AI images · checked 2026-10-19

Can I vibecode Ideogram?

KINDA· weekend project

Ideogram 4 is the unusual case where the model itself can be downloaded: Ideogram published 9.3B-parameter open weights, and its own README says they render text better than any open model it benchmarked. An agent can wrap them in a local studio with a prompt box, a queue and a searchable gallery over a weekend. You still need a CUDA GPU, a download of about 16 GB and a magic-prompt step that calls Ideogram's hosted service or an LLM of your choice. The weights are non-commercial, the local prompt expander isn't the one on ideogram.ai, and the video and editing tools stay behind. Good for personal use, risky for client work.

$20/moprice

$240you'd save a year

weekendbuild time

—people replaced it

Before you start

What you need on hand.

DifficultyAdvanced · Real infrastructure or OAuth work.

TimeA weekend, mostly drivers, the 16 GB download and getting the magic prompt working

Claude Code, Codex or CursorAny of them can run the prompt.
An NVIDIA GPUThe nf4 build targets CUDA and the README we read gives no memory figure, so check the model card and test. Rent a GPU by the hour if you have none.
Python 3.11 and disk spaceThe nf4 download is about 16 GB; the fp8 build is about 27.5 GB but isn't supported through diffusers.
A Hugging Face read tokenAccept the Ideogram 4 license on the model page first, or the download fails.
An Ideogram or OpenRouter keyThe default magic prompt calls Ideogram's hosted API, which is free; the repo's system prompts can also drive your own LLM.

Before you buildIdeogram Free gives weekly credits but makes images public: try it, and read the weights' non-commercial license, before building a local copy.

1. Install one agent: Claude Code, Codex, Cursor.
2. Make an empty folder and open it in the agent.
3. Paste the prompt below, press enter, and answer its questions as it builds.

The prompt

Build me a local text-to-image studio to replace Ideogram, running Ideogram 4's open weights. Requirements:

- Python 3.11 with FastAPI and a single-page front end in vanilla JS (no build step). `python app.py`
serves http://127.0.0.1:7880 and must not bind to any other interface.
- Model: install the `ideogram4` package from https://github.com/ideogram-oss/ideogram4 and run the
nf4 weights (`ideogram-ai/ideogram-4-nf4`) on CUDA, calling its inference code the way
run_inference.py does. HF_TOKEN goes in .env after the license gate is accepted on Hugging Face; the
weights download once into ./models/. With no CUDA GPU, stop with a clear message.
- Prompting: the model expects a JSON caption, so expand the plain prompt with the package's magic
prompt (the default hosted `ideogram-4-v1` with IDEOGRAM_API_KEY, or an OpenRouter configuration
with MAGIC_PROMPT_API_KEY) and show the JSON in an editable box before generating. A "raw JSON"
switch skips the expansion.
- Generate page: aspect presets (1:1, 3:2, 2:3, 16:9, 9:16) at sizes that are multiples of 16, the
sampler presets V4_TURBO_12, V4_DEFAULT_20 and V4_QUALITY_48, a seed field with a random option, and
1-4 images per request rendered one after another with progress.
- A single worker with a queue: a second request waits and shows its position.
- Storage: sqlite3 in ./data/gallery.db (prompt, JSON caption, seed, size, preset, seconds taken, file
path) with an FTS5 index on prompts. PNGs go to ./data/images/YYYY-MM/ with the settings in the PNG
metadata.
- Gallery page: infinite-scroll grid, prompt search, favourite, delete (moves the file to
./data/trash/) and "reuse settings" to refill the generate form.
- Single user. No accounts, no telemetry, local-first; the only outbound calls are the model download
and the magic-prompt request. Optional HIVE_TEXT_MODERATION_KEY and HIVE_VISUAL_MODERATION_KEY are
passed to the model code if set.
- Out of scope: the fp8 and Apple builds, other models, LoRA training, editing, upscaling, video and sharing.
- README: GPU and disk needs, the measured seconds per 1024x1024 image for each preset on this
machine, that the weights are non-commercial, that unset Hive keys turn safety screening off, and
that results will differ from ideogram.ai.

Opens your agent with the prompt filled in, you press enter. Or copy it and paste it into an empty folder.

Share on X

Max MiskowSponsoredAI agents, automations and MVPs, built for your company.Visit site ›

Then keep going

Next prompts, same folder.

The first prompt builds the core. Run these one at a time in the same folder to grow it into your own Ideogram.

Add batch runs to the image studio in this folder. Requirements:

- A Batch page takes a CSV with columns prompt, seed (optional), preset (optional) and
  an "n" count, and shows the total number of images and a time estimate from the
  measured seconds per image before it starts.
- Jobs run through the existing queue; a crash or restart resumes at the first unfinished row.
- Each row's images land in ./data/images/batch-<name>/ and in the gallery with a "batch" tag.
- A cancel button stops after the current image.
- Update the README.

Add a layout designer to the image studio in this folder. Requirements:

- A canvas at the chosen aspect ratio where I draw boxes, label each as a text element or a
  subject, and type the exact text for text boxes.
- A palette field takes up to six hex colours.
- Build the JSON caption from the boxes and palette following the schema in the ideogram4
  repo's docs/prompting.md, and show it in the editable box before generating.
- Save layouts by name in SQLite and reload them.
- A test checks that every box is inside the canvas and that the caption validates.
- Update the README.

Add a compare mode to the image studio in this folder. Requirements:

- One prompt, one seed and one size, rendered with V4_TURBO_12, V4_DEFAULT_20 and
  V4_QUALITY_48 one after another.
- Show the three images in a row with the seconds each took, with a synchronised zoom.
- Save the three as one group in the gallery so they open together.
- Update the README with the timings measured on this machine.

Add saved styles to the image studio in this folder. Requirements:

- A style stores a name, a hex palette of up to six colours and a free-text style note.
- Choosing a style adds its palette and note to the JSON caption before generation, and
  shows exactly where they were added.
- Import and export styles as a single JSON file.
- Update the README.

What you lose

Ideogram's video and older image models, Magic Fill, editing with AI and image reference
Character consistency, batch generation and the community gallery
The hosted prompt expander: the open one gives different results from ideogram.ai
Rendering on someone else's GPUs: a local card is slower unless it is a big one
A licence for commercial work: the downloadable weights are non-commercial

Why people still pay

People pay for typography that works: legible text inside posters, logos and mockups, in a hosted tool with editing, character consistency and video models and no GPU to buy. The open weights are free to run, but under a non-commercial agreement, so work for clients still needs the paid plans or a separate deal with Ideogram.

Moatproprietary models infrastructure scale execution polish

Do the math

Ideogram or your own copy?

Over

Your usage

Count my time12h ×$/h

Ideogram, 1 yr$240

Your own copy, 1 yr$130running + build session

You keep$110pays for itself in 1 months

Estimate · Free on a GPU you own, apart from electricity, and the hosted magic prompt costs nothing. Renting a GPU by the hour adds up: Plus is $20 for 2,400 credits, about 240 High-quality images, so light use favours the subscription and heavy private use favours owned hardware. The build session is token usage on your agent plan.

Leaving Ideogram

Take your data, then cancel.

1 · Take your data out of Ideogram

1Open your Library, select the items (Generations, Uploads or a collection) and click Download in the bottom bar. Ideogram builds a ZIP named ideogram-images-<date>.zip, with PNG files for paid subscribers and JPEG otherwise.
2The ZIP carries a CSV with one row per item: file name, prompt, Magic Prompt text, aspect ratio, model, visibility, seed, style, colour palette and a link to the asset. Single-image downloads don't include it.
3A selection holds up to 250 items and only one bulk download can run at a time; to get a whole collection, open its three-dot menu and click Download.
4Videos download as one MP4 each, up to 50 in a bulk download.

2 · Cancel Ideogram

1Sign in to the account that holds the subscription, go to Wallet & Billing and click Manage plan & billing; the billing portal opens in a new tab.
2Click Cancel subscription, choose a reason and click Continue to cancellation, then click Cancel subscription on the confirmation page.
3Cancel at least 5 days before the renewal date shown as Renews on the plan card, or the next charge may still go through.
4Subscribed through the iOS app: cancel through the App Store, because cancelling on the website won't stop Apple's charges.

You keep the plan to the end of the paid period. Ideogram doesn't offer refunds, including for partly used periods.
Afterwards the account moves to the free plan, or to pay as you go if you have bought credits. Unused monthly credits end with the plan; credits you bought separately stay usable for 365 days.
Existing private images stay private, but new images made on the free plan are public.
Deleting the account permanently deletes your generations and data, so download the library first.

Checked 2026-10-19 against Ideogram's docs: 1, 2, 3, 4, 5

The escape hatch

Don't feel like building it? These are already free.

Ideogram 4 (open weights)FreeIdeogram's own 9.3B model with Apache-2.0 code, run from the command line; the weights are gated on Hugging Face and non-commercial, and the nf4 build needs a CUDA GPU.★ 2.9k · active Jun 2026 · self-host, real ops
Linux, WindowsVisit ↗ComfyUIOpen sourceGPL-3.0 node-based app for running image and video models locally, with community nodes for most open models.★ 139k · active Oct 2026 · one-click install
Windows, macOS, LinuxVisit ↗InvokeAIOpen sourceApache-2.0 image studio with a canvas, a gallery and node workflows for open models, free to run on your own machine.★ 29k · active Oct 2026 · one-click install
Windows, macOS, LinuxVisit ↗

All 4 free alternatives to Ideogram ›

The numbers

Ideogram pricing

Plan	Monthly	Annual, per month	What you get
Free	$0	$0	Weekly credits for eligible accounts; new images are public by default.
Pay as you go	custom	—	One cent per priority credit, bought when you need it; one running or queued generation at a time.
Plus	$20	$15	2,400 priority credits a month, unlimited slow credits, all image and video models, private generation, unlimited character consistency, quality export, up to 8 running and queued generations.
Pro	$60	$42	7,500 priority credits a month, the best cost per credit, batch generation, the largest queue and up to 32 running and queued generations, plus limited custom model training.
Team	$30	$20	3,600 priority credits per user a month, central billing and administration, early access to collaboration features, limited custom model training.
Enterprise	custom	—	Postpaid custom credits, private models trained on your visual data, volume API discounts and priority support.

Free tier: A small weekly credit grant and public images; editing with AI and image reference need priority credits, which only paid plans and top-ups provide.
Billing: Plus bills monthly at $20 or yearly at $180 ($15 a month); Pro at $60 monthly or $504 yearly. You can cancel any time and keep the plan to the end of the paid period. Cancel at least 5 days before the renewal date, because the next charge may still go through. Ideogram doesn't offer refunds, including for part-used periods, and has no discounts or free trials for subscriptions.
Hidden costs: Credits are spent per image by model and quality: on the table we read, a High-quality model 4.5 image costs 10 credits and 22 with source images, so Plus's 2,400 credits cover about 240 or about 109. Unused subscription credits expire at the end of each billing cycle; only top-up credits carry over. Once priority credits run out you fall back to the slow queue, which has no AI editing or image reference.
Source: ideogram.ai

Also on the list

Midjourney$10/mo · AI imagesNOT REALLY

Photoroom$12.99/mo · Photo editingYES

Canva$18/mo · DesignKINDA

Questions

Ideogram, briefly

Not answered here? Tell us and we'll add it.

Ask a question ›

Eligible free accounts receive weekly credits when signing in with Google, Apple or Microsoft (the amount varies), and images made on the free plan are public.

Kinda. The core is a weekend build with the prompt on this page, but real gaps remain: Ideogram's video and older image models, Magic Fill, editing with AI and image reference, Character consistency, batch generation and the community gallery.

About $20 a month on the Plus plan, $240 a year (checked 2026-10-19).

Ideogram's video and older image models, Magic Fill, editing with AI and image reference; Character consistency, batch generation and the community gallery; The hosted prompt expander: the open one gives different results from ideogram.ai; Rendering on someone else's GPUs: a local card is slower unless it is a big one; A licence for commercial work: the downloadable weights are non-commercial. If any of those matter to you, keep paying.

Yes: Ideogram 4 (open weights) (free), ComfyUI (open source), InvokeAI (open source).