Upscale ×2. It reads no text. The AI Image Editor guide covers prompt edits, background removal, and upscaling in detail.

Presets: Avatar and Poster

The Preset field offers None, Avatar, Poster, Remove background and Upscale ×2 / ×4. Avatar sets the aspect ratio to 1:1 and adds a short brief for a head-and-shoulders portrait: centered, on a clean background, and readable as a small profile picture. Poster sets 2:3 and adds a brief for one strong subject, a bold composition, room for a title, and a print-ready result.

Your description is still what gets drawn. The preset brief is placed in front of it on the way to the model, while the chat shows your own words. See the dedicated guides for the AI Avatar Generator and AI Poster Generator.

Edit any image with one button

Every image you generate comes with four buttons beneath it: Download, Edit (the pencil), Remove background (the eraser), and Upscale (the magnifier).

The Edit button is what makes refining a picture easy. Instead of starting over, you click the pencil and the image is placed back into your message box as a starting point. You then simply describe the change you want — “make the sky sunset orange,” “add a small boat,” “remove the text in the corner” — and generate again. The model uses your existing picture as the basis and applies your edit, so you can keep adjusting an image step by step until it looks right.

This image-to-image editing works with the models that support it, and it's the quickest way to fine-tune a result without re-describing the whole scene from scratch.

Tips for better images

• Be specific. Mention the subject, setting, mood, lighting, and style you have in mind: the more detail, the closer the result.
• Use variants. Generating a few at once gives you options to choose from in a single step.
• Set the aspect ratio first so the composition fits where you will use the image.
• Draft small, finish large. Finding the right description takes several attempts, and there's no point paying for extra detail while you're still experimenting. Once a result is close, raise the resolution and generate it again. (On Nano Banana Pro there's nothing to save by drafting below 2K; that step costs no extra there.)
• Add a negative prompt (on models that support it) to keep out things you don't want.
• Refine with Edit. Get close, then use the Edit button to nudge the details rather than starting again.

Private by design

The Image Generator works the same private way as the rest of Secret Chat: it's an anonymizer. Your requests reach the image model with no name, account, or IP attached — the email you register with is used only for account access and payment, never to associate your images with you — and the pictures you create are kept in your browser's local storage on your own device, not a cloud gallery on our servers. The honest limit is the same too: this hides who is asking, not what the prompt says, so your description still reaches the chosen model provider to create the picture — keep personal or confidential details out of it. For more on how this works, see our private AI models page.

One more thing worth knowing when you refine a picture over several messages. Most image models can only be handed a single self-contained description (never a conversation), so a small helper model at OpenAI first rewrites your chat into that one description; without it, a follow-up like “make it blue” would mean nothing to a model that never sees what came before. (The Nano Banana models read the conversation directly and skip this step, as does the very first picture in a chat.) That helper call travels exactly like every other request (anonymously, under our own credentials, never linked to you), and every model that took part in a message, helper included, is listed in that message's Session Privacy Report.

FAQs about the AI Image Generator