Ideogram: The Essentials

Covers Ideogram V4, Ideogram V4 Image to Image, Ideogram V4 Tiling, Ideogram V3 Generate Transparent, and Ideogram V3 Layerize Text

The Ideogram V4 family on Scenario is built around one signature: typography that actually renders. The base text-to-image model puts spell-perfect words inside posters, menus, packaging, signage, UI, and brand surfaces. Two specialized siblings extend it: Image to Image restyles an existing frame from a text prompt while keeping the composition intact, and Tiling generates seamless textures and patterns that repeat perfectly across any surface. All three share the same speed tiers, the same Magic Prompt expansion control, and the same multilingual typography engine.

Two more tools round out the family for design production work: Generate Transparent creates new images with a native alpha channel, so no background-removal step is needed, and Layerize Text takes an existing image with text and removes it, returning a clean background ready for new copy.

Which Model Should I Use?

Use the base V4 when the deliverable is a one-shot graphic with text on it. Use Image to Image when you already have a source frame and want to swap the style, mood, or medium without losing the subject. Use Tiling when the result must wrap, like wallpaper, fabric prints, game-ready ground textures, brand surfaces, and material libraries. Use Generate Transparent when the deliverable needs a built-in transparent background (logos, icons, stickers). Use Layerize Text when you need to strip text out of an existing design to reuse the background.

How to Use the Models

Ideogram V4 (Text to Image)

The base V4 model excels at design work that needs accurate text rendering. The minimum call is just a prompt:

prompt:         "A vintage travel poster for TOKYO, JAPAN, 1950s midcentury modern style"
imageSize:      "square_hd"       // default, 1024x1024
renderingSpeed: "BALANCED"        // default
expansionModel: "Medium"          // default Magic Prompt

The signature feature is in-image typography. Quote the literal words you want rendered. "FRESH BREAD" beats "the words fresh bread": Ideogram's text engine treats quoted strings as the literal type to render. Describe placement and hierarchy alongside the quoted text: which line is the headline, which is a tagline, where on the canvas they sit.

Three speeds trade time for letterform sharpness:

• TURBO: fastest. Right for thumbnails and rapid iteration when typography is short and bold.

• BALANCED: default. The right middle ground for everyday work.

• QUALITY: sharpest letterforms and texture. Use for final delivery, multilingual scripts (Japanese, Arabic, Cyrillic), and intricate layouts.

Turn Magic Prompt (expansionModel) off when literal output matching your wording is essential. Magic Prompt set to Medium or Large can introduce stylistic detail you did not ask for, including stray decorative typography. For final spelling verification on the live model page, leave expansionModel: None.

Six aspect ratios are available: square_hd (1024 by 1024, default), square (512 by 512), portrait_16_9 (576 by 1024), portrait_4_3 (768 by 1024), landscape_16_9 (1024 by 576), landscape_4_3 (1024 by 768). Pick portrait for social, landscape for banners and key art, square for icons and avatars.

Featured Examples

The pinned examples on the Ideogram V4 model page show the text engine across formats. The comic splash below renders "THIS ENDS TONIGHT!" inside a jagged speech bubble plus "SECTOR 9" in a caption box, all baked into the artwork at QUALITY rendering:

Retro print ad, landscape 4:3

Jazz night flyer, portrait 9:16

Tarot card, portrait 3:4

Metal wolf badge, square HD

More on the Ideogram V4 model page: 39 pinned examples spanning posters, brand identity, signage, character cards, retro print, and game key art.

Ideogram V4 Image to Image

The Image to Image variant takes a source image and a prompt, then restyles the source while keeping the underlying composition. The minimum call is a prompt plus a source:

image:    asset_xxx
prompt:   "Transform into a pencil sketch on cream paper with cross-hatching"
strength: 0.85

Below, a photoreal perfume bottle becomes a soft graphite study at strength: 0.85. The bottle's silhouette and gold liquid level stay where they were; the medium and surface treatment swap entirely:

Click to open asset_Fff9MQcbTRYhLkseEniUwmXz on Scenario.

Tuning the Strength Dial

strength is the single most important control. It runs from 0 to 1 and decides how far the result departs from the source:

• 0.3 to 0.5: subtle. Lighting, color grading, mood adjustments. The subject is preserved.

• 0.55 to 0.7: moderate. Style and medium swap clearly, the subject's shape and pose stay recognizable. The "CITY STYLE" magazine cover example below used strength: 0.55.

• 0.75 to 0.95: dramatic. Pose is preserved but the medium, palette, and fine geometry get reinvented. Use for medium swaps that turn a photograph into a print or painting.

Default is 0.8. Lower it when the source must remain identifiable; raise it when you want the model to commit to the new look.

Re-rendering Text in Images

Ideogram's text rendering carries through Image to Image. Quote text inside the prompt and the model bakes it cleanly into the restyle. Below: a watercolor postcard with "VAN LIFE" hand-lettered in the sky, and a magazine cover restyle with "CITY STYLE" set across the top. Lower the strength (around 0.55 to 0.65) on text-heavy restyles so the model has room to draw the type without warping the underlying scene:

"T" magazine cover, strength: 0.55

Featured Examples:

Eight pinned examples from the model page, each linking back to the source asset on Scenario. Same composition, different medium:

Dancer as Y2K rave poster

Red dress as editorial pencil

More pinned restyles live on the Ideogram V4 Image to Image model page.

Ideogram V4 Tiling

The Tiling variant generates seamless textures and patterns that repeat perfectly across any surface. The simplest call needs nothing more than a prompt:

prompt:      "Hand-painted gouache polka dots, mustard yellow on cream"
tilingMode:  "both"          // default, repeats in every direction
imageSize:   "square_hd"     // default, 1024x1024

To prove a tile is really seamless, paint it as a CSS background and watch the seams disappear. The block below is one Art Deco tile repeated. Click the tile name underneath to open the original asset on Scenario:

Art deco gold tile (asset_GEqb7PNntRpKVnV76pvVUAeV) repeated 4 by 2 at 200px. No seams visible.

Choosing the Tiling Mode

The tilingMode input decides which axes repeat:

• both (default): repeats horizontally AND vertically. Wallpaper, fabric prints, ground textures.

• horizontal: repeats left-to-right only. Borders, banner backgrounds, marble counter strips.

• vertical: repeats top-to-bottom only. Vertical wood plank, column patterns, stripe materials.

Featured Examples:

Moroccan zellige mosaic

Damask navy and gold

Otomi embroidery

More pinned tiles on the Ideogram V4 Tiling model page.

Optional Image-to-Tile

Pass an image to the image input and Ideogram will convert your reference into a seamless tile rather than starting from scratch. The strength dial controls how far the tile departs from the source: low values (around 0.3) tile the original closely, high values (0.8 to 0.95) treat it as inspiration. Combine with a mask (black-and-white image) to redo only specific regions of an existing tile.

Ideogram V3 Generate Transparent

Generate Transparent runs the Ideogram V3 image model with transparency enabled at the generation level. The alpha channel is produced natively during generation, not by post-processing, so the cutout is clean and part of the original render rather than an approximation applied afterward. The minimum call is a prompt:

prompt:         "Coffee shop logo with 'BREW' text, vintage badge style, warm brown and cream tones, circular design"
renderingSpeed: "BALANCED"   // default
expandPrompt:   true          // default, Magic Prompt

Rendering Speed controls both cost and output quality: QUALITY uses more diffusion steps for detailed output, FLASH uses fewer steps for the fastest, lowest-cost result, and BALANCED is a good default for most production work. Disable Expand Prompt (Magic Prompt) when the output needs to match a specific brief exactly.

Ideogram V3 Layerize Text

Layerize Text analyzes an image, identifies which elements are text, and removes them, returning the background as a clean PNG. There is no mask to draw: the model reads the full image and infers what to remove automatically. It works best on images with clearly readable, typeset text on a distinct background (ad creatives, posters, product labels, game UI); heavily stylized or hand lettering may be removed only partially.

A common pipeline: use Generate Transparent to produce a base visual with a headline and CTA, feed that image into Layerize Text to strip the text, then bring the clean background into Figma, Photoshop, or another compositing tool and place new copy on top, in any language.

Parameters

Shared Across the Family

prompt

Required, up to 2048 characters. Describe the subject, style, mood, and any text that should appear. Quote literal strings for in-image typography.

renderingSpeed

One of TURBO, BALANCED (default), or QUALITY. Trades speed for letterform sharpness and texture detail.

expansionModel

One of None, Medium (default), or Large. Whether Ideogram's Magic Prompt enriches your input with extra style detail. None obeys your prompt exactly; Large is the most descriptive.

acceleration

One of none (default), low, regular, high. Extra speed boost on top of renderingSpeed. Keep at none for dense typography and high-detail renders.

numOutputs

Range 1 to 4, default 1. How many variations to generate from the same prompt.

seed

0 to 2147483647. Reusing the same seed and settings reproduces the same image. Leave empty for fresh variation.

Ideogram V4 (txt2img)

imageSize

One of square_hd (default, 1024 by 1024), square (512 by 512), portrait_16_9 (576 by 1024), portrait_4_3 (768 by 1024), landscape_16_9 (1024 by 576), landscape_4_3 (1024 by 768). Match the dimension to the publication format.

Ideogram V4 Image to Image

image

Required. Source image asset on Scenario.

strength

Range 0 to 1, default 0.8. How far the result departs from the source. The most important dial in the model.

imageSize

Default auto matches the source dimensions. Override with any of the seven shaped sizes when the restyle should reshape the canvas at the same time.

Ideogram V4 Tiling

image (optional)

Optional starting image. When supplied, the model converts it into a seamless tile.

strength

Range 0 to 1, default 0.8. How far the tile departs from the input image.

mask

Black-and-white image marking regions to regenerate (white) versus keep (black). Requires an input image.

tilingMode

One of both (default), horizontal, or vertical. Which axes repeat seamlessly.

imageSize

Default square_hd. Six sizes available; portrait and landscape for directional patterns.

Ideogram V3 Generate Transparent

prompt

Required, up to 2048 characters. Describe the subject, style, and color; the alpha channel is added automatically.

negativePrompt

Optional, up to 2048 characters. Elements to exclude, useful for avoiding specific styles, colors, or content types.

renderingSpeed

One of FLASH, TURBO, BALANCED (default), or QUALITY. Trades cost and speed for output detail.

aspectRatio

Default 1:1. Fifteen presets from 1:3 to 3:1, covering Stories, banners, and standard formats.

expandPrompt

Default true. Runs Magic Prompt to expand your description before generation; disable when prompt precision matters more than richness.

imageCount

1 to 8, default 1. Number of images to generate per job.

seed

0 to 2,147,483,647, default random. Fixed seed for reproducible output.

Ideogram V3 Layerize Text

image

Required. The image to remove text from. Works best on designs with clearly legible, typeset text on a distinct background.

prompt

Optional, up to 2048 characters. Guidance to influence how the model identifies and removes text.

seed

0 to 2,147,483,647, default random. Fixed seed for reproducible layerization output.

The schema also exposes font name and font file URL parameters for H1, H2, body, and small text levels. These are passed to the underlying model but do not affect Scenario's output, which is always the background PNG with text removed regardless of font settings. Leave them empty for standard use.

Use Cases

• Marketing and brand graphics (V4 base). Posters, lookbook covers, social key art, packaging, signage, UI mockups. Wherever the deliverable carries readable typography baked into the image.

• Multilingual typography (V4 base). Japanese calligraphy, Arabic, Cyrillic, Korean, Devanagari, chrome movie titles. The text engine handles non-Latin scripts at QUALITY without falling back to placeholder glyphs.

• Editorial restyles and campaign variants (Image to Image). Take one product or character shot and ship multiple stylistic versions for a campaign: editorial pencil, sumi-e, stained glass, risograph.

• Concept art and game character mood passes (Image to Image). Generate alternate medium versions of a finalized character for marketing or pitch decks.

• Marketing collateral with text overlays (Image to Image). Write the headline INTO the restyle rather than pasting on top. One-shot social posts, postcards, lookbook covers.

• Surface design and wallpaper (Tiling). Seamless wallpaper, gift wrap, packaging surfaces. Ships at native 2K.

• Fabric and apparel prints (Tiling). Otomi, damask, polka dots, chintz patterns tile cleanly for textile production.

• Game-ready ground and material textures (Tiling). Drop into Unity, Unreal, or Blender as a tiling material. The cyberpunk circuit and honeycomb brass examples above repeat without seam artifacts at any UV scale.

• Interior and architectural visualization (Tiling). Marble, brick, wood plank tiles slot directly into a 3D scene's UV-tiled materials.

• Logo and brand assets (Generate Transparent). Logos, wordmarks, and brand icons with a transparent background ready for any surface, no clipping required.

• Sticker and overlay packs (Generate Transparent). Sticker sheets, emoji-style icons, and overlay graphics at scale, up to 8 variants per job.

• Design template extraction (Layerize Text). Strip the text from an existing poster, label, or ad creative to produce a reusable background template.

• Localization workflows (Layerize Text). Remove text from a finished design, then re-composite with translated copy while the visual layout stays consistent.

• Game UI and HUD assets (Generate Transparent). Buttons, badges, and frames with native transparency for direct integration into a game engine sprite sheet.

Tips for Better Results

1. Quote literal text for in-image rendering. Use double quotes around the words you want rendered: "VAN LIFE" beats "the words van life". Works in all three variants.

2. Disable Magic Prompt for spelling-critical text. Set expansionModel: None when you need the model to obey your wording exactly, especially for brand names, dates, addresses, and prices.

3. Start at Balanced, finalize at Quality. Iterate quickly with renderingSpeed: BALANCED, then re-render the chosen seed at QUALITY for delivery. Quality matters most for intricate typography and complex patterns.

4. Image to Image: start at strength: 0.7 and adjust. 0.7 is the sweet spot for a clear stylistic shift that still preserves the source. Push higher for medium swaps; pull lower if the result drifts from the subject.

5. Name the medium concretely, not the "style". "Risograph print, two-color overprint of fluorescent pink and electric blue" outperforms "stylized print look". The more material the prompt names (paper, ink, brush, print process), the more committed the restyle.

6. Tiling: match tilingMode to the pattern's logic. Marble veins should not loop vertically; wood planks should not loop horizontally. Pick the mode that matches real-world repetition.

7. Tiling: default to square_hd unless the pattern has a direction. Square is the most flexible. Reach for landscape or portrait when there is an obvious primary axis.

8. Lock a seed when iterating colors or text. Reuse the seed and change only the color words or the quoted text to keep the composition or pattern structure constant while swapping detail.

9. Tiling: test the tile by repeating it. Before shipping, paint the tile as a CSS background or load it into a 3D viewer at the intended UV scale. Seams hidden in a single tile become obvious across a four-by-four repeat.

10. Keep acceleration at none for dense typography. Higher acceleration shaves time but can soften letterforms.

11. Generate Transparent: describe subject, style, and color together. "Coffee shop logo with 'BREW' text, vintage badge style, warm brown and cream tones, circular design, transparent background" produces a well-defined output; short prompts produce less controlled results.

12. Layerize Text: feed designs with clear, typeset text. Typeset text on a contrasting background produces the cleanest removal. Hand-lettering or text deeply embedded in illustration detail may be removed only partially.

13. Layerize Text works on any image, not just Scenario generations. Uploaded photos, screenshots, or rendered designs all work as input.

Known Limitations

• Long text passages degrade. All three variants excel at short labels, headlines, and brand words. Multi-sentence paragraphs still produce errors. Keep on-image type concise.

• Image to Image preserves composition, not fine geometry. At high strength the model can drift on small details (number of fingers, exact letter shapes in long passages, jewelry detail). Drop strength or use QUALITY when those details matter.

• Tiling guarantees edges, not perfect symmetry. The edges line up by design, but the pattern within the tile is not always perfectly symmetrical. For strict geometric symmetry, prompt explicitly for it.

• Plan access restrictions apply. Some workspaces may not have access to the full Ideogram family.

• Layerize Text returns only the background image. The underlying API also returns structured text position/font/color data, but Scenario does not expose it; only the cleaned background PNG comes back as an asset.