Ideogram 4.5: Precise Edit, Closed Weights, and What the Vendor Has Published

Ideogram 4.5 is Ideogram's precise-edit image model, live since 30 September 2026. Two questions bring people here, so both get answered up front. The first is what the model does: it edits an existing image and returns the result at that image's own width and height, with the pixels the edit did not touch copied exactly rather than re-rendered, which is the property the vendor is selling under the heading "precise edit". The second is whether you can download it. You cannot. Ideogram 4.5 is a closed, hosted model; the vendor's open-weights programme, which is real and has published repositories, currently covers Ideogram 4.0 and stops there.

Everything below is drawn from Ideogram's own model page, its developer documentation, its licensing page and its public Hugging Face and GitHub organisations, read on 13 October 2026, plus one independent leaderboard. Where the vendor publishes nothing, this page says so rather than filling the gap.

What Precise Edit actually does

The vendor's own one-line description is unusually concrete: "Tell Ideogram 4.5 what to change and it leaves everything else alone." The developer documentation spells out what that means in pixels — the output always matches the input image's width and height, and "pixels the edit did not meaningfully change are copied exactly from it." That is a stronger statement than the usual "preserves detail" phrasing, and it is the difference between an editor that re-synthesises a whole frame around a small change and one that patches a region.

The model page frames the same idea as a claim about drift: "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." The API overview adds a third piece — "pixels the edit doesn't touch are copied exactly from your image, and the result comes back at your image's own width and height" — and pitches it at four areas: targeted colour, material or detail changes; consistency across chained edits; reference-guided edits; and whole-image work in Studio.

Practically, the capability set is narrow and specific:

• Targeted edits — describe a change in natural language or a structured JSON prompt and keep the rest of the frame.

• Masked edits — supply a black-and-white mask to confine the change to one area. Black marks the area to edit, white the area to keep, and in-between values are rounded to the nearer of the two.

• Reference-guided edits — up to four reference images on the dedicated edit endpoint, or three if a mask is also supplied, to steer the edit with a product, material or style. No more than five source images on the generation endpoint, the first of which is treated as the one being edited.

• Edits at source size — on the edit endpoint the output is fixed to the input's dimensions. Images too large for the model are scaled down proportionally and returned as rendered. Images whose aspect ratio falls outside 1:6 to 6:1 are rejected outright.

The grouping the vendor uses for its own examples tells you who this is for: colour and lighting, text modification, product photography, interior design, architecture, restoration of old photographs, sketch-to-image, style reference, reframe and depth-to-image.

The open-weights question, answered

No page on the model's own stem answers this, which is why it gets its own section. Here is the sourced position as of 13 October 2026.

The position: closed. Ideogram 4.5 is served from Ideogram's own hosted API and its own apps. There is no checkpoint, no repository, and no licence published for it. The strongest evidence is negative and unambiguous: Ideogram's developer documentation, read today, describes every v2 endpoint as a hosted call to api.ideogram.ai with an Api-Key header, and no download path exists for 4.5 anywhere in it.

What is open is the previous generation. Ideogram's licensing page describes an open-weights programme for "Ideogram 4.0" — the vendor's own summary is "Quantized model weights on Hugging Face", "Apache 2.0 inference code on GitHub", "Fine-tunes and modifications permitted", with outputs remaining subject to an acceptable-use policy. Three tiers are documented: a free non-commercial licence for research, evaluation and personal projects; a self-serve commercial licence at $300 per month covering 10,000 images a month of self-hosted public quantized weights, rising through higher allowances, with the vendor stating plainly that this tier "does not include full-precision weights"; and an Enterprise tier which does include full-precision access.

The repositories confirm the split. Ideogram's Hugging Face organisation lists exactly three models — ideogram-4-fp8, ideogram-4-nf4 and ideogram-4-nf4-diffusers — and its GitHub organisation lists two repositories, ideogram4 and ComfyUI-Ideogram4. All five belong to 4.0. Nothing named 4.5 appears in either organisation.

There is one loose thread worth stating precisely, because it is the source of the confusion. Reporting on the launch by The Decoder on 1 October 2026 recorded that "Ideogram also says an open-weight release is coming soon." That is a third party relaying a statement of intent. It is not a published artefact, and as of today nothing has been released against it. Treat any claim that Ideogram 4.5 has open weights as unverified until a repository with 4.5 in its name exists.

Inputs, outputs and the size envelope

The modality picture is narrower than the model's fame might suggest, and it is worth being exact about it because the two entry points behave differently.

• Text in — a natural-language prompt or a structured JSON prompt, up to 10,000 characters in the app. Natural language is converted to a structured prompt server-side; valid JSON is used as written.

• Images in — JPEG, PNG or WebP, max 25 MB each. Up to five on the generation endpoint, up to four on the edit endpoint, three of those four when a mask is present.

• Masks in — one JPEG, PNG or WebP mask, same width and height as the source image, containing both black and white areas.

• Images out — a URL to a generated image plus its resolution, seed and a safety flag. Image and video links expire, so the vendor tells you to download anything you want to keep.

• Generation sizes — auto, source, or an exact WIDTHxHEIGHT. Without source images an exact size must be one of the supported 1K or 2K presets. With source images, an exact size must have both sides a multiple of 32 and at least 256 px, total no more than 2048 × 2048 pixels, and an aspect ratio no wider than 6:1. source returns the output at the first source image's own width and height. Source images must each fall between 1:6 and 6:1.

• Edit size — not a choice. The precise-edit endpoint returns the result at the source image's exact width and height. In the app, attaching a reference image swaps the aspect-ratio and resolution controls for a Size menu with Auto, Source image size, and a Custom width and height under the same 32-px-multiple, 256-px-minimum, 2048 × 2048-capped rules.

• Quality tiers — four in the API: very_low, which requires source images, then low, medium and high. The generation endpoint defaults to high without source images and medium with them. The app shows Very low (references only), Low, Medium and High, with High as the default on paid plans and Low on the free plan.

• Numbers and seeds — num_images per request, and a seed for reproducible generation. The app exposes 1 to 8 images per run.

Text rendering

Ideogram's reputation was built on typography, and the 4.5 documentation treats text as a first-class element rather than a side effect. The vendor's guidance is consistent across pages and mostly about restraint: put the exact words in quotation marks so the model treats them as literal text, mention the text early in the prompt because text named near the start comes out more reliably than text appended at the end, say where it appears, keep it short, and keep the scene simple.

The limits are stated as plainly as the capabilities: "Ideogram handles titles and short blocks well. Add body copy or full documents later in a design tool," and "Text comes out most accurately in English. Non-Latin scripts and accented letters can come out wrong or not at all." You also cannot pick a font by name — the model has learned what text styles look like, not a font library — so the documented approach is to describe weight, shape and mood instead.

The precise lever is the JSON prompt, which 4.5 shares with 4.0. Each text element gets its own text, desc for how it looks, color_palette, and bbox for exact placement, with up to 16 hex colours for the whole image and up to five per element. The vendor is explicit that plain language gets you close but the layout can shift run to run, and that a structured prompt is what holds it still — and equally explicit that very long JSON prompts can lose their trailing elements, so the most important elements go first.

Editing fidelity versus generation, in the vendor's words

Ideogram is not claiming this is the best image generator on the market. The model page sells one behaviour and demonstrates it, and the comparison it chooses is telling: a side-by-side of the same sequence of edits against three rival editors, one of them named outright — GPT Image 2.5 Sunburst — with the claim that those outputs "become unusable within a few edits" while Ideogram 4.5 "stays clean edit after edit."

Two things about that claim deserve to be separated. It is vendor-reported: the comparison is Ideogram's own demo, on Ideogram's own page, with Ideogram's own choice of images and edit sequence, and this page has not reproduced it. And it is specific, which makes it testable in a way that most launch claims are not. The claim is not "better taste" or "better photorealism". It is that the edit boundary holds. That is why the vendor's other demonstration — an edit performed on a 4,016 × 6,016 pixel source, 24.2 megapixels, without downscaling the file first — matters more than any leaderboard row: it is the case where an editor that re-renders the whole frame is most obviously unusable.

The generation endpoint is, by the vendor's own framing, the plain one: "Create images from a prompt with Ideogram 4.5 and 4.0, known for prompt fidelity and crisp typography." Notably, the app treats 4.5 as the editing entry point rather than the default generator — it is the only Ideogram model in the app that accepts reference images, and in the app's Auto mode, attaching reference images is what causes the system to pick a model that treats the prompt as an edit instruction.

How you get access, and what the API looks like

There are four documented routes, and they are not equivalent in what they cost you.

• The web and iOS apps — the model is a selectable model in the Image app, with Precise Edit reachable from the Edit button on any image. Precise Edit needs a paid plan or bought credits; the slower queue that paid plans fall back into does not cover edits or generations that use a reference image.

• The API — the v2 surface is documented and the model is in the URL: POST /v2/image/generate/ideogram-4-5 and POST /v2/image/precise-edit/ideogram-4-5, both multipart form data, authenticated with an Api-Key header. Results return inline by default; set async or supply a webhook_url to get a generation_id back and poll GET /v2/generations/{generation_id}. Webhook deliveries are signed with Ed25519 and verifiable against public keys published at api.ideogram.ai/v1/.well-known/jwks.json. Requests must be HTTPS, and private, loopback and cloud-metadata hosts are rejected.

• Other endpoints — Ideogram's documentation also routes upscaling (including Topaz models), background removal and replacement, describe, layerise and video generation through the same v2 pattern. Those are separate models under one key, not extra capabilities of 4.5.

• MCP — Ideogram publishes an MCP integration and a machine-readable documentation index at docs.ideogram.ai/llms.txt.

Two API details are worth knowing before you budget anything. First, dry_run=true validates and prices a request without running it — nothing is generated, stored or billed, and you get a price quote instead of an image. Second, a rejected request comes back with a machine-readable reason: insufficient_funds, subscription_required, daily_limit, priority_credit_required, inflight_limit or feature_limit, with the in-flight cap and queue name returned when the limit is the constraint. API billing is prepaid: credits purchased from $1 to $300, a balance ceiling of $300, and optional auto-recharge.

On price, be careful about which number you are reading. Artificial Analysis's board, independently of the vendor, lists $220.00 per 1,000 images for the Ideogram 4.5 (High) editing configuration and $100.00 per 1,000 images for the text-to-image configuration. Launch reporting on 1 October 2026 put the vendor-reported ladder at 0.8 to 22 cents per image across the quality tiers, and the documented tiering rule is that generation bills by the resolved output pixels — up to 1024 × 1024 as 1K, above that as 2K, and an auto size as 2K. Ideogram's own API pricing page renders its numbers client-side and returned no figures to a text fetch on 13 October 2026, so this page does not quote a vendor per-image price as vendor-confirmed.

The independent numbers, and the fact that they moved

Artificial Analysis carries Ideogram 4.5 (High) on both of its image boards. Read on 13 October 2026, the rows are:

• Editing — AA-Image-Editing v2.0: 1,062 Elo from 2,913 votes, a 95% interval of 1,052 to 1,072, ranked 24th with the board's own interval spanning 22nd to 26th, at $220.00 per 1,000 images, listed as released September 2026. The leaders are GPT Image 2.5 Sunburst at 1,184 and GPT Image 2.5 Flare at 1,165.

• Text to image — AA-Image-T2I v2.0: 1,016 Elo from 2,479 votes, an interval of 1,005 to 1,027, ranked 35th, at $100.00 per 1,000 images.

These are not the figures this site published on 1 October 2026, which recorded 1,064 Elo from 2,382 votes at 23rd on editing and 1,014 Elo from 2,247 votes at 34th of 167 on text to image. The board has moved: 2 Elo down on editing as the sample grew by roughly 530 votes, and 2 Elo up on text to image as its sample grew by roughly 230. That is the drift you expect from a small sample settling, and it is a reason to treat both rows as provisional rather than as a verdict. The confidence intervals remain wider than the leaders' — that is what a few thousand votes buys.

One caveat applies to both numbers and it is not a small one. Neither board asks voters to judge the tenth edit in a sequence against the first. What they measure is preference between pairs of images, which is not the property Precise Edit is claiming. The vendor's actual claim is currently unmeasured by anyone independent.

What is not published

This is a short list, and it is short on purpose. Ideogram has published the behaviour, the interface, the size rules and the licence position, and has declined to publish several things a spec sheet would normally carry.

• Architecture and parameter count for 4.5 — nothing. By contrast, the vendor's June 2026 technical post for Ideogram 4.0 describes "a 9.3B single-stream Diffusion Transformer, trained from scratch, with a vision-language text encoder and structured JSON prompts", names Qwen3-VL-8B-Instruct as that encoder, and details a 34-layer DiT. None of that has been published for 4.5.

• Training data and process for 4.5 — nothing.

• Any vendor benchmark — nothing. Every quality claim on the model page is demonstrated rather than scored, and the one head-to-head is the vendor's own demo against three named rivals.

• A drift or multi-turn fidelity benchmark — nothing, from the vendor or from a third party. The single property the model is built around has no published measurement.

• Latency figures — nothing documented. The only guidance is that higher quality "takes longer".

• A vendor per-image API price — the pricing table is client-rendered and did not return figures to a text fetch on 13 October 2026.

• Ideogram 4.5 weights — nothing published, and the vendor's own open-source navigation points at the 4.0 repositories.

What this means in practice

Three things are settled. Ideogram 4.5 is a closed, hosted model with a real and documented API, and its distinguishing capability is that it returns your image at your image's own size with untouched pixels copied rather than re-rendered. The open-weights line at Ideogram currently ends at 4.0, and anyone claiming otherwise is ahead of the repositories. And the independent scores place it in the middle of a crowded board at a price near the top of it — $220 per 1,000 edits on the editing board, against $60 per 1,000 for Grok Imagine Image 2.0 and $19.50 for MAI-Image-2.6-Flash.

What is not settled is the only thing that matters for the purchase decision. The claim Ideogram is making is a claim about the tenth edit, and no published measurement covers the tenth edit. If your work is a single prompt to a single image, this model's distinguishing feature is not the feature you are buying, and the text-to-image row — 35th, at $100 per 1,000 images — is the honest summary of what you get. If your work is colourway variants across a product catalogue, translating signage across a campaign, or restoring a photograph without repainting what survived, the vendor has demonstrated exactly your case and nobody has independently reproduced it.

That is the experiment to run, and it is cheap to scope: the same image, the same ten edits, run through Ideogram 4.5 and through whatever you currently use for multi-turn work, with the outputs diffed against each other turn by turn. Ideogram's API makes this unusually easy to instrument — num_images, seed and dry_run mean you can price the run before you pay for it — and the endpoint that matters is the edit endpoint, not the generation one. Note too that Ideogram 4.5 is not on our own catalogue: OrcaRouter's image line today is built from Open​AI's GPT-​Image models, x​AI's Grok Imagine image model and Googl​e's Imagen 4 family, and Ideogram is not among them, so this is a model you would call directly rather than through us. Placing it against a model you already route is still the comparison worth making, and the routing layer is what makes "run both and diff" a single-key exercise rather than a procurement decision.

The one development that would change this page fastest is a repository with 4.5 in its name. If Ideogram publishes the weights, the economics of every self-hosting question here change in a single day, and this page would need a rewrite rather than an update. Until then the answer to the question that brought most readers here is a plain no.