Stability's Stable Image edit tools, matched to Sume image routes

Three of Stability's Stable Image tools have a close match on Sume: inpainting through mask_url on ChatGPT Image 2.5, background removal through RMBG 1.0, and upscaling through Sume Image Upscale 1.0. Outpaint, search and replace, relight, sketch, structure and style control do not exist as named Sume routes.

The tool names below come from Stability's own page, Stability AI Image Models (read 2026-10-11). That page lists no prices, so this post compares capabilities only. The Sume side is read from the Image API docs and the MCP tools list.

What Stability's page lists

The page groups its tools into three families, with model names Stable Diffusion 3.5 (Large, Turbo, Medium), Stable Diffusion XL and Stable Diffusion XL Turbo.

• Edit: Erase object, Inpaint, Outpaint, Remove background, Search and recolor, Search and replace, Replace background and relight.
• Upscale: Creative, Conservative, Fast.
• Control: Sketch, Structure, Style.

Side by side with Sume

The table maps each family to the nearest Sume route. A match means the job is doable, not that the output is identical to Stability's.

Where the three matches work

Inpainting is a normal POST /v1/images call. mask_url is a public HTTPS URL and is accepted by ChatGPT Image 2.5 only, so a model that does not list it returns 400 unsupported_parameter rather than ignoring the mask. On an edit, aspect_ratio: "auto" matches the source image shape; leaving it out is not the same.

Background removal and upscaling are separate job routes, not parameters on the image call. Sume Image Upscale 1.0 takes a factor from 1 to 4 and costs $0.20 per image flat, and RMBG 1.0 costs $0.0225 per image. Both are exposed as MCP tools, so an agent can call them without custom code. Because they are jobs, you read the result from the standard job endpoints.

• Generation and edit billing is the fal list price times 1.25, rounded up, and a failed or cancelled request is not billed.
• Results come back as signed Sume-hosted URLs, not base64.
• A call that outlasts 30 seconds returns 202 with a job envelope instead of the image.

What to do about the gaps

If your workflow needs outpaint or a sketch-driven layout, Sume is not a drop-in replacement, and this post does not claim otherwise. The closest workarounds use only listed features: describe the extension in the prompt of a reference edit and pass a wider aspect_ratio, or feed several reference images to steer composition. Both give a new image, not a pixel-exact extension of the old one.

Check before you port. GET /v1/images/models lists every model with its capability descriptors, and a parameter missing from the descriptor is a 400, so one test call per model tells you what is real. For a broader view of open-weights versus hosted options, see Stable Diffusion alternatives in 2026, and for the mask workflow in practice, AI object remover with mask_url.

Picking a route

Use the mask edit when you can draw the region, RMBG when you only need a cutout, and Upscale 1.0 when the picture is right but too small. If your list needs more than those, keep that one step on Stability and use Sume for the rest, since each is a plain HTTPS call.