> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryblend.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Image Gen
> Generate or edit images from a prompt and optional reference images, using the image model you pick.
The Image Gen node turns a text prompt, and any reference images you feed it, into a generated image.
You pick an image model in the panel, write your prompt, and connect whatever inputs the model exposes. Its handles are **model-aware**: change the model and the handles update to match what that model accepts.
## Inputs
The inbound handles are derived from the selected model, so the exact set changes per model. The two you'll always work with are the prompt and the model selector; reference-image and parameter handles appear when the model supports them.
| Handle | Type | Required | Notes |
| ------------------ | --------------------------------------------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Prompt | `text` | Usually | The main instruction. Most image models expose a primary prompt field. If it contains `{{variables}}`, each variable becomes its own `text` handle instead of one whole-field handle. |
| Reference image(s) | `image` or `file` | No | Appears when the model accepts image input (for edits, style references, etc.). Narrows to `image` when the field is image-only; otherwise it's the generic `file` handle. Accepts **multiple** incoming connections. |
| Model parameters | `text` / `number` / `boolean` (per parameter) | No | Any other connectable setting the model exposes (size, aspect, count, and so on) gets its own handle so you can drive it from upstream nodes. |
| Model | `model` | No | The synthetic handle for the model selector. Connect a [User Input](/workflow/nodes/input) here to let callers choose the model at run time. |
Inputs are on the **left**. File/reference handles accept several edges at once; single-value handles take one. Standard connection rules still apply.
## Outputs
| Handle | Type | Notes |
| ------ | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------- |
| Image | `image` | The single output, on the **right**. Carries the generated image(s) — a model that returns several variations emits them all through this handle. |
One model call that returns several variations usually produces one output item with several image parts.
If the Image Gen node receives multiple input items from a Batch or structured list, it runs once per item and produces separate output items.
## Settings
The panel mirrors the model you choose:
* **Model** — a selector limited to models that output images. Switching models remaps the parameter fields and handles; compatible values you'd already set are carried over.
* **Prompt** — the primary text field for the selected model. Type `{{variable}}` tokens to expose per-variable input handles for templating.
* **Reference images / files** — file pickers for any image or file parameters the model supports, so you can set them directly on the node or connect them.
* **Model parameters** — the remaining exposed settings (such as count, size, or aspect ratio) render as their native controls. Less-common ones are tucked under a collapsible **Advanced Settings** section.
Only the controls the chosen model actually exposes appear. There is no fixed parameter list — a minimal model shows just a prompt, while a richer one adds size, count, and reference inputs.
## When to use
Reach for Image Gen whenever a step should produce a picture — generating art or product shots from a description, or editing/restyling an incoming image when the model accepts a reference. Pick the model that fits your quality, speed, and cost needs.
## Connects to
* **Upstream:** a [User Input](/workflow/nodes/input) feeding the prompt, model, or a reference image; an [LLM node](/workflow/nodes/llm) producing a refined prompt; any node emitting an `image`/`file` for the reference handle.
* **Downstream:** a [User Result node](/workflow/nodes/result) to return the image from your published workflow, or any node that consumes an `image` (or a `file` target).
## Notes / limits
* **Runs on the server.** Image Gen executes server-side. You can run it on its own to preview output without running the whole workflow.
* **Handles follow the model.** Changing the model can add, remove, or retype handles; edges pointing at a handle that no longer exists are cut automatically.
* **Multiple references allowed.** Image/file input handles accept more than one connection, so you can combine several references for models that support it.
* **Item behavior follows upstream input.** A Batch, multi-select Choose Output, or LLM structured list makes Image Gen run once per incoming item. See [Lists & single values](/concepts/lists-and-values).
## Next steps
How nodes, handles, and connections fit together on the canvas.
The same model-driven pattern, for generating video instead of images.