Niches · Image generation

Edit an existing image by instruction or mask, often over several turns

One company's docs show a full fit.

1companies, full fit
12companies, fit with a limit
26companies carry the tag
16share of the category's work
8first-hand accounts

What counts as done

An image where only the requested region or attribute changed, unmasked pixels visibly unchanged, from one call per edit with the input accepted as URL, base64 or file.

Measured, for our own tests: no quality loss after three chained edits. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: change the hairstyle in this photo, remove the person on the left, make this low-poly render look photoreal, extend this image to 16:9, now make the sky darker

Does the job 1

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Ideogram API D
Ideogram AI, Inc.
Ideogram 4.5 Precise Edit takes the image as a multipart file with an optional mask or box, copies untouched pixels precisely, returns the source's own size, and is documented for chained edits with minimal drift. source

Does the job, with a limit 12

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Adobe Firefly API D
Adobe Inc.
fill-async takes an image and mask and fills only the masked area, but inputs must be an uploadId or a pre-signed URL on four allowed storage domains (no base64), and nothing covers quality across chained edits. source
Black Forest Labs FLUX API B
Black Forest Labs Inc.
FLUX 3 Image edits take one to ten images as URL or base64, with bounding boxes to change one element while the rest stays; FLUX.1 Fill takes a mask. Chained-edit quality rests on vendor examples, and acc-26 saw saturation rise after outpainting on FLUX.2 klein. source
Bria API BB
Bria Artificial Intelligence Ltd.
/v2/image/edit takes a URL or base64 image plus a mask on Fibo 2, and the docs say outside the mask the image is kept as it was, but masking is generative replacement and nothing covers quality across three chained edits. source
fal image models BB
Features & Labels, Inc.
Edit models such as Nano Banana Pro edit take image_urls (URLs or data URIs) with per-model schemas, but locality and chained-edit quality depend on the model chosen, and acc-15 saw quality degrade over successive edits on fal. source
FLORA B
FLORA AND FAUNA FLOWERSHOP, INC.
Hosted edit models include Ideogram 4.5 Precise Edit and its inpainting variant, but image inputs must be a public https URL or a FLORA asset, not base64, and locality depends on the model chosen. source
Leonardo.Ai API F
Leonardo Interactive Pty Ltd
Reference images go in guidances as an id, URL or base64 data, and edit-capable models (Kontext, Nano Banana, GPT Image) sit behind one endpoint, but no mask or inpaint parameter was found in the v2 schema and chained-edit quality isn't covered. source
OpenAI Image API BB
OpenAI OpCo, LLC
/v1/images/edits takes a file, an HTTPS or base64 image_url or a file_id, with markup or a mask for regional edits, but acc-02 saw a whole photo regenerated despite a mask and acc-03 grainy output after chained edits. source
Qwen-Image 2.1 (Model Studio) B
Alibaba Cloud (Singapore) Private Limited
The model card describes local edits by circles, painted annotations or separate masks with up to 10 references, but the Model Studio API reference couldn't be reached, so the API's mask and input forms weren't confirmed. source
Recraft API D
Recraft Inc.
/v1/images/inpaint takes image_url and mask_url as URL, data URL or multipart and regenerates only the white area, but only on recraftv3 models, and nothing documents quality across chained edits. source
Replicate image models B
Replicate, LLC
Edit models run through the same predictions API, but each model sets its own input rules, and acc-27 found wan-2.7-image, grok-imagine-image-2 and qwen-image-3 rejecting PNGs other models accept; locality and chained quality are per model. source
Seedream on BytePlus ModelArk C
BytePlus Pte. Ltd.
Reference images go in as a URL or a Base64 data string, and editing is prompt-driven, but there is no mask parameter, so nothing confines the change to a region, and chained-edit quality isn't covered. source
Stability AI Image API F
Stability AI Ltd
/v2beta/stable-image/edit/inpaint takes the image and a mask (or the image's alpha) as multipart files at fixed credits, plus search-and-replace and erase, but output is set to 4 megapixels and nothing covers quality across chained edits. source

Doesn't do this job 2

Listings in the category that do something else, with the reason.

ListingWhy
Iris-3B C
Speridlabs
The text-to-image model takes text prompts only; the released fine-tunes do depth and 4x restoration, not edits or masks. source
SurgePix API F
SurgePix
The public API reference documents only chunked upload and /tasks/remove-background; text translation and other edits are skills with no documented parameters for masks or regions. source

What stops agents finishing

  • mask accepted but the whole image regenerated
  • quality degrades with each chained edit
  • edit endpoint rejects the input format the agent holds
  • multi-turn editing works only with server-side stored state

Around it

Before: image-generation.text-to-image, storage.download. After: video-generation.image-to-video, storage.upload.

Where this comes from

Weight 16 of 100 in Image generation: 8 of 37 accounts; mask edits that regenerate the whole photo, grain after chained edits, input-format failures on OpenAI, fal and Replicate

The same as data: /api/v1/use-cases/image-generation.edit-existing-image.json, or the MCP tool get_use_case with id image-generation.edit-existing-image. Other jobs in this category: /api/v1/niches?category=image-generation.

Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.