Image Generation (Beyond Flyers)
The current image models (ChatGPT Images 2.5, Nano Banana 2 and Pro, Midjourney V8.1 and others), which to use for what, and how to keep images accurate to Chabad practice.
On this page
This page covers AI images in general: illustrations, infographics, social posts, logos, renders and photo edits. Event flyers have their own page, Flyer Design.
The image models, October 2026
| Model | Where you use it | What to know |
|---|---|---|
| ChatGPT Images 2.5 | ChatGPT (all plans), ChatGPT Work, Codex | Released 2026-09-08. OpenAI says it gives sharper detail, keeps the people in your reference photos more faithfully, follows edits more reliably over several rounds, and runs up to 50% faster than Images 2.0. Adds a Sketch tool for drawing a rough layout in the chat (9to5Mac). |
| ChatGPT Images 2.0 (gpt-image-2) | Replaced by 2.5 | Released 2026-04-21. Added 2K output, more aspect ratios, a "thinking" mode that can research before drawing, and better non-Latin scripts. |
| Nano Banana 2 | Gemini app (default), Google AI Studio, Google Search | Released 2026-02-26. Fast, uses web search to get real subjects right, outputs up to 4K. See Nano Banana. |
| Nano Banana Pro | Gemini app (AI Pro and Ultra plans), AI Studio, NotebookLM slides and infographics | Released November 2025. Google's highest-fidelity image model. |
| Midjourney V8.1 | midjourney.com | Released 2026-04-30, according to Midjourney's model documentation. Strong for artistic illustration. The group rarely discusses it. |
| Adobe Firefly | firefly.adobe.com | Adobe's own models plus partner models, including Google's Nano Banana and OpenAI's image models, in one app. Best known in the group for generative fill on an existing photo. |
Claude does not generate images. As of October 2026 members still confirm this. Use Claude to write the image prompt, or use Claude Design to lay out a design from your own photos and art.
Two others come up in the chats as community experience only: Grok, which members used in 2025 for quick batches of variations and photo edits, and Meta AI in WhatsApp, which one member said in October 2026 "is great for flyers."
Which one for what
From the group's experience, not a lab benchmark:
| Job | What members use |
|---|---|
| Ads and flyers you will edit several times | ChatGPT. Since January 2026 the admin has preferred it because it keeps quality through edits. |
| Realistic backgrounds, renders of a real building or room | Nano Banana (Gemini). Upload a real photo for accuracy. |
| Infographics from an article or source | ChatGPT ("Make a bold infographic for...", December 2025) or Gemini; NotebookLM for a source you have uploaded. See How do I turn a photo of a source into a clear infographic? |
| Logos | Several rounds: brainstorm, then prompt, then image. In 2025 members found Nano Banana better than ChatGPT. For a clean transparent logo, Canva's background remover and upscaler. |
| Editing a real photo (remove an object, fix a detail) | Canva Magic Eraser, Google Photos editing, or Adobe Firefly. Not a generator, which redraws faces. |
| Collages of many faces | Canva. Image generators regenerate the faces rather than pasting them (September 2026). |
| Comparing several models on one prompt | OpenArt, which runs several image models side by side (February 2026). |
| Upscaling for print | Upscayl (free), Upscale.media, or Canva's upscaler. |
The safe habit, given how often the lead changes: run your prompt in both ChatGPT and Gemini and keep the better result.
Getting images right for a Chabad House
Image models default to generic religious imagery. Every image on a shliach's page or flyer should show what a Chabad House actually does, and someone who knows the practice should look at it before it goes out. The common corrections:
- Shabbos and Yom Tov candles: the mother's candles, plus a smaller candle for each young girl, who lights her own from about age three. In February 2026 a shliach reminded the group that the Rebbe was particular that Neshek publicity show a smaller candle with the two larger ones. AI usually draws two.
- Menorah: straight diagonal branches as in the Rambam's drawing, not rounded arms; eight lights and a shamash for Chanukah.
- Luchos: square tops.
- Lulav and esrog: the lulav bound with five rings of lulav leaf (the Chabad custom), never a woven holder; the esrog held pitam up.
- Tefillin: models draw them wrong. Give a real photo and ask for a prompt that reproduces it (August 2026).
- People: add tznius instructions to edit prompts ("boys with kippah and tzitzis, modest dress"), and check that every figure in a group scene matches (February 2026).
For describing people and scenes accurately, see Prompting AI Image Tools to Depict Your Community.
Consistent images across a series
The JSON method, repeated in the groups since 2025:
- Give a reference image to ChatGPT or Gemini.
- Ask it to describe the image as JSON: every visible detail, the style, the colors, the composition.
- Paste that JSON as the starting prompt for each new image in the series.
Members also simply upload the same style images with every request (May 2026). In May 2026 the same JSON trick was described for video: describe a reference video in detail as JSON, then paste it into Gemini's video tool.
Let one AI write the prompt for another
A tip that keeps coming back. In February 2026 a member's version:
Help me write a great prompt for nano-banana for making a flyer for [EVENT]. Ask me clarifying questions first. Then write the prompt, audit it against what I asked for, and improve it before giving it to me.
The same idea works for any image tool: tell the chatbot your goal and ask, "What is the best prompt for this for [tool]?"
Keeping a real face or photo unchanged
"What's the best prompt to have ChatGPT not change the photos when inserting into a pamphlet/flyer? Sometimes it listens, other times it changes the faces."
The answer that worked: "Just take the image, draw on it, and circle the area that should remain unchanged." Since September 2026 ChatGPT also has a Sketch tool for drawing a rough reference right in the chat. Another tip (April 2026): give the real photo again at the end and say "replace with real photo and be accurate to the detail." When accuracy matters, place the real photo yourself in Canva.
A Grok trick from December 2025: to fix a photo where one student was looking the wrong way, turn the photo into a short video and screenshot a good frame.
Maps
To get an accurate map in an image, give the model an image of the real map (the admin, June 2026). Asked to draw a city map from memory, the model makes a generic one.
Related
- Flyer Design
- Nano Banana
- ChatGPT
- Canva
- AI Flyer Aesthetics, for the tells that apply to every AI image.
- Hebrew & Yiddish, for Hebrew inside images.