Edit with instructions
Edit any image by describing the change
Type what should change. Everything else stays as it was. It runs free on your own GPU with Qwen-Image-Edit, or in the cloud with GPT Image.
Model
price of the model you pick, per run
$0
on your own machine
"Turn the winter scene into a lush green summer forest. Keep the man and the wolf." Qwen-Image-Edit, 12 seconds.
Plain words, no masks
Say "make it summer" or "remove the man on the left". No brush, no layers.
Up to 2 reference images
Put the jacket from one photo on the person in another, or copy a style.
Private on your machine
Run locally and the image never leaves your computer.
Reference
Inputs and parameters
The same on every surface: the desktop app, Claude through MCP and the API. Unknown parameters are refused, so a typo never runs with a default.
Inputs
| Name | Type | Limits | Notes |
|---|---|---|---|
| image required | image | 30 MB, 4 MP | Image to edit |
| references | image (list) | 30 MB, 4 MP, 1 to 2 files | Optional reference images (image 2, image 3 in the instruction) |
Parameters
| Name | Type | Default | Accepted | Notes |
|---|---|---|---|---|
| model required | string | · | up to 128 characters | An instruction-edit model (catalog models with engine qwenimage-edit, e.g. qwen-image-edit-2511). |
| instruction required | string | · | up to 2000 characters | The change to make, e.g. "replace the sky with a sunset". |
| negative_prompt | string | · | up to 2000 characters | |
| steps | int | · | 1 to 80 | Denoising steps (default: the model's recipe, 4 for the Lightning build of qwen-image-edit-2511). |
| guidance | float | · | 0 to 10 | True CFG scale (default: the model's recipe, 1.0 for the Lightning build). |
| keep_size | bool | true | true, false | Resize the result back to the input size (the model works at 0.5 to 1 MP: a small input is rendered at its own size, at least 0.5 MP). |
| seed | int | · | 0 to 4294967295 | Fixed seed for a reproducible result (default: random). |
Outputs
- imageimage
API
Run Edit with instructions from the API
curl https://imference.com/tools/instruct-edit \
-H "Authorization: Bearer $IMFERENCE_API_KEY" \
-F image=@image.png \
-F 'params={"model": "gpt-image-2.5-flare-edit", "instruction": "replace the sky with a sunset"}'
Files go in multipart fields named after the inputs; JSON works too, with {"inputs": {...}, "params": {...}} and https URLs or base64 data: URIs. The answer is a request_id: poll GET /status?request_id=... until it returns the outputs. Pay per call from a wallet instead at POST /ondemand/tools/instruct-edit. A run that fails is refunded to your credits. Details in the reference.
Specs
The details
- Price
- Same as a generation with the model
- Model
- Needs an instruction-edit model (model_type image-edit, GET /api/models?type=image-edit), e.g. qwen-image-edit-2511.
- Runs
- Local and cloud
- License
- Apache-2.0
- Tool id
- instruct-edit 1.0.0
- MCP tool
- edit_with_instructions
More tools
Often used with it
FAQ
Good to know
Which models can edit images?
Qwen-Image-Edit 2511 runs locally on an NVIDIA GPU with about 8 GB of VRAM, or in the cloud. GPT Image 2, 2.5 Flare and 2.5 Sunburst Edit run in the cloud only.
Does it keep the rest of the image?
Yes. Describe only what should change. The subject, the framing and everything you did not mention are kept.
Is it free?
Yes on your own GPU. In the cloud you pay the price of the model you pick, shown before you run it.