# Ideogram 4.5 Prompt Guide: Typography, JSON Layouts and Precise Edits

Canonical page: https://fuser.studio/articles/ideogram-4-5-prompt-guide

How to prompt Ideogram 4.5: quoted text, Expand Prompt, JSON layouts with bounding boxes, and precise edits with masks and references, tested on 17 real runs.

[All guides](https://fuser.studio/articles) · [Ideogram 4.5 in Fuser](https://fuser.studio/models/ideogram-4-5)

**Quick answer:** for generation, put every word you want on the image in quotes, say where each line goes and how it should look, end with "No other text", and switch Expand Prompt off when the exact wording and letter case matter. For layout control, paste an Ideogram JSON prompt with bounding boxes. For edits, name the one change and add "Keep everything else exactly the same." ([Ideogram's own example](https://developer.ideogram.ai/ideogram-api/api-overview)), set Edit Precision to High, and add a mask when the change replaces something large. In our tests Ideogram 4.5 spelled all 31 requested text lines correctly across seven generations, and with High precision 96% of the area no edit touched was still byte-identical after three edits in a row.

## What Ideogram 4.5 is

Ideogram announced Ideogram 4.5 on 30 September 2026 ([launch post](https://x.com/ideogram_ai/status/2105327223431737780)) and calls it "the most precise edit model". Its pitch is multi-turn editing: "With each edit, image models can introduce pixel shifts, color changes, and texture artifacts. Ideogram 4.5 reduces this drift, preserving details across multi-turn edits" ([model page](https://ideogram.ai/models/4.5/)). Generation is the second job: the [API overview](https://developer.ideogram.ai/ideogram-api/api-overview) describes Ideogram 4.5 and 4.0 as "known for prompt fidelity and crisp typography". Ideogram publishes open weights for Ideogram 4, but its [Hugging Face collection](https://huggingface.co/collections/ideogram-ai/ideogram-4) listed no 4.5 weights when we checked on 2 October 2026, so treat 4.5 as a hosted model for now.

In Fuser it is one node, [Ideogram 4.5](https://fuser.studio/models/ideogram-4-5), that switches mode by what you connect:

- **No image connected:** text to image. Controls are Prompt, Quality (Low, Medium, High), Image Size, Expand Prompt and Seed.
- **Image connected:** edit. You also get Mask, up to four Reference Images, Edit Precision (Regular or High; Regular is the default) and a fourth quality tier, Very Low. Expand Prompt is hidden because it only applies to generation.

Medium is the default quality in Fuser for both modes. Ideogram's own API defaults to High when there is no source image ([generate reference](https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5)), so if you are copying settings from Ideogram's docs, set Quality explicitly.

## Write generation prompts like a typesetter

Ideogram 4.5 renders text well when you treat every line as a spec. The prompt we used for the poster below:

> A screen-printed poster for a summer night market. A string of paper lanterns hangs across a deep teal night sky above silhouetted food stalls. Headline across the top in tall condensed orange letters: "NIGHT MARKET". Below it in smaller cream letters: "Fridays in July · 6–11 pm". Along the bottom edge: "Harbour Square · Free entry". Two-colour risograph texture. No other text.

1. **Describe the piece and the scene first:** medium, subject, palette, texture.
2. **Give each line of text its own sentence:** where it sits, its size relative to the others, its colour, then the exact words in double quotes.
3. **Close with "No other text."** None of our seven text generations added stray words.
4. **Type the words in the case you want printed.** See the Expand Prompt section: the rewrite can override it.

Across seven generations (this poster at three quality tiers plus one with Expand Prompt off, a chalkboard menu twice and a JSON banner), every one of the 31 requested lines came out spelled correctly, including the middle dots, en dashes, prices and the apostrophe in "today's". Typeface descriptions were followed more loosely than words: "tall condensed" letters came back as a condensed sans at Low and Medium but as a wide serif at High.

![The same Ideogram 4.5 poster prompt and seed at Low, Medium and High quality; all three spell every text line correctly.](https://statics.fuser.studio/cms/f93401ac-d0bb-41b4-bcec-29fcf5b52580)

_One prompt, seed 4521, Portrait 3:4 (864 × 1152), Expand Prompt on. One run per tier, generated 2 October 2026._

## Low, Medium or High quality

All three tiers spelled every line in our poster test. The differences were texture, detail and composition: even with the same seed, each tier laid the poster out differently, so a seed does not lock a layout across tiers. In Fuser credits, Medium costs about twice Low, and High costs about 3.7 times Medium (about 7 times Low). Edits are priced by the same tiers, and Edit Precision does not change the price. Very Low, available only for edits, costs about a quarter of Low; our Very Low version of the date edit below was spelled correctly and left 89% of the poster's pixels identical.

A practical split: iterate on wording and layout at Low, keep Medium as the working default, and render the version you will ship at High if its texture is worth the price for that piece.

## Expand Prompt: when to switch it off

Expand Prompt switches on prompt expansion, the step Ideogram's API calls magic prompt. When magic prompt is on, the prompt is rewritten and expanded before generation; when it is off, Ideogram "keeps your wording and only converts it into a structured prompt" ([generate reference](https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5)). It is on by default in Fuser.

We ran two prompts both ways with the same seed:

- **Poster footer:** typed as "Harbour Square · Free entry". With Expand Prompt on it printed as HARBOUR SQUARE · FREE ENTRY at all three quality tiers; with it off, the case matched what we typed.
- **Chalkboard menu:** the prompt asked for "white chalk capitals" but quoted the items in mixed case. With Expand Prompt off the board used the quoted mixed case and stayed a flat, straight-on board. With it on, the items came out in capitals and the board was placed in a café with a window and a pendant lamp that the prompt never mentioned.

![Ideogram 4.5 with Expand Prompt off and on: a chalkboard menu and a poster footer, showing how the rewrite changes letter case and scene.](https://statics.fuser.studio/cms/d8ef989f-85ba-4cbd-b640-88a6bc9b37eb)

_Same prompt and seed with Expand Prompt off and on, Medium quality. The rewrite changed letter case in both prompts._

So: switch Expand Prompt **off** when you have written a full prompt and the exact case, wording and scene matter (packaging, menus, brand lines). Leave it **on** for short prompts where you want Ideogram to fill in the scene, and check every text line afterwards.

## Place type exactly with a JSON prompt

Ideogram 4.5's generate endpoint accepts "a structured Ideogram 4.0 JSON prompt" as well as natural language ([generate reference](https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5)). Ideogram documents the format in its [Ideogram 4 prompting guide](https://github.com/ideogram-oss/ideogram4/blob/main/docs/prompting.md):

- `high_level_description`: one or two sentences about the whole image.
- `style_description`: for photographs `aesthetics`, `lighting`, `photo`, `medium`; for everything else `aesthetics`, `lighting`, `medium`, `art_style`; then an optional `color_palette` of up to 16 uppercase hex colours.
- `compositional_deconstruction`: a `background` description and a list of `elements`. Each element is `"type": "obj"` or `"type": "text"`, with an optional `bbox` of `[y_min, x_min, y_max, x_max]` in 0–1000 units from the top-left corner, a `desc`, the literal `text` for text elements, and up to five hex colours of its own.

The guide says key order matters because the model was trained on a fixed order, and that colours must be uppercase `#RRGGBB`. Ideogram's API skips its prompt rewrite for a valid JSON prompt unless magic prompt is explicitly set to on ([generate reference](https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5)); Expand Prompt is on by default in Fuser, so we switched it off. This is the prompt we pasted into the Fuser node's Prompt field with Expand Prompt off (we sent it as a single line; indented here for reading, and still valid JSON):

```
{
 "high_level_description":
  "A wide web banner for an autumn knitwear sale with bold type on the left and a stack of folded sweaters on the right.",
 "style_description": {
  "aesthetics": "clean, warm, editorial, generous whitespace",
  "lighting": "soft diffuse studio light from the upper left",
  "photo": "straight-on product shot, sharp focus",
  "medium": "photograph",
  "color_palette": ["#F3E9DC","#C8553D","#3B2F2F","#E0B589"]
 },
 "compositional_deconstruction": {
  "background": "A seamless warm cream paper backdrop with a faint shadow under the products.",
  "elements": [
   {"type": "text", "bbox": [180, 60, 420, 520],
    "text": "AUTUMN SALE",
    "desc": "Very large bold dark brown condensed sans-serif headline.",
    "color_palette": ["#3B2F2F"]},
   {"type": "text", "bbox": [450, 60, 560, 480],
    "text": "30% off knitwear",
    "desc": "Medium rust-red sans-serif subheading.",
    "color_palette": ["#C8553D"]},
   {"type": "text", "bbox": [700, 60, 800, 300],
    "text": "Ends Sunday",
    "desc": "Small cream text inside a rounded rust-red pill button.",
    "color_palette": ["#C8553D","#F3E9DC"]},
   {"type": "obj", "bbox": [150, 580, 900, 960],
    "desc": "A neat stack of four folded chunky knit sweaters in rust, oatmeal, camel and chocolate brown."}
  ]
 }
}
```

![An Ideogram 4.5 autumn sale banner made from a structured JSON prompt, with the prompt's bounding boxes drawn over the result.](https://statics.fuser.studio/cms/4ccd0e10-a930-49ea-85e5-6329885f79a7)

_A JSON prompt pasted into the Prompt field, Expand Prompt off, Landscape 16:9, Medium. One run._

Every block landed inside its box, the stack had exactly four sweaters in the four listed colours, and the case of "30% off knitwear" was kept. That was a single run, so treat it as a strong start rather than a guarantee; it is the most direct way we found to control where text sits. Write the JSON in a Text node so you can reuse it with different words.

## Edit prompts: name the change, protect the rest

Ideogram's own edit example states the change, then adds "Keep everything else exactly the same." ([API overview](https://developer.ideogram.ai/ideogram-api/api-overview)). We used that pattern on the Medium poster for three turns, feeding each output back in:

1. "Change the date line to "Fridays in August · 6–11 pm". Keep everything else exactly the same."
2. "Make the paper lanterns warm red. Keep everything else exactly the same."
3. "Replace the headline NIGHT MARKET with the French "MARCHÉ DE NUIT" in the same font, colour and position. Keep everything else exactly the same."

Quote any new text, and say which existing element it replaces.

![Three Ideogram 4.5 high-precision edits to one poster: date, lantern colour, then a French headline with and without a mask.](https://statics.fuser.studio/cms/6aa4097e-d364-4f8f-9035-e96061c4f3df)

_Each edit fed the previous output back in. Edit Precision High, Medium quality, Auto size. One run per step._

**Edit Precision decides how much of the image is redrawn.** On High, Ideogram's Precise Edit returns the image at its own width and height and copies "pixels the edit did not meaningfully change" from the source ([precise edit reference](https://developer.ideogram.ai/api-reference/images/precise-edit/ideogram-4-5)). Regular redraws the whole frame, and in Fuser it is the only mode that can change the output size (High and masked edits require Image Size Auto). We ran the same three prompts both ways and compared every pixel with the original:

- **High:** after turn 1, 91% of the poster was byte-identical to the source. After three turns, 96% of the lower half (below the headline block, which no prompt touched) was still identical, and only 6 of its 483,840 pixels had changed by more than about 3%.
- **Regular:** after turn 1, 0.3% of pixels were identical, though the change looked just as clean. After three turns, none of the lower half was identical and 17.7% of it had changed by more than about 3%: the stars, the stall silhouettes and the footer had all been redrawn.

![Pixel difference maps after three Ideogram 4.5 edits: regular precision changes the whole poster, high precision only the edited areas.](https://statics.fuser.studio/cms/97b13b57-d0b3-4369-8f00-af3d65710ecc)

_Red marks pixels that differ from the original poster after the same three edit prompts. Lower half = everything below the headline block._

Use High for any edit you may build on later. Use Regular when you want a new aspect ratio from the edit, and expect the whole image to be re-rendered.

## Use a mask when the change is big

Turn 3 failed on both settings: the new headline was longer than the old one, and part of a letter from NIGHT survived inside the new "MA" (on Regular the word came out garbled, closer to "MRCHÉ"). A mask fixed it in one run. Ideogram's rules for the mask: black marks the area to edit, white the area to keep, it must be the same width and height as the source image, and it must contain both colours ([precise edit reference](https://developer.ideogram.ai/api-reference/images/precise-edit/ideogram-4-5)).

To make one in Fuser, set up a Compositor node at the source image's exact size (864 × 1152 here), fill it white, draw a black rectangle or brush over the area to change, and connect its output to the Ideogram 4.5 node's Mask input. Our mask covered the whole headline block, lanterns included; the result was a clean MARCHÉ DE NUIT with the red lanterns redrawn in place. Below the mask, 93% of pixels were identical to the input; the changes sat in a band just under the mask edge, and below the date line 99.9% were identical. Cover the whole old element with a margin, as we did, so none of it is left for the edit to keep.

## Steer an edit with reference images

Connect up to four images to Reference Images (three when a mask is connected). Ideogram uses them to guide the edit and never edits them ([precise edit reference](https://developer.ideogram.ai/api-reference/images/precise-edit/ideogram-4-5)). Say in the prompt exactly what to take from the reference, and what to keep:

> Change the top sweater in the stack to the navy, mustard and cream Fair Isle knit from the reference image. Keep its shape, fold and position. Keep everything else exactly the same.

![Ideogram 4.5 reference edit: a Fair Isle swatch guides a change to the top sweater in a sale banner while the type stays the same.](https://statics.fuser.studio/cms/6a8b4223-46ac-4d2d-af4c-89d2ad80b0b2)

_One reference image, Edit Precision High, Medium quality. 84% of the banner’s pixels came back identical._

The colours and knit style carried over and the type was untouched, but the motif was simplified rather than copied stitch for stitch. If a pattern must match exactly, a reference alone is not enough; check the result at full size.

## Sizes and what the node does not do

- **Generation sizes:** Square HD and Square (both give 1024 × 1024), Portrait 3:4 (864 × 1152), Landscape 4:3 (1152 × 864), Portrait 9:16 (720 × 1280) and Landscape 16:9 (1280 × 720). Auto falls back to Square HD for generation. Ideogram's API also offers 2K sizes such as 2048 × 2048 ([generate reference](https://developer.ideogram.ai/api-reference/images/generate/ideogram-4-5)), but the Fuser node's presets stop at the sizes above.
- **Edit sizes:** Auto keeps the source's proportions; Ideogram scales images that are too large for the model down ([precise edit reference](https://developer.ideogram.ai/api-reference/images/precise-edit/ideogram-4-5)). All our poster edits came back at exactly 864 × 1152.
- **Larger or exact sizes:** build the canvas in the Compositor for exact pixel dimensions, and run the result through an upscaler such as [Topaz](https://fuser.studio/models/topaz-image-upscale) or [Recraft Upscale](https://fuser.studio/models/recraft-upscale) for print sizes.
- **What the node does not do:** it has no style presets, negative prompt or transparent background, and makes one image per run. Ideogram's API index lists transparent generation for its 4.0 and 3.0 models only ([API reference index](https://developer.ideogram.ai/v2/llms.txt)). For a cut-out, pass the result to [Background Remover](https://fuser.studio/models/background-remover).
- **Seed:** same seed and prompt for repeatable runs at the same quality tier.

## Prompts to copy

- **Event poster (Expand Prompt off):** "A risograph poster for a jazz night in a basement club, warm red and navy. Headline across the top in tall condensed red letters: "LATE SET". Below it in small navy letters: "Thursdays · 9 pm". Along the bottom: "The Cellar · Free entry". No other text."
- **Menu board (Expand Prompt off):** "A café chalkboard menu photographed straight on. Title in yellow script: "Morning Menu". Six lines in white chalk with prices aligned right: "Flat white 3.80", "Oat cortado 3.60" … Small line at the bottom: "Ask about today's bake". No other text."
- **Text change:** "Change the date line to "Saturday 14 June". Keep everything else exactly the same." (Edit Precision High.)
- **Colour change:** "Make the lanterns warm red. Keep everything else exactly the same."
- **Translation:** "Replace the headline "OPEN LATE" with the Spanish "ABIERTO HASTA TARDE" in the same font, colour and position. Keep everything else exactly the same." (Mask over the old headline.)
- **Material from a reference:** "Change the sofa's upholstery to the green velvet from the reference image. Keep its shape and position. Keep everything else exactly the same."

## Fix what goes wrong

- **Letter case changed:** switch Expand Prompt off and type the text in the case you want.
- **The scene has things you did not ask for:** Expand Prompt adds detail; switch it off or describe the full scene yourself.
- **Text in the wrong place:** describe the position in words, or use a JSON prompt with a `bbox` per line.
- **Wrong typeface style:** describe the type, but expect it to drift between quality tiers; lock the tier before refining the type description.
- **Leftover strokes after replacing text:** add a mask that covers the whole old text block, with a margin.
- **Colours or texture drift after several edits:** set Edit Precision to High and keep Image Size on Auto.
- **"Masked and high precision edits require Auto image size":** set Image Size back to Auto, or switch to Regular without a mask if you need a new size.

## Build it as a workflow

Ideogram 4.5 is strongest as the step that finishes an image: make or adjust the base in Fuser, lock its text and layout with Ideogram, then keep editing it turn by turn without re-rendering what already works. Masks come from the Compositor on the same canvas, and the result can go straight on to an upscaler, [Background Remover](https://fuser.studio/models/background-remover) or a video model. For a head-to-head on the same prompts, see [Ideogram 4.5 vs GPT Image](https://fuser.studio/articles/ideogram-4-5-vs-gpt-image); for the full poster process, the [AI poster typography workflow](https://fuser.studio/articles/ai-poster-typography-workflow); and for how it compares with other models on lettering, [the best AI models for text in images](https://fuser.studio/articles/best-ai-models-for-text-in-images) and [the best AI image editing models](https://fuser.studio/articles/best-ai-image-editing-models).

## Ideogram 4.5 cheat sheet.

What to write, and the setting that goes with it in Fuser.

### Generate

| Goal | Write in the prompt | Set on the node |
| --- | --- | --- |
| Exact words | Each line in double quotes, with its position, size and colour. End with "No other text." | Quality: Medium (default) |
| Exact letter case | Type the text exactly as it should print. | Expand Prompt: off |
| Quick idea from a short prompt | Subject, style and the quoted text. | Expand Prompt: on |
| Exact layout | JSON prompt with a bbox for every text block and object. | Expand Prompt: off |
| Cheap drafts | Same prompt; lock wording and layout first. | Quality: Low |

### Edit

| Goal | Write in the prompt | Set on the node |
| --- | --- | --- |
| Small change | "Change X to Y. Keep everything else exactly the same." | Edit Precision: High, Image Size: Auto |
| Replace a large element or long text | Same pattern; name the old element and quote the new text. | Mask over the old element, with a margin |
| Material, product or style | Say what to take from the reference and what to keep. | Up to 4 Reference Images (3 with a mask) |
| New aspect ratio | Describe the change; expect a full redraw. | Edit Precision: Regular, no mask |

## Questions, answered.

### What is Ideogram 4.5?

An image generation and editing model Ideogram announced on 30 September 2026. Ideogram calls it the most precise edit model and says it reduces the pixel shifts, colour changes and texture artifacts that build up over repeated edits. In Fuser one Ideogram 4.5 node generates from text, and edits when you connect an image.

### Is Ideogram 4.5 good at text in images?

In our tests it spelled all 31 requested text lines correctly across seven generations, including prices, punctuation and an apostrophe. Typeface descriptions were followed more loosely than the words themselves, and Expand Prompt sometimes changed letter case.

### What does Expand Prompt do on Ideogram 4.5?

It is Ideogram's magic prompt: the prompt is rewritten and expanded before generation. It is on by default in Fuser and only applies to generation. In our runs it changed mixed-case text to capitals and added scene details, so switch it off when exact wording and case matter.

### Can Ideogram 4.5 follow a JSON prompt?

Yes. Ideogram's API accepts a structured Ideogram 4.0 JSON prompt with bounding boxes for each text block and object. In our run a JSON prompt pasted into the Fuser node with Expand Prompt off put every element inside its box.

### What is the difference between Regular and High edit precision?

High copies unchanged pixels from the source and keeps its size; Regular redraws the whole image and can change the output size. After three edits to one poster, High left 96% of the untouched area identical; Regular left none of it identical. Both cost the same.

### Are Ideogram 4.5 weights available?

Ideogram publishes open weights for Ideogram 4, but its Hugging Face collection listed no Ideogram 4.5 weights when we checked on 2 October 2026. In Fuser it runs as a hosted model.

## Set type and edit it on one canvas.

Generate with Ideogram 4.5, mask in the Compositor and keep editing without losing what works.

[Open Ideogram 4.5 in Fuser](https://fuser.studio/models/ideogram-4-5) · [Explore all guides](https://fuser.studio/articles)

## More articles

- [Best AI Models for Text in Images (2026)](https://fuser.studio/articles/best-ai-models-for-text-in-images.md)
- [FLUX.2 Prompt Guide: Structure, JSON Prompts, Hex Colours and References](https://fuser.studio/articles/flux-2-prompt-guide.md)
- [Hunyuan3D Guide: Image to 3D, Low Poly Topology and Multiview](https://fuser.studio/articles/hunyuan3d-guide.md)