Ideogram 4.5 Precise Edit: Pixel Test and Prices

Ideogram 4.5 launched on 30 September 2026 as an image editing model built for one job: editing the same picture again and again without it falling apart. It is live in the Ideogram app, the Ideogram API and launch partners including fal, Pika, Krea, Picsart and ComfyUI. On fal it costs $0.008 to $0.22 per image depending on quality, and Ideogram says open weights are coming, with no date.

We checked two things you can verify without spending money. First, the claim behind the model: in Ideogram's own recolor sample, 72.3% of the pixels outside the edited shoes are byte-identical to the original, and 99.5% are within 8 of 255 levels. Second, the price. The new ComfyUI nodes charge 1.43x fal's price on three quality tiers, but only 1.30x on High, so a High edit costs $0.286 in ComfyUI against $0.22 on fal.

What Ideogram 4.5 ships

Ideogram exposes the model through two endpoints on its new v2 API. Generate creates images from a prompt, or edits up to five source images with one. Precise Edit is the new part: the output always matches the input's width and height, and "pixels the edit did not meaningfully change are copied exactly" from the source.

Two details matter in practice. Images with an aspect ratio beyond 1:6 or 6:1 are rejected, and large images are scaled down before editing. Ideogram's model page shows an edit on a 4,016 x 6,016 pixel source (24.2 MP) by cropping a region, editing it and stitching it back, which keeps the rest of a print-size file sharp.

Where to run it, and what each route charges

The same model costs different amounts depending on where you call it. fal's edit page lists $0.008 at Very Low, $0.03 at Low, $0.06 at Medium and $0.22 at High, and says image size does not change the price. Ideogram's own API reference says its billing is tiered by output size instead: up to 1024 x 1024 bills as 1K, anything larger as 2K. Its pricing page loads the actual figures in the browser, so we could not read them.

ComfyUI added three nodes in pull request 16689, shipped in the v0.38.1 tag the same day: Ideogram 4.5 Text to Image, Edit and Precise Edit. They run on Ideogram's servers and bill Comfy credits. We read the price badge in the node source.

The 1.43x factor is the same one we found on Recraft, Claude and GPT nodes in our ComfyUI v0.38 price check. High is the odd one out: $0.286 is exactly 1.43x $0.20, not $0.22, so ComfyUI appears to be pricing High from a lower base than fal charges. Neither Comfy nor Ideogram explains the gap. One more ComfyUI difference: its nodes default to Medium, while Ideogram's API defaults to High when you generate from text alone.

We checked the "untouched pixels" claim

Ideogram's API overview shows one sample chain: a pair of coral knit running shoes, then "Change the product's colour and finish to slate-blue knit", then two edits later with a gum sole and a pale lilac backdrop. We downloaded the three published images (1024 x 576 WebP) and compared them pixel by pixel.

In the recolor step, 9.2% of the frame changed by more than 32 levels on some channel: that is the shoes. Across the rest of the frame, 72.3% of pixels are byte-identical to the original, 93.7% are within 2 levels and 99.5% within 8. The mean difference is 0.67 of 255. The small differences sit in faint blocks across the background, which is what lossy WebP compression of two separate files looks like, plus a slight change in the shadow under the shoes. Counting the whole frame, shoes included, 65.7% of pixels did not change at all.

That is close to what Precise Edit promises, with two limits. These are Ideogram's own chosen samples, compressed for its docs, not a run on our images. And the third image replaces the backdrop, which is most of the frame, so pixel identity cannot test whether the shoe shapes held. We tried a shape metric and dropped it, because a copy shifted 8 pixels still scored 97%. Ideogram's model page compares 14 chained edits against GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2, saying the rivals "become unusable within a few edits". RuntimeWire notes that the comparison names no independent evaluator, protocol or failure rate.

Ten edits: Ideogram 4.5 vs Nano Banana and FLUX Kontext

Multi-turn editing multiplies whatever one edit costs. Here is a 10-edit chain at 2K on each model's fal page, plus Ideogram 4.5 through ComfyUI.

The price spread inside Ideogram 4.5 is wider than the spread between models. Very Low is the cheapest editing option in the table and High the most expensive. Precise Edit's promise to copy unchanged pixels is not tied to a quality tier in Ideogram's API reference; the tiers trade speed and price for render quality. So the sensible chain is cheap tiers while you iterate and one High pass at the end.

On quality, AlphaSignal reports a 1,351 score on Image Edit Arena, 18th overall, with 14th in 3D imaging and 15th in product design. That puts it in the pack on single edits. The case for Ideogram 4.5 is the tenth edit, not the first, and that is the part nobody has measured independently yet.

Workflow: a multi-turn product edit

A product colorway set is the job this model was built for: one hero shot, several variants that must match exactly. Here is the chain in the API, with the ComfyUI equivalent noted.

1. Start from a finished image. Use your own photo, or generate one with the Generate endpoint at High. Keep it near 2K (about 4 MP) or below, or it is scaled down first. In ComfyUI, load it into the Ideogram 4.5 Precise Edit node.
2. Quote before you run. Send the exact request with ?dry_run=true. Ideogram returns a price quote and bills nothing, which is the only way to see its direct price per request.
3. Make one change per turn. Write the edit as a single instruction ending in "Keep everything else exactly the same", the phrasing in Ideogram's own quickstart.
4. Iterate at Very Low or Low. At $0.008 an edit on fal, ten tries cost 8 cents. Set a fixed seed so a retry with the same prompt returns the same result.
5. Mask anything that must not move. If a logo or label drifts, add a mask (black where you edit, white where you keep) at the same size as the image.
6. Feed the output back in. Each turn's output is the next turn's image. Download every step: Ideogram's result URLs expire.
7. Finish with one High pass. Re-run only the final edit at High for delivery. In ComfyUI that is $0.286; on fal it is $0.22.

For a quick drift check of your own, difference each turn against the previous one in any image editor. Areas you did not ask to change should come out nearly black.

What is still missing

The weights. Ideogram's launch post says "open weights soon", and Ideogram 4 did follow that path: its GitHub repository went up on 28 May and quantized checkpoints followed on Hugging Face within days. For 4.5, the newest model in Ideogram's Hugging Face account on 1 October is still an Ideogram 4 checkpoint from 4 June. Until weights land, every route is a paid API.

Two smaller gaps. fal's edit page mentions a very_high tier for high-precision edits but prices only four tiers. And no outside tester has published a long edit chain comparing Ideogram 4.5 with GPT Image 2.5 or Nano Banana Pro. Ideogram's 14-edit video is the only evidence so far.

How we checked

We downloaded Ideogram's three published sample images and compared them with a short Python script (Pillow and NumPy), treating any pixel that changed by more than 32 levels as edited. We read the ComfyUI price badge from comfy_api_nodes/nodes_ideogram.py at the v0.38.1 tag, and took fal prices from its model pages on 1 October. We made no paid API calls, so we did not run our own edit chains, and the costs above are listed prices, not measured invoices. The script and output are kept with this article.

Frequently asked questions

What is Ideogram 4.5?

An image generation and editing model from Ideogram, released on 30 September 2026. Its focus is multi-turn editing: changing one thing at a time while keeping the rest of the image the same, without the color shifts and artifacts that build up over repeated edits.

How much does Ideogram 4.5 cost?

On fal, $0.008 per image at Very Low (edits only), $0.03 at Low, $0.06 at Medium and $0.22 at High, at any size. In ComfyUI the same tiers cost $0.01144, $0.0429, $0.0858 and $0.286. Ideogram's own API bills 1K and 2K outputs separately; use dry_run=true to get a free quote.

What is the difference between Edit and Precise Edit?

Edit goes through the Generate endpoint and can resize or recompose the image. Precise Edit always returns the input's exact width and height and copies the pixels the edit did not change straight from the source. Use Precise Edit when the rest of the image must match exactly.

Is Ideogram 4.5 open source?

Not yet. Ideogram says open weights are coming but has given no date, license or hardware requirements. Ideogram 4 is available as open weights, and the Ideogram 4 local setup still works for text-to-image.

Can I use Ideogram 4.5 in ComfyUI?

Yes, from v0.38.1. Search for Ideogram 4.5 Text to Image, Edit or Precise Edit. They are Partner Nodes, so they need a Comfy account with credits and run in the cloud, not on your GPU.

Is Ideogram 4.5 better than Nano Banana Pro for editing?

For long edit chains, Ideogram's own comparison says so, but it is self-reported. On single edits it ranks 18th on Image Edit Arena. For a single edit, the arena puts 17 models ahead of it. If you make ten edits to the same image, try Precise Edit at Low first: ten edits cost $0.30 on fal.