Midjourney inpainting via API: mask_url on GPT Image 2.5
Midjourney's edit model now changes only selected pixels. Sume lists no Midjourney model; its masked inpainting runs on GPT Image 2.5 through mask_url.
Sume has no Midjourney model, so it cannot run Midjourney inpainting. What it does have is a masked edit: send a source image and a public HTTPS mask_url to ChatGPT Image 2.5. Its docs describe no outpainting mode.
Midjourney facts are from its September 24, 2026 post; Sume facts from the Image API docs and request code. Read 2026-09-30.
What did Midjourney change?
The post says inpainting and outpainting with the edit model are improved: targeted edits change only the pixels you selected, which allows repeated edits without degrading the image. The post does not announce an API.
How does Sume do inpainting?
Through mask_url, an optional public HTTPS mask for edits on openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst. Any other model refuses it: the request code throws an unsupported-parameter error for mask_url on models without that capability. Send the source in input_references and, on edits, prefer aspect_ratio: "auto"; the docs note that omitting the field is not the same as auto. A full request is in GPT Image 2.5 mask edits.
| Term | Midjourney post | Sume docs |
|---|---|---|
| Inpainting | Edit model; changes only selected pixels | mask_url on the two 2.5 ids |
| Outpainting | Edit model; improved | No outpainting parameter documented |
| References | Not stated in the post | Up to 16 images on the 2.5 ids |
Can I outpaint on Sume?
The docs describe no mode for extending a canvas, so do not plan on one. Reference URLs must be public HTTPS, and a request with a parameter the model does not list returns 400 unsupported_parameter.
Will repeated edits keep the rest of the image intact?
That claim belongs to Midjourney's edit model. Sume's docs make no equivalent promise for GPT Image 2.5, so compare the result with the source after each pass.
Sources
Related posts
More in Use cases
- New York synthetic performer ad law and AI avatar ads
NY S8420A requires a conspicuous disclosure of synthetic performers in ads. What it covers, what it exempts, and how to burn a line onto an avatar ad.
- Turn product photos into a silent video with Timeline static holds
Timeline 1.0 treats a still as a static hold, so product photos become a silent video with fades. Shopify lists MP4 up to 10 minutes; here is the request.
- Pull product stills from a video with the video-frames API
Video frames returns jpeg or png stills from a Sume-hosted clip at set seconds. Google lists 500 x 500 minimum, so extract from the clip before captions.
- Seedance 2.5 draft mode: a 480p-then-1080p loop on Sume
Higgsfield added Seedance 2.5 Draft Mode. On Sume there is no draft model: render at resolution 480p, pick a take, then re-request at 1080p.
Written by Sume