Ideogram 4.5 Magic Fill and Extend vs Sume mask_url and aspect ratio
Ideogram's docs list Magic Fill and Extend as editing features. Sume has no Extend endpoint; here is what its mask and aspect-ratio fields cover instead.

Ideogram's documentation describes Magic Fill (replace or fill areas without regenerating the whole image) and Extend (grow the image beyond its borders). Sume does not publish endpoints with those names. What it does document is a mask_url for ChatGPT Image 2.5 edits, up to 16 references, and an aspect_ratio field with an auto value for edits.
Facts about Ideogram come from its full API documentation text, read on 2026-10-03; facts about Sume come from the Image API docs. That Ideogram page does not carry dates for September or October 2026, so treat it as the current description, not a changelog.
What does Ideogram document?
Ideogram 4.5 is described as the only Ideogram model that takes reference images, up to 5 per generation. With references you can set custom dimensions: at least 256 pixels per side, multiples of 32, up to 2048 by 2048 or a 6:1 ratio. Without references you choose an aspect ratio at 1K or 2K. Quality tiers are Very Low (references only), Low, Medium and High, and a prompt can be plain text or JSON up to 10,000 characters.
For editing the page lists Magic Fill and Extend, plus canvas and editor workflows with layers. It lists an API Playground for the Generate, Remix and Remove Background endpoints, and does not state prices for 4.5.
What can Sume do for an in-place fill?
ChatGPT Image 2.5 on Sume (openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst) accepts text-to-image, up to 16 references, an optional mask_url, and background set to auto, transparent or opaque. A mask edit is the closest equivalent to a fill: mask the region, describe what belongs there, and repeat what must not change.
The masks are guidance. OpenAI's guide, read on the same day, says the image and mask must match in format and size, both must be under 50 MB, the mask needs an alpha channel, and the model may not follow the exact shape. On Sume other families reject the field: Nano Banana, Seedream, FLUX and Grok return 400 unsupported_parameter when sent mask_url.
What can Sume do for an extend?
There is no documented outpainting operation. The nearest documented levers are the two size fields. On an edit, aspect_ratio: "auto" keeps the reference's shape, and omitting the field is not the same as auto. Setting a wider ratio than the source asks the model to produce a wider picture, but the docs do not promise that the original pixels are preserved at their position, so verify the result.
A safer route for a strict extend is to place your own picture on a larger canvas outside Sume, mask the empty area, and use a mask edit to fill it. That keeps the source pixels under your control.
How do the two sets of features line up?
Read the catalog before choosing: GET /v1/images/models lists each row's supported_parameters, and anything not listed returns a 400 rather than being ignored.
| Task | Ideogram documentation | Sume documentation |
|---|---|---|
| Fill a region | Magic Fill | mask_url on ChatGPT Image 2.5 only |
| Grow the canvas | Extend | No Extend endpoint; aspect_ratio auto or a wider ratio |
| Reference images | Up to 5, Ideogram 4.5 only | Up to 16 on ChatGPT Image 2.5; per-model on others |
| Custom size | Multiples of 32, up to 2048 or 6:1 | Multiples of 16, max edge 3840, ratio up to 3:1 on GPT |
| Prompt format | Text or JSON, 10,000 characters | Plain string prompt |
Which should you choose?
If Ideogram's editor tools are the workflow you want, use them at Ideogram, and read its pricing page for costs. If you want one API for generation, masked edits and references with Sume's job model, ChatGPT Image 2.5 covers the masked fill, and you handle canvas extension yourself. Either way, test one real image, since both vendors describe masks and sizes as guidance within limits.
Sources
Related posts
More in Comparisons
- Instagram's Devanagari caption fonts: what Sume burns for Hindi
Instagram added Devanagari and Bengali-Assamese caption fonts. Sume documents Latin and Hangul faces only, so test Hindi cues on one clip first.
- Reels run to 20 minutes; Sume's audio and caption limits to plan for
Instagram says Reels can reach 20 minutes but over 3 minutes is not recommended. The Sume Timeline audio, avatar and caption limits that matter at that length.
- Kling 3 vs Seedance 2.5 vs Wan 3.0: 5-second 720p price on Sume
A 5-second 720p clip on Sume: Wan 3.0 $0.625, Kling 3 with audio $1.05 (audio off $0.70), Seedance 2.5 $2.889. List x 1.25 for every model.
- Midjourney alternative for product stills by API: what to send on Sume
Sume's image catalog has no Midjourney model. For product stills it lists reference-based edit models instead; here is which id fits which job, and how to test.
Written by Sume