Kosher AI image generator - modest pictures for the classroom and home

A kosher AI image generator has to do one thing a general image tool does not: produce a picture that can go on a parsha sheet, a classroom wall or a shul newsletter without anyone checking it twice for tznius. This is a practical guide to how the Image tool in Navōn handles that, and how to ask it for something you can actually use.

The deeper question, why a modesty standard is harder to enforce in a picture than in a sentence, is covered in our longer piece, Tznius and AI image generation. This one is about what happens when you press the button.

Modest AI images by default

Before your request reaches the image model, Navōn puts a set of instructions in front of it. Every person in the picture is to be fully and modestly dressed: sleeves past the elbow, a high neckline, skirts below the knee, nothing tight, no swimwear. Men and boys wear shirts, and any shorts reach the knee. Beach and pool scenes are ruled out. If a scene cannot be drawn within those limits, the model is told to change the scene, not the standard.

There is a second default, about who appears. If you ask for a family at the Shabbos table and say nothing more, the people drawn will be observant Orthodox Jews: a kippah and tzitzis on the men and boys, hair covered on married women, sefarim on the shelf and a mezuzah on the doorpost. If your request names someone else, an Egyptian taskmaster or a Roman soldier, that default steps aside and you get what you asked for.

When you set up your account you are asked which community your household belongs to: Ashkenazi, Sephardi, Litvish, Chassidish, Modern Orthodox, or a few words of your own. That answer shapes how people are dressed in your pictures, so a Chassidish household is not shown a Modern Orthodox family, and the reverse.

What a kosher AI image generator checks: the prompt and the picture

Instructions are a request, not a guarantee. Image models sometimes ignore them. So Navōn checks twice.

• Before anything is drawn, your prompt goes through the same checks as a chat message: the platform's own content rules, plus any topics a parent or school has blocked for your account. A request that fails is refused with a short reason.
• After the picture is drawn, a separate model looks at the image itself: how any people are dressed, what is in the background, and any text visible in it. A picture that fails is withheld before it is saved, so there is no copy in your library and no link to share.
• If that second check cannot run, because a service is down or slow, the picture is withheld as well. Navōn fails closed: it would rather show you nothing than show you something it did not look at.

The same checks run on edits. Cutting out a background, extending a scene or turning a photo into a painting all produce a new image, and each one is scanned before you see it.

None of this is perfect. A check made by a model can miss something, and it can also withhold a picture that was fine. If an image is withheld that should not have been, rewording the request usually solves it. If something comes through that should not have, use the Report option on the picture, because that is how the rules get better.

Asking well for a parsha sheet or class poster

Most disappointing pictures come from a request that leaves out what you were picturing. A few habits help.

Describe the scene, not the topic

"Parshas Noach" gives the model almost nothing to work with. "The teivah resting on a mountaintop after the flood, a dove flying back with an olive leaf in its beak, a rainbow across a clearing sky" gives it a picture. Name the objects, the setting, the time of day and the mood.

Choose a style and a shape

The Image tool offers styles, among them Illustration, Watercolour and Storybook, and shapes: Square, Wide, Tall, Classic and Portrait. Storybook in Portrait suits the top of a parsha sheet. Wide suits a banner across a bulletin board. Leave the style on As written if your prompt already says how the picture should look.

Say who is in it, and how they are dressed

The modesty instructions are always there, but it costs nothing to be explicit, and explicit prompts come back closer to what you meant. For example:

• "Three boys in white shirts, dark trousers and black kippot, sitting around a table learning from a large sefer, warm afternoon light, storybook illustration."
• "A bochur in a dark suit and hat blowing a shofar in a shul, seen from the side, watercolour."
• "An elderly man with a white beard, in a long dark coat and hat, lighting a Chanukah menorah in a window at night."

If your school's practice is not to show the faces of the Avos or the neviim, say so in the prompt. "Seen from behind" or "a figure in the distance, face not shown" works better than hoping. Often the simplest answer is to leave people out and draw the objects instead: the luchos, the menorah of the Mishkan, the tents of the camp around it.

Leave the lettering to yourself

Image models are still unreliable at drawing words, and Hebrew letters in particular tend to come out malformed. Ask for the picture with empty space at the top or bottom, then add the title and the posuk yourself in the document. It looks better, and you know it is spelled right.

Ask more than once and choose

The same prompt gives a different picture each time. Two or three tries and one good choice is usually quicker than rewriting a long prompt five times.

Settings a parent or school can change

On a child's account, a parent or school can turn the creative tools off entirely, which covers images, video and stories. The child sees a short note that the tool has been turned off, and the rest of Navōn works as before. Blocked topics apply to the Image tool exactly as they do in chat, so a subject a family has closed is closed in pictures too.

That switch is worth considering for younger children. Trying to list in advance every way a picture could go wrong is a losing game. Turning the tool off until a child is older is simple, and it can be turned back on later.