Text-to-Image Workspace

Route: /#/image/text2image

Use this workspace when you want to generate images from text only, with no reference image.

First-time rule of thumb

If both are true, this is usually the right page:

If you already have an input image, use Image-to-Image Workspace.

Even though the main mode here is “text only,” recent releases also connected reference-image-assisted prompt work into this workspace.

Near the left-side header, the current UI can expose two reference-image actions:

Replicate: ignore the current prompt and infer a reusable prompt plus variables from the reference image
Style Learn: keep your current subject goal, but learn style, composition, and color language from the image

These actions are especially useful when:

you already have a finished or style reference image and want to turn it back into reusable prompt material
you already know what subject you want, but want to borrow visual style without switching to image-to-image

Reference-image actions are not normal right-side generation. They depend on a separate image recognition model.

So if you want to use:

you need to configure an image-recognition-capable model separately in model management.

If that model is not configured, normal text-to-image generation can still work, but the reference-image actions will not be fully available.

The left side edits the image prompt itself.

The left side uses a text model, not an image model.

The right side tests:

If you use the reference-image actions, you can think about the workflow as three different steps:

reference-image actions: pull prompt clues from the image
left-side analysis / optimization: rewrite those clues into a cleaner prompt
right-side testing / comparison: check whether the real images now match the goal

If your starting point is a reference image, a better sequence is: