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 COMPARISON

Compare models across multi-turn edits.

Here's a side-by-side comparison of the same edits on Ideogram 4.5 vs. GPT Image 2.5 Sunburst, Nano Banana Pro, and Nano Banana 2. GPT Image and Nano Banana outputs become unusable within a few edits, while Ideogram 4.5 stays clean edit after edit.

Precision for every kind of edit

Explore the kinds of edits Ideogram 4.5 handles with exceptional precision, from refining color and restoring old photos to making targeted changes while preserving the rest of your image.

Edit at any resolution

Edit high-resolution images without downsizing them. Ideogram 4.5 preserves the edges of an edited crop, so you can stitch it seamlessly back into the original. Create new colorways or fix a detail while keeping the rest of your image sharp and untouched—ready for close-ups and large-format print.

Source image: 4,016 x 6,016 pixels (24.2 MP)

Make your next image your best yet.

Start with an image you love. Tell Ideogram 4.5 what to change. See how far your next idea can take it.

Original source

Remove backgrounds from images with AI

Upload an image and get a clean transparent PNG in one click. Ideogram preserves fine edges on product photos, logos, typography, hair, glass, and intricate cutouts.

How to remove a background with AI

Step 1

Upload your image

Start with a product photo, portrait, logo, illustration, or graphic that needs a transparent background.

Step 2

Remove the background

Ideogram separates the subject from the scene and preserves fine details like soft edges, type strokes, hair, and glass.

Step 3

Download the PNG

Export a transparent PNG ready for marketplaces, catalogs, ads, design mockups, and creative tools.

Transparent PNGs with clean edges

Most background removers struggle at the border between foreground and background, leaving halos, fringing, or leftover color. Ideogram focuses on clean alpha channels so logos, typography, product silhouettes, hair, fur, and glass can stay sharp without manual cleanup.

Built for product photography and catalogs

Remove distracting studio, tabletop, and lifestyle backgrounds from product images so every asset is ready for listings, launch pages, ads, and marketplace refreshes. Teams can process one image in the app or send high-volume jobs through the API.

Keep typography, logos, and detailed graphics intact

Background removal is hardest when the subject is made of fine strokes and negative space. Ideogram preserves graphic detail in lettering, badges, illustrations, and print-on-demand artwork so the cutout can be reused on new colors and layouts.

Remove backgrounds through the API

POST an image to /v1/remove-background and receive a signed URL for the transparent PNG. The endpoint is built for tools, marketplaces, and production workflows that need consistent background removal across thousands of images.

API

Remove backgrounds via the API

POST an image to /v1/remove-background and receive a signed URL to the transparent PNG. The endpoint is built for teams running volume across product launches, marketplace listings, and catalog refreshes that need consistency over thousands of images. The cutout quality is the same on the first request and the ten-thousandth. Pricing is per image, with volume tiers for enterprise teams.

Get API Key

cURL

curl -X POST https://api.ideogram.ai/v1/remove-background \ -H 'Api-Key: YOUR_API_KEY' \ -F '[email protected]'

Frequently asked questions

Why is Ideogram the best background remover?

Ideogram produces cleaner edges than remove.bg, Photoroom, BiRefNet, and RemBG on the hardest subjects: hair, fur, glass, reflections, intricate cutouts, and fine product silhouettes. The output is a transparent PNG with no fringing, no halo, no manual cleanup.

Original source

Why your old prompts behave differently, what JSON actually is, and how to get precise results

text, layout, colour, and style.

If you've moved from a previous version of Ideogram to 4.0 and found that your go-to prompts are giving you inconsistent results, layouts that shift between generations, styles that feel generic, text that garbles, you're not doing anything wrong. Our model has changed in a fundamental way to bring you stronger control over the output. This guide explains what changed, why it matters, and how you can take advantage of it.

4.0 was trained on structured JSON captions: a format that describes an image as organized, labeled data rather than a single run-on sentence. It doesn't mean you have to write code. It means the model has a preferred language now, and the closer you get to it, the more reliably it does what you ask.

Why your prompts might feel different in 4.0

Ideogram 4.0 is more literal than you might expect. Words like "moody," "vibrant," or "editorial feel" carry less weight now. The fix isn't to write more, but to write more specifically. "Warm directional light from the upper left" instead of "nice lighting." "Hasselblad medium format, slightly desaturated" instead of "film look." Named references, specific light sources, exact colours. That's the kind of vocabulary that 4.0 responds to, and the JSON format is just a way of making sure every part of your image has that kind of specificity attached to it.

WHY AM I HITTING THE SAFETY FILTER?

Plain-text prompts trigger significantly more false positives with the safety filter than structured prompts. We are aware of this and may make a future checkpoint update to improve it. If you're getting blocked on something innocuous, adding more detail to the JSON is the fastest fix.

JSON explained — no coding experience needed

Before 4.0, your prompt was one long sentence and the model filled in the gaps however it liked. Sometimes that was great. Other times, the logo ended up in the wrong place and the background was nothing like what you described, and you had no way to tell the model exactly which part was wrong.

JSON gives every part of the image its own labeled slot. What's important for Ideogram 4.0 is that your prompts need to be organised, labeled information. Each field in your JSON prompt tells the model exactly one thing: what the style is, what the background looks like, what text to render and where.

Here's the same image described in both formats:

PLAIN TEXT

A lush tropical greenhouse interior, Rousseau-style oil painting, jungle plants, parrot, bright flowers

Works, but the model infers a lot — results vary

JSON STRUCTURED CAPTION

{
  "high_level_description": "Lush tropical greenhouse interior, Rousseau-style oil painting",
  "style_description": {
    "aesthetics": "dreamlike, abundant, deeply saturated",
    "lighting": "dappled greenhouse light through glass ceiling",
    "medium": "oil painting",
    "art_style": "Henri Rousseau, naive jungle painting, rich impasto"
  },
  "compositional_deconstruction": {
    "background": "dense emerald and cadmium green foliage filling the frame",
    "elements": [
      {"type": "obj", "desc": "broad tropical leaves in deep green, layered and overlapping"},
      {"type": "obj", "desc": "exotic orchids in crimson and gold, scattered throughout"},
      {"type": "obj", "desc": "a parrot perched mid-frame, vivid red and blue plumage"}
    ]
  }
}

Each part of the image is explicitly labeled — more consistent

PLAIN TEXT RESULT

Same oil painting, but the model chose the palette, the flowers, and the layout. Run it again and those choices drift.

JSON RESULT

The oil medium, the deeply saturated palette, and each named element — orchids, parrot, layered foliage — are spelled out, so they all arrive.

Think of JSON as organizing your prompt into four clear jobs:

• What's the overall scene? (high_level_description)
• What's the visual style? (style_description)
• What's the background? (background)
• What specific things are in it? (elements)

You don't have to memorize the syntax. Magic Prompt (covered below) can translate natural language into JSON prompts that you can edit later. But once you understand the four jobs, you'll know exactly where to change something when a generation isn't quite right.

Breaking down the JSON format

If you've ever gotten a generation that was almost right — the composition was good but the style felt generic, or the lighting was flat when you wanted dramatic — the culprit is usually a missing or thin style_description. The model has no default style. If you don't tell it what kind of image this should be, it guesses. Here's what each field is actually responsible for:

FIELD DESCRIPTION TYPE high_level_description One or two sentence scene summary. Reads like a natural prompt — start with the subject, name recognizable properties. Keep it under 50 words. string style_description Visual style: aesthetics, lighting, medium (photograph, illustration, graphic_design), and optionally a colour palette. Key order is strict — see below. object compositional_deconstruction Background description plus an elements array. This is where layout, text placement, and bounding boxes live. object

KEY ORDERING MATTERS

The style_description fields must appear in a specific order — the model was trained that way.

Photography:

aesthetics → lighting → photo → medium → color_palette.

Illustration/design:

aesthetics → lighting → medium → art_style → color_palette. The open-source CaptionVerifier catches ordering mistakes before you generate.

Shifting from natural language to JSON

If switching to JSON sounds like learning a new language when you just want to make images, you don't have to. Magic Prompt is a built-in translator — you type a normal sentence, and before it ever reaches the model, an LLM rewrites it into a full structured JSON caption. On the live version of 4.0, Magic Prompt will turn ON or OFF automatically: the model decides whether to expand your input based on what you sent. Plain text gets expanded. A valid, well-formed JSON prompt passes through as-is.

Here's what that translation actually looks like:

WHAT YOU TYPE

a product shot

of

a perfume bottle, elegant

5 words in

WHAT GETS SENT TO THE MODEL

{
  "high_level_description": "A perfume bottle product shot, luxury cosmetics photography",
  "style_description": {
    "aesthetics": "editorial, luxury, minimal",
    "lighting": "single soft key light upper left, subtle fill right",
    "medium": "photograph",
    "photo": "studio"
  },
  "compositional_deconstruction": {
    "background": "clean white surface, slight shadow",
    "elements": [
      {"type": "obj", "bbox": [100, 300, 900, 700], "desc": "cylindrical frosted glass perfume bottle, gold metal cap, embossed logo"}
    ]
  }
}

Full structured caption — this is what the model actually sees

Magic Prompt is opinionated — it makes decisions on your behalf that you might not want for controlled output:

• Populates sparse scenes with secondary elements you didn't ask for
• Upgrades vague references to specific ones ("Nike Air Jordan 1" instead of "retro sneaker")
• Removes hedging language — "could include," "various," or "similar" get stripped out
• Drops render metadata like bokeh and film grain unless you explicitly name them
• Commits to one value when you gave it two — "cream or ivory" becomes "cream"

For quick exploration, this is exactly what you want. For a brand deliverable where hex colours and text placement are non-negotiable, write JSON directly. Magic Prompt won't expand a prompt that's already structured.

Same scene, two approaches. Natural language gives general placement cues; JSON locks every element to an exact zone.

Why is the text in my image coming out wrong?

You tried to generate a logo mockup. The brand name came out as a smear of letters that vaguely resembles the right word. Or you asked for a poster headline and the model gave you something grammatically adjacent but spelled wrong. Or the text was fine but landed in completely the wrong font.

AI image models have struggled with text rendering for years. 4.0 has made it largely solvable through literal text rendering: during training, the model learned to treat text inside quotation marks as exact rendering instructions. To take advantage of this, quote exactly what you want rendered. With this method, Ideogram 4.0 achieves 97% accuracy on large English text benchmarks.

DESCRIBING THE TEXT

A vintage naturalist collector card

for

a mantis shrimp

with

its species name and a short description

Text may be garbled, hallucinated, or omitted entirely

QUOTING THE TEXT EXACTLY

Vintage naturalist collector card,

"MANTIS SHRIMP"

in

bold display serif at top,

"Odontodactylus scyllarus"

in

green italic below the illustration,

"Possesses the most complex eyes in the animal kingdom."

in

smaller roman,

"8 of 50"

bottom right,

"OCEAN MARVELS"

in

banner at bottom

Every text element lands exactly as written

THE DOUBLE-QUOTE RULE

Put specific text you want rendered in double quotes inside your prompt — "MANTIS SHRIMP" not the species name. 4.0 treats quoted strings as literal rendering instructions. Keep each quoted string to 5–7 words maximum, as accuracy drops off on longer strings. In JSON, use "type": "text" with a "text" field for the literal string and "desc" for how it should look visually.

Rough accuracy guide by size and language:

• Large English text — ~97% accurate
• Small English text — ~90% accurate
• Large Arabic text — works well for single words and short phrases
• Small Arabic text or full paragraphs — expect garbling and phantom letters

For numbers inside charts or labels, default to using words. Prompting with "forty-two percent" renders more reliably than "42%" at small sizes.

How to control exactly where things appear

If you've ever tried to generate a movie poster or a product ad, you know the pain. The headline drifts into a corner. The logo sits on top of the main subject instead of below it. The credit block is unreadable soup. You regenerate, and everything moves to a completely different position.

To solve this, Ideogram 4.0 was trained with bounding box coordinates. The model learned to couple plain-language descriptions with exact spatial locations on a 0–1000 grid, so a "credit block" lands where you put it and stays there across seed variations.

The format is [y_min, x_min, y_max, x_max], with 0 at the top-left and 1000 at the bottom-right. A box covering the lower-right quadrant would be [500, 500, 1000, 1000].

Elements land inside their declared box consistently, even across seed variations. Use this for anything with a layout requirement: posters, packaging, social media templates, infographics.

Three things to know:

• Ground, floor, and pavement are always background — never put a surface as a bounding-box element
• Keep total elements to 5 or fewer per image for best quality
• Boxes are layout guides, not pixel-perfect crops — leave some margin

How to use exact brand colours

You described your brand colour as "deep forest green with a warm undertone." The model gave you something olive-ish. You regenerated. Now it's teal. You added "like a Rolex dial" and now it's darker but still not right. Twenty minutes gone trying to get a specific green out of a sentence.

Ideogram 4.0 supports hex colour conditioning to break out of that cycle. Pass the exact hex code and it anchors the image's entire palette around that value. A hex code like #2D5A27 will stick between generations.

FIELD DESCRIPTION LIMIT color_palette (style) Image-wide dominant palette. Include background and shadow values — not just your primary brand colour. The model anchors the full tone range against this list. Up to 16 hex codes color_palette (element) Per-element colour steering. Use this to lock a specific text or object element to a brand colour while letting the rest of the image breathe. Up to 5 hex codes

FORMAT IS STRICT

Hex codes must be uppercase #RRGGBB, with no shorthand (#f63) and no lowercase (#ff6b35). The model was trained on a specific serialisation. The CaptionVerifier will flag anything non-conforming. Always include a background colour and a shadow/highlight value in your palette — a complete set like ["#0A0A0A", "#C9A84C", "#FFFFFF"] outperforms ["#C9A84C"] on its own.

Identical prompt, different color_palette arrays. The model anchors the entire tone range — highlights, shadows, and reflections — against the hex values you provide.

How to keep the same look across multiple images

You generated a perfect image — the grain was right, the colour grade was exactly on-brand, the lighting felt like your reference. Then you tried a second image in the same series. Same style description. It looked nothing like the first one.

That's the core problem with style-in-language. The same words produce different interpretations across generations. Style Codes are Ideogram 4.0's answer to this. They're a fingerprint extracted from a specific generation that you can apply to any future prompt. Generate one hero image you love, copy its Style Code from the panel, and pass it as style_code. Every generation after that shares the same visual DNA.

Same prompt — Three style codes

"A fox portrait, clean background"

Dutch oil — SC_7f3d2

Woodblock — SC_2a8c1

Bauhaus — SC_9e4f3

Style codes extracted from higher-quality presets (V4_QUALITY_48) tend to produce better results when reapplied. Preset quality levels also affect consistency:

• V4_TURBO_12 (~21s at 1024²) — exploration and drafts
• V4_DEFAULT_20 (~32s at 1024²) — daily production work
• V4_QUALITY_48 (~72s at 1024²) — finals, typography-critical output, style code extraction

The settings that actually change your output

SETTING DESCRIPTION VALUES style_code Locks in an exact visual aesthetic by code. Extract from any generation you love in the app generation panel. SC_xxxxx style_type Broad aesthetic category when no style code is set. AUTO lets the model infer from the prompt. REALISTIC — DESIGN — AUTO magic_prompt Always set to AUTO on live Ideogram 4.0. Plain text gets expanded into JSON; a well-formed JSON prompt passes through as-is. Configurable when self-hosting. AUTO negative_prompt Elements to suppress. Works best when specific. Pair with positive framing — "clean white background, no other objects" is more reliable than a long exclusion list. string seed Reproducibility anchor. Same seed + same prompt = same composition. Copy from a generation you want to vary — change the prompt, keep the seed. integer sampler_preset Quality/speed tradeoff. Use V4_QUALITY_48 for typography-critical output — the extra steps meaningfully improve small-text accuracy. V4_TURBO_12 — V4_DEFAULT_20 — V4_QUALITY_48

Plain text vs. JSON: which one should I use?

Magic Prompt is always AUTO on live 4.0 for now, so the real choice is whether to write plain text or JSON. Plain text gets expanded automatically. JSON goes through as-is, giving you full control. Here's when to use one over the other:

WHEN TO USE THIS APPROACH First idea, exploring Plain text — Magic Prompt expands it Brand asset with exact colours JSON with color_palette Poster or packaging with specific text JSON + text elements + bbox Full campaign consistency JSON + style_code from hero image Getting blocked by safety filter JSON prompt Self-hosted / local inference JSON (Magic Prompt backend is separate)

THE TOOLING IS OPEN SOURCE

The JSON schema, CaptionVerifier, and all three Magic Prompt backends are at github.com/ideogram-oss/ideogram4. The backends are swappable: hosted Ideogram API, Claude Opus/Sonnet via OpenRouter, and Gemini Flash-Lite.

What to try next

The techniques in this guide compound. Start with a plain-language prompt and get something close. Add a style_code from that result to lock the look. Switch to JSON when you need exact colours or placed text. Most workflows land on a combination of all three.

• Generate at scale with the API. Every technique here works over the API — batch runs, seed pinning, style code reuse across a full campaign.
• Explore the open-source tooling. CaptionVerifier and all three Magic Prompt backends are available at github.com/ideogram-oss/ideogram4 if you want to run them locally or swap in your own LLM backend.

Start generating

Open Ideogram 4.0 and try your first structured prompt.

Original source

IDEOGRAM 4.0

Open image model at the forefront of design.

Built for design: open weights, multilingual text, precise layout control, editable elements, and realistic 2K images.

THE OPEN APPROACH

The future of generative AI is open source.

Chromium outran every closed browser engine, PyTorch is the dominant ML framework, and most of the internet runs on open source software. Over the past year, proprietary image models have set a new standard on text rendering, prompt adherence, and photorealism, while open weight models have fallen behind.

Today, we are releasing Ideogram 4.0 as a state-of-the-art open weight image model for developers and enterprises to build with us. The weights are yours to download, fine-tune, and run on your own hardware. Commercial deployments come with a license that matches your scale. We believe openness drives innovation, and we invite the research community to innovate with us on the forefront of visual intelligence.

MODEL TRAINING

Teaching the model to read structure before recreating it.

A picture is worth a thousand words. Ideogram 4.0 was trained with a describe-to-structure-to-recreate loop: first reading scenes, backgrounds, text, and objects as structured data, then learning to rebuild images from that representation.

COMPOSITION CONTROL

Design with every element in the right place.

We trained Ideogram 4.0 with bounding boxes coupled to plain-language descriptions, teaching the model where each object, text region, and layout element belongs before it paints the final image. That structure lets the model learn tighter composition in dramatically less training time, while giving creators fine-grained control over dense, compelling layouts.

Hover a layer to trace the same bounding box from the prompt plan to the generated poster.

FOR DESIGNERS

Generations come out as editable files, not flat frames.

Production design rarely ends at a single pixel layer. Headlines change before launch and cutouts drop onto new backdrops. Ideogram already ships these workflows as separate tools, and the next 4.0 release brings them native to the model itself.

SHIPPING TODAY

Background removal already returns transparent output. The Background Remover produces a clean alpha cutout from any generation, so the result drops onto a new backdrop without manual masking or Photoshop cleanup.

Layerize already extracts editable text. Headlines, body copy, and graphic elements come back as separate editable layers, so the typography stays revisable after the model is done.

COMING TO 4.0

The next release of 4.0 returns alpha channels and editable text layers directly from inference. No second pass, no masking step. The model's output is the editable file your team can hand to production.

WHY THIS MATTERS

Choose the open model designed for the work your team actually ships.

Open-weight image models are no longer scarce. The decision an enterprise faces is which one to standardize on, and which capabilities will hold up against the next two years of brand work and platform shifts.

• Brand work is the lane it was built for.
  Ideogram has led on text rendering since launch, and 4.0 adds bounding-box layout control. Headlines stay readable, packaging copy says the right words, and logos land where the brief asked for them.
• It converges on your house style, not on generic taste.
  Open weights and a commercial license let your team fine-tune on style guides, product photography, and historical campaigns until the model defaults to your look instead of fighting it.
• It runs where your CIO needs it to run.
  Deploy on your hardware, behind your firewall, in the region your residency rules name. Inference cost scales with the compute you provision, not with how many images marketing ships next quarter.

ENTERPRISE

Frontier visual intelligence built for your brand.

Ideogram 4.0 is built for enterprises across every sector. It ships with open weights, a commercial license, and the customization options enterprises need. Run it on your hardware, train it on your data, and keep your output behind your firewall.

API

Use Ideogram 4.0 commercially through the hosted API.

The fastest way to integrate Ideogram 4.0 into your product is through the API. You get hosted access to the model for commercial use, with three quality tiers so you can choose the right tradeoff between speed, cost, and output fidelity.

API AND PRICING

• Turbo: $0.03 / image
• Default: $0.06 / image
• Quality: $0.10 / image

Per-image pricing, no subscription required. Start with the hosted API and scale up based on the volume your platform needs.

Start building with Ideogram today

Think it. Make it. Own it.

Ideogram 4.0 gives visual ideas somewhere to go: into products, agents, campaigns, and finished design systems.

Original source

IDEOGRAM MCP

Add Ideogram to any AI agent

Available on all plans. OAuth via your Ideogram account.

INSTALL

Connect any MCP client.

The server speaks streamable HTTP at https://mcp.ideogram.ai/mcp. Authentication is OAuth via your Ideogram account. The first time a client connects, a browser window opens for sign-in.

Add to claude_desktop_config.json. Restart the app.

IN PRACTICE

Try one of these to start.

Drop these into your assistant after connecting. The agent picks the right tool from the catalog.

WORKFLOWS

See what teams build with it.

Each of these is a single conversation. The agent chains the right tools without anyone writing glue code.

Asset packs from a brief

Hand the agent a one-paragraph campaign brief. It returns a hero, four social cards, and two print posters in a single conversation, saved into a named collection.

• generate_images_bulk
• create_collection
• upscale_image

Catalog cleanup

Feed product photos into the chat. The agent removes backgrounds, reframes for marketplace ratios, and uploads the cleaned assets back to a shared collection.

• edit_image
• reframe_image
• upload_image

Train and iterate

Drop a folder of brand references. The agent creates a dataset, uploads the assets, kicks off training, and pings you when the model is ready to use.

• create_dataset
• upload_dataset_assets
• train_model

QUESTIONS

Frequently asked

Start building with Ideogram today

Think it. Make it. Own it.

Ideogram 4.0 gives visual ideas somewhere to go: into products, agents, campaigns, and finished design systems.

Original source

AI background remover built for graphic design.

Production-ready background remover for designers, setting a new standard for background removal. Obtain transparent PNGs with clean edges on letters, logos, and intricate cutouts.

NEW APPROACH

A generative background remover, not a segmentation model.

Most background removers are segmentation models that classify pixels as foreground or background without a deep semantic understanding of the image. The main challenge in background removal lies at the borders, where pixels naturally blend the foreground and background, and the correct output is a partial transparency paired with the foreground color underneath. This is the matting problem, and segmentation models are not built to solve it. The translucent edge of a glass carries the color of the wall behind it, hair scatters light into a soft halo, and an anti-aliased typographic stroke is by design indistinguishable from its surrounding. The failures show up in the alpha channel as halos, fringing, and residual color from the discarded background.

Ideogram approaches the problem differently. The background remover is a generative model trained from the ground up, which carries a deep semantic understanding of the scene it depicts. Because the model already knows what a bottle, a strand of hair, or a typographic stroke looks like, it can reconstruct each border pixel as a partial transparency paired with the foreground color it should carry. Transparency is preserved through glass and other refractive surfaces, fine detail survives unaltered, and the output is a transparent PNG with no halos, no residual color, and no need for cleanup in Photoshop.

COMPARISON WITH THE STATE-OF-THE-ART

No manual cleanup needed with Ideogram BG Remover.

Same input run through Ideogram and the leading commercial and open-source background removers. Look at the typography stroke, the negative space, and the soft edges. The difference is in the alpha channel.

API

One API call for a production-ready transparent PNG.

POST an image to /v1/remove-background and receive a signed URL to the transparent PNG. The endpoint is built for teams running volume across product launches, marketplace listings, and catalog refreshes that need consistency over thousands of images. The cutout quality is the same on the first request and the ten-thousandth. Pricing is per image, with volume tiers for enterprise teams.

Original source

Your custom image model, trained on your brand.

Train your model on 15-100 on-brand images with a specific style. Every new generation from your custom model will follow your art direction, typography, and visual identity.

Available on Team, Pro, and Enterprise plans and via the API.

Trained on 27 illustrations from The Tale of Peter Rabbit by Beatrix Potter, this model learned soft English watercolor washes, delicate pen linework, anthropomorphic character proportions, and the pastoral garden compositions that define the series. Every output carries the same muted earth tones and countryside warmth.

THE PROBLEM

A model trained on the internet doesn't know your brand.

Off-the-shelf models are trained on the entire internet: billions of unrelated images blended together. The outputs look almost right but carry no defined aesthetic. When your brand runs through a model built for everyone, it starts to look like it belongs to no one.

A custom model is fine-tuned exclusively on your assets, to learn your style, your visual identity, and your aesthetic choices. Feed the best outputs back into training, and the model evolves as your brand does. Train a different model for every expression of your brand.

WHY IDEOGRAM

Your LoRA is only as good as its base model.

Open-source LoRAs are similar in nature to Ideogram Custom Models, but they inherit the limitations of open-source base models. Inconsistent text, generic composition, limited style range. The foundation matters.

Ideogram was started by the researcher behind Google's first text-to-image system. We cracked legible text rendering first. Then character consistency. Now we're solving custom model training on a foundation built for it, already powering enterprise workflows in production.

Same prompt across 9 custom models:

Woman seated in bistro chair, coffee cup in hand, Paris street with Eiffel Tower in background. Text "Paris is always a good idea" is displayed across the top.

ENTERPRISE

Want more control over your custom model?

Talk to our team about advanced customization.

• 01 White-glove setup and training.
  Access to our research, engineering, and annotation teams. From asset sourcing to annotation to deployment.

• 02 Large datasets, unlimited iteration.
  Train on thousands of images with custom captioning pipelines. Iterate until the model matches your standards.

• 03 Custom typography and exact logos.
  Your brand marks rendered as layerized text and SVG. Pixel-perfect on-brand identity, every time.

• 04 Your data, your model.
  Training data and models are not used to improve our foundation models. Your model and data are yours alone.

100K+ Branded concepts generated by Skechers
30+ Custom models trained for enterprises
3B+ Designs generated on Ideogram
30M+ Creators, designers, and marketers

TRUSTED BY

Netflix, Adobe, Skechers, WPP, Vistaprint, Wix, Gamma, Canva

HOW IT WORKS

Upload. Train. Generate.

No infrastructure to manage. No GPU provisioning. You bring the images. Ideogram builds the model.

• 01 Upload images
  Select the Models tab from the navigation bar. Click + to create a new model. Upload 15 to 100 images of your brand, subject, or style.

• 02 Train your model
  We auto-generate captions for your images. Review and refine them, then kick off training. Quality matters more than quantity.

• 03 Generate at scale
  Select your Custom Model and start prompting. Stack Style References and Color Palettes for full brand control.

API SUPPORT

Train and generate programmatically.

Create datasets, upload training images, kick off model training, and generate at scale. All from your own infrastructure, no UI required.

• 01 Create a dataset and upload 10–100 images
• 02 Start training with a single POST request
• 03 Generate images using your custom_model_uri

FOR TEAMS

Custom Models, built for your whole team.

Train a Custom Model on your visual identity once. Share it across your organization. Every team member generates from the same foundation, whatever the output needs to be.

CUSTOM MODELS ON TEAM PLAN

Train it. Share it. Everyone generates from it.

Train on your approved assets. Share across the organization. Generate at scale.

• Any team member can train and share
  No gatekeeper. Anyone builds a model, shares it across the org, and the whole team generates from it.

• Unlimited models, no cap
  Product photography, social content, print campaigns. Each team trains the model they need.

• Trained from shared collections
  Pull from collections of approved assets to train. The brand source stays central.

Train once, share with your whole team.

Frequently asked questions

How many images do I need to train a custom model?
What can I train a custom model on?
What if my custom model cannot generate certain scenes?
Can I combine a custom model with other Ideogram features?
Is my training data used to improve Ideogram's base models?
Can I share a custom model with my team?
Can I train a model on my brand's custom font or logo?
How does Enterprise differ from self-serve custom models?
How can I access Ideogram Custom Models?

Original source

MODEL OVERVIEW

Ideogram model 3.0

Photorealism. Legible text. Reusable style control. Ideogram 3.0 is the release that turned image generation into something creative teams could actually use for campaigns, brand systems, and polished design work.

Pushing the limits of image generation

Ideogram 3.0 pushes the boundaries of generative media through significant advancements in image-prompt alignment, photorealism, and text rendering quality. In human evaluations, Ideogram 3.0 consistently outperforms other text-to-image models, scoring highest in ELO rating over a set of diverse prompts that probe a wide variety of capabilities, subjects, styles, use cases, and composition difficulty.

Consistent Styles

Ideogram 3.0 introduces Style References. Upload up to 3 reference images to control generations to follow your preferred aesthetics. For fresh style inspiration, the Random style feature explores a unique mix from our library of 4.3 billion presets. Once a desirable style is identified, it can be reused via its Style Code.

Creative Designs

Game-changing text and layout generation for graphic design, advertising, and marketing. Stylized, accurate text with remarkable precision, including complex and lengthy compositions that other models struggle with.

Brand Graphics

Professional logos, promotional posters, landing pages, product photography. Ideogram 3.0 generates complete brand assets with accurate text and consistent visual identity. Use Batch Generation to create multiple variants in a single session.

Stunning Realism

Natural skin tones, accurate reflections, proper lighting physics. Spatial compositions, environmental detail, and photographic quality that holds up at full resolution. From surreal concepts to documentary-style portraits.

Original source

Our first open-weight foundation model. A 9.3B single-stream Diffusion Transformer, trained from scratch, with a vision-language text encoder and structured JSON prompts.

Weights.

Code.

Overview

Ideogram 4.0 is a 9.3B parameter open-weight text-to-image model. Recent open-weight releases have converged on a single self-attention sequence over text and image tokens, and Ideogram 4.0 follows the same pattern: text and image tokens share the same projections at every layer of a 34-layer DiT. Two design choices distinguish it from peer releases. First, the text encoder is Qwen3-VL-8B-Instruct, a vision-language model, and the DiT consumes hidden states from 13 of its intermediate layers concatenated along the feature dimension, instead of a single hidden state or no external encoder at all. Second, the model is trained exclusively on structured JSON captions with per-element styling and optional bounding boxes and color palettes, and the reference inference pipeline parses every prompt as JSON and validates it against the schema before generation.

Architecture at a glance

The pipeline has four components: a frozen vision-language text encoder, the trainable single-stream Diffusion Transformer, a flow-matching Euler sampler with asymmetric classifier-free guidance, and a frozen KL autoencoder that decodes latents to pixels.

Figure 1. End-to-end inference pipeline. The 9.3B parameter DiT is the only trained component; the text encoder and VAE are reused frozen.

Model spec

FIELD VALUE
parameters 9.3B
emb_dim 4608
num_layers 34
num_heads 18
intermediate 12 288 (SwiGLU)
adaln_dim 512
rope_theta 5 000 000
mrope_section (24, 20, 20)
latent channels 32 × 2² = 128
max text tokens 2048
resolution range 256–2048 px per side
aspect ratios flexible
quantization nf4 and fp8
sampler Euler flow-matching, asymmetric CFG

Model

The text encoder is Qwen3-VL-8B-Instruct in text-only mode. The DiT consumes hidden states from 13 of its intermediate layers, concatenated along the feature dimension, instead of a single hidden slice.

The DiT is 34 transformer blocks operating on one sequence of text and image latent tokens. Each block uses self-attention with QK-RMSNorm, 3D Multimodal RoPE that places text and image tokens in one positional frame, and a SwiGLU MLP. AdaLN from the flow-matching timestep modulates the attention and MLP residual paths in every block.

The training objective is flow matching: the DiT predicts a velocity field v(z_t, t) that defines an ODE from pure noise to a clean latent. At inference, an Euler sampler integrates the ODE over a logit-normal noise schedule whose mean auto-adjusts with resolution.

Asymmetric classifier-free guidance

Classifier-free guidance combines a conditional pass (text + image latents) with an unconditional pass. Ideogram 4.0 makes the unconditional pass asymmetric: it drops the text tokens entirely instead of replacing them with padding, so the unconditional pass runs only over image tokens. The two branches can also be tuned independently, which lets us schedule prompt adherence and image quality separately across the sampling trajectory.

The V4_QUALITY_48 preset, for example, runs 45 steps at gw=7 followed by 3 polish steps at gw=3 near t=0. The shorter V4_DEFAULT_20 and V4_TURBO_12 presets follow the same shape with two and one polish steps respectively. The polish tail tightens fine detail without over-saturating the global composition.

Structured JSON prompting

The model is trained exclusively on structured JSON captions. Each training caption exhaustively describes every element in the image, with a style block and optional bounding boxes and color palettes. Training and inference share one prompt format, and the more relationships each caption pins down, the more grounded supervision the model extracts per training pair.

The reference pipeline validates every prompt against the JSON schema before generation and rejects inputs that do not parse, so the input format at inference time is the same one the model saw during training.

PROMPT EXAMPLE

{
  "high_level_description": "A cinematic 35mm film photograph of a lone wooden sailboat on a glassy lake at sunset, the boat on a right-third vertical with the horizon at the lower third, in a cool muted blue palette.",
  "style_description": {
    "aesthetics": "Cinematic, minimal, serene, quiet stillness.",
    "lighting": "Cool overcast dusk light with a small warm sun low at the horizon; muted and low-contrast.",
    "photo": "35mm motion-picture film still, 16:9 framing, subtle grain, slightly desaturated.",
    "medium": "Photograph.",
    "color_palette": ["#1B3A5C", "#5B8FB9"]
  },
  "compositional_deconstruction": {
    "background": "Windless evening on a wide lake; horizon at the lower third. Dusty blue-violet sky with a small amber sun at the horizon and a thin gold streak across the glassy teal water. Subtle 35mm grain.",
    "elements": [
      {
        "type": "obj",
        "bbox": [380, 590, 660, 720],
        "desc": "Lone wooden sailboat on the right-third vertical in the midground, dark varnished hull and a single tall mast with a slack white sail hanging limp in the still air. No visible crew."
      },
      {
        "type": "obj",
        "desc": "Mirror-perfect reflection of the sailboat and its mast doubling straight down into the glassy teal water directly beneath the hull."
      },
      {
        "type": "obj",
        "desc": "Out-of-focus reed tips slicing into the lower-left foreground, soft dark blades against the teal water, blurred from proximity to the camera."
      }
    ]
  }
}

Three things this surface gives you that a plain prompt does not:

• Color palette conditioning. Up to 16 hex colors per image (and 5 per element) steer the dominant colors directly, not through descriptive language.
• Bounding-box layout. Any element can be placed by bounding box, given as [y_min, x_min, y_max, x_max] in 0–1000 normalized coordinates (image origin at the top-left). The model honors the boxes through the shared MRoPE positional space.
• Typed text elements. A text element carries the literal string to render and a separate visual description for its styling. This is the mechanism behind multi-line, multi-font in-image text.

Benchmarks

We evaluate Ideogram 4.0 across five capabilities against the strongest closed-source models (GPT Image 2, Nano Banana 2) and every leading open-weight release. The two figures below give the headline picture; the model card has the full breakdown.

• LAYOUT CONTROL 0.69 (7Bench mIoU: how tightly generated objects sit inside their requested bounding boxes.)
• TEXT RENDERING 0.97 (X-Omni English OCR accuracy on in-image text.)
• SPATIAL REASONING 0.76 (SpatialGenEval accuracy, combining the spatial-relationship and basic-object question splits.)
• PROMPT ALIGNMENT 0.89 (Prism-bench alignment track. Following the prompt across long, compositional inputs.)

• We judge SpatialGenEval with gemini-2.5-flash instead of the leaderboard's Qwen2.5-VL-72B. Applied uniformly across all models, so cross-model comparisons remain valid.

Figure 2. Five-axis capability radar — Ideogram 4.0 vs leading closed-source models on layout control, spatial reasoning, object fidelity, prompt alignment, and text rendering. Layout control = 7Bench mIoU · Spatial reasoning = SpatialGenEval spatial_acc (8 spatial Qs) · Object fidelity = SpatialGenEval basic_acc · Prompt alignment = Prism alignment track · Text rendering = X-Omni OCR. *With Magic Prompt ON.

Figure 3. Parameter efficiency on text rendering (X-Omni EN OCR). Ideogram 4.0 (9.3B) sits alone in the small-but-strong corner, ahead of every other open-weight release model. Hunyuan Image v3 reports 80B total parameters with ~13B active per token.

Figure 4. Designer-preference ELO from an internal arena where graphic designers pick the better of two generations without being told which model produced either image. Bar length is the ELO score; the cap-line on each bar marks the 95% confidence interval. Ideogram 4.0 sits in second place overall and first among open-weight models.

Examples

A handful of qualitative samples grouped by capability. Every image below was generated with seed = 0 from the fp8 quantized checkpoint; click an image to view it larger (click again inside the viewer to toggle 1:1 zoom), and View prompt for the JSON.

Resolution flexibility

A single set of weights covers everything from square thumbnails to ultrawide banners. The noise schedule auto-adjusts per resolution, so you do not need a dedicated model for phone wallpapers, social headers, or 2K finals.

Safety

Open weights ship with three layers of mitigation. Pre-training data is filtered through NSFW classifiers across multiple categories before the model ever sees it, so the base distribution does not contain the content we explicitly do not want it to reproduce. Post-training procedures further reduce the probability of unsafe generations, even for prompts that explicitly request them. At inference time, the reference pipeline screens every prompt and every output through Hive moderation, and we require equivalent or stronger filtering for any redistributed deployment.

Open weight

Weights, inference code, full prompting guide, and sampler presets are public. The repository ships both fp8 and nf4 checkpoints; the nf4 variant fits on a single 24 GB GPU.

Get the model

Weights on Hugging Face. Inference code, prompting guide, and license on GitHub.

Original source

Meet Ideogram 3.0 — stunning realism, creative designs, and consistent styles, all in one powerful model. Now available to all users on ideogram.ai and our iOS app!

Ideogram 3.0 pushes the boundaries of generative media through significant advancements in image-prompt alignment, photorealism, and text rendering quality. In human evaluations, Ideogram 3.0 consistently outperforms other text-to-image models, scoring highest in ELO rating over a set of diverse prompts that probe a wide variety of capabilities, subjects, styles, use cases, and composition difficulty.

Consistent Styles

Ideogram 3.0 introduces Style References. Creators can upload up to 3 reference images to control generations to follow their preferred aesthetics. This enables creators to quickly specify styles that are difficult to describe with text-based prompting, resulting in a more efficient and expressive creative workflow. For fresh style inspiration, the Random style feature offers the ability to explore a unique mix from our library of 4.3 billion presets. Once a desirable style is identified, it can be reused via its Style Code.

Try a style:

• Bowling Alley
• Cadillac Diner
• Treehouse

A neon-lit bowling alley from the 1960s, with a rack of colorful bowling balls and retro decor.

Creative Designs

Ideogram 3.0 offers game-changing text and layout generation capabilities for graphic design, advertising, marketing, and other professional use cases. Stylized, accurate text can be created with remarkable precision, including complex and lengthy compositions that other models struggle with. The created text is beautiful and artistic, going far beyond the capabilities of typesetting in traditional design platforms.

Brand Graphics

Ideogram 3.0 revolutionizes design capabilities and approachability for small businesses and entrepreneurs. It’s now possible to generate professional-quality logos, promotional posters, landing page concepts, product photography, and more — all in mere seconds and at a fraction of traditional costs. Teams can also customize graphics at scale and ideate quickly and effortlessly using our Batch Generation feature.

Stunning Realism

In addition to design capabilities, Ideogram 3.0 brings a new level of realism that blurs the line between generated and real imagery. Creators can effortlessly craft sophisticated spatial compositions with intricate backgrounds, precise and nuanced lighting and colors, and lifelike environmental detail.

Thank You to Our Community

We extend our sincere appreciation to our beta testers and Creative Partner Program members for their valuable contributions in testing Ideogram 3.0. Special recognition goes to @josiassevero, @sragy, @areejthg, @atheist_american, @ylucic, @yeclanixon13, and @gabemabazza for providing several of the exceptional images featured in this post.

We're Hiring!

We're on a mission to develop generative media models to help people become more creative. If our mission resonates with you, then we’d love to hear from you. We're hiring for roles in AI research, engineering, and growth. Join us to build the future of AI for creativity. Please apply at Ideogram Jobs.

Original source

Enhance your creative workflow with Batch Generation

Enhance your creative workflow with Batch Generation, now available for Ideogram Pro users on ideogram.ai ! With this launch, we’re also increasing Pro plan credits to 3,500 per month, giving you even more flexibility to explore new ideas. See pricing details.

Batch Generation amplifies your image creation and ideation processes, removing the need for manual interaction and constant supervision. By reducing repetitive tasks, it enables you to focus on more creative, high-impact work. Whether you’re designing social media visuals, conceptualizing logos, developing product mockups, or running marketing campaigns, you’ll produce higher-quality designs more quickly and efficiently.

Simply upload a spreadsheet with your prompts and settings (such as model version, style, aspect ratio, color palette, and render quality) and generate images in bulk. For more details, read the full documentation here. You can even combine the power of Ideogram Batch Generation with other AI tools to efficiently iterate and inspire new design concepts, as discussed below in two distinct examples.

Example 1: Ideate logo designs from scratch

AI assistants, such as ChatGPT, can generate a CSV file for Batch Generation with just a template file and a text prompt. Here’s how:

• Download the sample template CSV file from the Batch Generation page.
• Upload the template to ChatGPT and ask it to produce a CSV file following the same structure.
• Provide ChatGPT with a design task and a few effective prompt examples to enhance the quality of the output.
• Download the generated CSV and upload it to Batch Generation. Ideogram will handle the rest.

Below is a sample text prompt for a logo design task, which helped us generate four high-quality logo concepts included below.

Example 2: Experiment with ad creatives using a consistent tagline

Here’s an example of a text prompt for an ad design task. By experimenting with diverse variations that feature a consistent tagline "Where Imagination Meets Purrfection," marketers can conduct A/B testing to identify which creative style best engages their target audience.

Try it today!

We believe Batch Generation will be a powerful addition to your creative workflow. Try it today at ideogram.ai/batch.

Original source

Ideogram Canvas, Magic Fill, and Extend

Ideogram Canvas is now available on ideogram.ai! Access unlimited canvases and industry-leading Magic Fill and Extend with any paid plan, or get Ideogram Plus or Pro to edit uploaded images. Developers can use the Ideogram API to integrate Magic Fill and Extend into their applications. Watch a 2-minute tutorial video for Canvas here.

Today, we’re introducing Ideogram Canvas, an infinite creative board for organizing, generating, editing, and combining images. You can upload your own images or generate new ones within Canvas, then seamlessly edit, extend, or combine them using industry-leading Magic Fill (inpainting) and Extend (outpainting) tools. Use Magic Fill and Extend to bring your face or brand visuals to Ideogram Canvas and blend them with creative, AI-generated elements. Perfect for graphic design, Ideogram Canvas offers advanced text rendering and precise prompt adherence, allowing you to bring your vision to life through a flexible, iterative process.

Magic Fill

Magic Fill, our inpainting tool, allows you to edit specific regions of your images to replace objects, add text, fix imperfections, change backgrounds, and more. Ideogram Canvas enables zooming into portions of images to generate new high-resolution details. You can upload multiple images to your canvas and use Magic Fill to combine them into a unified composition. As shown below, to use Magic Fill: 1) select the region to edit, 2) adjust the generation window to focus on a specific area if needed, and 3) enter a text prompt to guide Magic Fill.

Extend

Extend, our outpainting tool, allows you to expand images beyond their original borders while maintaining a consistent style. Adjust the image composition and aspect ratio to fit any screen size without compromising the original content. Combine the power of Magic Fill and Extend to create stunning, creative content. As shown below, to use Extend: 1) adjust the generation window, and 2) enter a text prompt to guide Extend.

An AI-native interface for your iterative, creative process

Ideogram Canvas is a newly designed interface, built from the ground up, to enhance your iterative creative process with the magic of AI. Whether you're a seasoned designer or just starting your design journey, our flexible platform empowers you to bring your vision to life with remarkable speed and precision, expanding the boundaries of creative freedom.

Thank you to our community

We'd like to thank our beta testers, Ideogram Creators Club members, and the whole Ideogram community for helping us test Ideogram Canvas. Special thanks to: AbandonedMuse, aitext2023, alexwoow, atheist_american, caiojordao84, josiassevero, phil_ozoff, queencaffeine, snowcookies, sragy, trinityturret.

We're hiring!

We're on a mission to develop generative media models to help people become more creative. If our mission resonates with you, we’d love to hear from you. We're hiring for roles in AI research, engineering, marketing and finance. Join us to build the future of AI for creativity — please apply at Ideogram Jobs.

Oct 22, 2024

Ideogram AI

Original source

Ideogram 2.0 is now freely available to all users on ideogram.ai and our new iOS app! Premium features are available via subscription plans. Developers can now build with Ideogram 2.0 using our new Ideogram API.

We are excited to release Ideogram 2.0, our new frontier text to image model with industry leading capabilities in generating realistic images, graphic design, typography, and more. Trained from scratch like all Ideogram models, Ideogram 2.0 significantly outperforms other text to image models across many quality metrics, including image-text alignment, overall subjective preference, and text rendering accuracy.

Today's milestone launch also includes the release of the Ideogram iOS app, the beta version of the Ideogram API, and Ideogram Search. One year ago, we announced the formation of Ideogram AI to advance the mission of making everyone become more creative. In the past year, our users have generated more than 1 billion images. This vast collection is now publicly accessible through Ideogram Search, providing a rich source of inspiration for all users. Additionally, the Ideogram iOS app brings our cutting edge platform to mobile users, and the Ideogram API brings our technology at the most competitive pricing to developers looking to enhance their applications and workflows.

Enhanced Control over Image Style

You can now choose from a number of distinct styles, including Realistic, Design, 3D and Anime. These styles are crafted specifically to render unique genres of images and have significant influence on your creations. You can now generate images in any aspect ratio, including 3:1 and 1:3.

Realistic Style

The Realistic style of Ideogram 2.0 enables you to create images that can convincingly pass as real photos. Textures are significantly enhanced, and human skin and hair appear lifelike.

Design Style

The Design style of Ideogram 2.0 significantly boosts the accuracy of text in the generated images. This enables you to create premium graphic designs for greeting cards, print on demand, posters, illustrations, and marketing and social media content with long, stylized text. This advancement will substantially improve the efficiency of the workflows of graphic designers and businesses.

Color Palette Control

You can now generate images that adhere to your specific color palette. This enables granular control of visual tones for purposes such as artistic control and brand consistency.

iOS App

Ideogram now offers an official dedicated iOS application, bringing Ideogram's powerful image generation capabilities to users on-the-go. An Android application is planned for release at a later date.

Download now on the App Store.

Build on the Ideogram API

Our new API brings Ideogram's unique capabilities to developers and businesses looking to integrate Ideogram into their own applications. The API offers superior image quality at a lower cost compared to other offerings. We're excited to see the creative applications that will emerge from users of the Ideogram API.

Join our public beta here, and learn about pricing here.

Search through Community Images

Ideogram's library of public creations are now available for text-based search. Now easily accessible for inspiration are over 1 billion publicly viewable images generated by Ideogram users in the last year.

Advanced Prompting with Describe and Magic Prompt

Magic Prompt and Describe are significantly improved to boost the creative process. You can use Describe to generate a detailed text prompt based on any original image, use Magic Prompt to write creative variations of the prompt, then use Generate to bring those prompts to life as fresh new images. This combination allows for endless creative iterations and reimaginings of any visual concept.

Step 1: Input image

A vintage poster with the text "Me travel? ...not this summer" and "VACATION AT HOME" at the bottom. There's a relaxed man dressed in a blue shirt and brown trousers reclining in a red armchair, reading a newspaper with text "USA Government". Beside him is a table with a glass jar filled with golden coins and a bottle. Above the man, there's a framed portrait of a young man in a military uniform. The overall theme suggests the idea of enjoying a comfortable and leisurely stay at home during the summer months.

Step 2: Describe

Step 3: Use the description as the prompt to generate a new image

Industry-Leading Image Generation

Ideogram 2.0 features significant advancements in image-prompt alignment, photorealism, and text rendering quality. Human evaluations consistently rate Ideogram 2.0 as a significant improvement over Flux Pro and DALL·E 3. Despite our improvement in image quality, our API pricing remains extremely competitive, lower than DALL·E 3 and on par with Flux Pro.

We're Hiring!

We're on a mission to develop generative media models to help people become more creative. If our mission resonates with you, then we’d love to hear from you. We're hiring for roles in AI research, engineering, and data operations. Join us to build the future of AI for creativity. Please apply at Ideogram Jobs.

Thank You to Our Community

We'd like to give special thanks to our beta testers, Ideogram Creators Club members, and the whole Ideogram community for helping us test Ideogram 2.0 and providing creative images used in this blog post. Special thanks to: AbandonedMuse, afranci, aitext2023, Alexeverywhere, alexwoow, Anaconda, atheist_american, Blademort1988, butterflygirl, Dee_nkai10, destiny_rose, dirtracerk27, doctorpangloss, donjorge, dshapiro, emerald5, foobly, funkyraf, Jaicraft, jcreations, josiassevero, Kamyar, Kryptomonark, manifold, masdevallia, mysocialeq, nagelkarle, neocandor, phil_ozoff, philmcc, queencaffeine, red_poppy__, Redrojo, saletlux, sragy, successfuelmedia, tanaj, thegrimreaper79, thelasertwins, thrivinglioness, toyartuk, trinityturret, tron, whatsthisbutondo, ylucic.

Original source

Enhance every pixel with state-of-the-art detail and photorealism

We are excited to release Ideogram Upscale, a new feature to enhance your images and increase their resolution. With Ideogram Upscale, you can elevate every pixel to perfection, with state-of-the-art capabilities in enhancing color, details, text rendering, and photorealism. Like all other Ideogram models, Ideogram Upscale is trained from scratch using our proprietary Generative AI technology.

Upscale is now available for Basic, Plus, and Pro subscribers on ideogram.ai. Ideogram Upscale not only increases the image resolution, but also rectifies artifacts present in the original image. Users can enhance existing Ideogram images and increase their resolution by a factor of 2X. Plus and Pro users can upload their own images, including AI-generated images from other platforms, and enhance the details and resolution.

Ideogram Upscale provide several ways for user to control the trade-off between resemblance to the original image and the introduction of new details:

• Resemblance: Control the degree of similarity to the input image.
• Detail: Adjust the level of creative enhancement.
• Prompting: Inform Ideogram Scale the desired content and style.

Experience the next level of creative image enhancement today, now available for Basic, Plus, and Pro subscribers on ideogram.ai!

Try it out

Upscale is available for Ideogram Basic, Plus, and Pro plans. But to celebrate the launch, we're making Upscale available for all Ideogram users (yes, even Free) until June 29, 2024, 12 noon ET! Here's how to use Upscale:

Step 1

Go to any Ideogram image you want to upscale, then press the Upscale button.

If you're an Ideogram Plus or Pro member:
With the Image Upload feature, you can start by uploading your own image from the prompt box and select Upscale.

Step 2 (Optional)

In the prompt box, you can adjust Resemblance and Detail to customize the behavior of Upscale. You can also edit the prompt to make small changes.

Step 3

Press Upscale! You'll see your original image on the left, and upscaled result will be generated on the right.

We're hiring!

If our mission of developing generative media models to help people become more creative resonates with you, then we’d love to hear from you. We're hiring for roles in engineering, research, design, and operations. Please apply at Ideogram Jobs.

Original source

Enhanced text rendering and photorealism

We're excited to release a major upgrade to Ideogram 1.0, our most advanced text-to-image model, alongside several new features, including Describe for image captioning, negative prompt, and image rendering quality and speed control.

All updates are now available at ideogram.ai!

Since the public beta of Ideogram in September 2023, we've welcomed more than 7 million creators and generated over 600 million images. Here is a highlight of a few Ideogram 1.0 images from our community:

Ideogram continues to offer state-of-the-art text rendering in AI-generated images, which is critical for design applications. The new upgrade reduces text error rates by an additional 15%, compared to the model released on Feb 28. In addition, we've enhanced image coherence and photorealism significantly — human raters prefer images generated by the upgraded model 30-50% more than the prior version in prompt alignment, image coherence, and text rendering quality.

Describe

Ideogram 1.0 introduced Magic Prompt, a creative assistant that automatically enhances, extends, and translates your prompts to result in beautiful, creative images. Today, we launch Describe, a captioning assistant that turns images into detailed descriptions. You can use such descriptions as excellent prompts to generate high quality images similar to the reference images, but you can also adjust the aspect ratio and descriptions to suit your needs. Describe supports two methods of input: any image from the Ideogram feed, or images uploaded from your device.

In the following examples, the reference image on the left is converted to a text description on the right. Then the description is used as a prompt to generate the image at the bottom.

Step 1 - Input image

A vintage New Year's greeting card. In the foreground, a young child dressed in a red coat and hat is blowing a large horn. The child stands on a snowy hill, with a picturesque town covered in snow in the background. Above the town, the night sky is dotted with stars, and there's a crescent moon visible. Above the child, there's a branch with a few green leaves. The words 'Wishing You A Happy New Year' are prominently displayed at the bottom of the image.

Step 2 - Describe

Step 3 - Use the description as the prompt to generate a new image

Step 1 - Input image

A serene and ethereal scene at what appears to be dawn or dusk. A misty golden haze blankets the water, reflecting the soft light. On the left, a bird with dark feathers and a long beak stands perched on a rocky outcrop, surrounded by sparse vegetation. In the center, a white egret or heron is captured mid-flight, its wings spread wide, with its long legs trailing behind. The bird seems to be approaching the rocky perch. On the right, a lone duck floats calmly on the water, its silhouette contrasting against the misty backdrop.

Step 2 - Describe

Step 3 - Use the description as the prompt to generate a new image

Step 1 - Input image

A vintage scene, possibly from the 19th century, set in a dimly lit room. A woman in a dark dress sits on a chair, looking down, while another woman in a lighter dress reclines on a bed, appearing to be in a state of rest or illness. A third woman, dressed in a white blouse and dark skirt, stands beside the bed, attending to the reclining woman. The room features a window with heavy drapes, a table with a vase, and a patterned rug on the floor. The overall mood of the image is somber, with the individuals appearing contemplative or concerned.

Step 2 - Describe

Step 3 - Use the description as the prompt to generate a new image

Rendering quality and speed control

Ideogram now allows you to switch between three rendering options: Fast, Default, and Quality. Fast mode prioritizes speed, and generates images in approximately 5 seconds. On the other hand, Quality mode prioritizes finer details, generating images in approximately 20 seconds. This control enables creators to quickly ideate using the Fast mode, and add quality to the best generations by switching to Quality mode with the same seed number. This feature is now available to all Ideogram subscribers.

• Fast (5 secs)
• Default (12 secs)
• Quality (20 secs)

Negative prompt

We now support negative prompting for you to tell the model what you don't want to see in the output images. You can use this feature to remove specific objects, or even refine the style of the image.

We're hiring!

If our mission of developing generative media models to help people become more creative resonates with you, then we’d love to hear from you. We're hiring for roles in engineering, research, design, and operations. Please apply at Ideogram Jobs.

Original source