Image to Image

When to use this mode

Image to Image is the right pick when:

• You have source images you want transformed (restyled, recoloured, expressions edited, set in a new environment).
• You want to fan out variations of a hero asset without re-uploading it every time.
• You want filename-based pairing — hero-frontal.png matched to “hero frontal, painted style”.

The reference image library

When you pick the Image to Image mode tile, a Reference image(s) dropzone appears above the prompts textarea.

1. Drop image files into the dropzone, or click it to pick from your filesystem.
2. Each image becomes a numbered tile. Drag-and-drop to reorder; click an image’s filename label to copy it to clipboard.
3. The library persists across browser restarts. Per-mode: the images you upload for Image to Image stay with that mode and don’t bleed into Text to Image (which has no library) or Image to Video (which has its own).

Two ways to pair prompts with images

A toggle just below the prompts textarea controls how prompts find their images:

Auto-attach matching reference images (on by default)

“Match image filenames against words in each prompt and attach them automatically.”

The matcher reads each prompt as plain text and looks for any reference filename (minus extension) that appears as a substring. So with the library:

• hero-frontal.png
• hero-side.png
• villain-frontal.png

The prompt "hero frontal in a noir alleyway" matches hero-frontal.png. The prompt "villain frontal, oil painting style" matches villain-frontal.png. The prompt "a generic noir alleyway" matches nothing and runs with no reference image (you’ll see an unmatched warning on that row).

This is the right mode when filenames are meaningful — you’ve named your library deliberately, and you want one image per prompt to be picked by name.

Manual mode (toggle off)

Turn the checkbox off and a new control appears: Max input images per prompt (1–6).

The matcher is gone — instead, every prompt gets the first N images from the library, in order. Prompt 1 gets images 1–N, prompt 2 also gets images 1–N. This is the right mode for one-to-many fanout: you have one hero asset, you want 12 stylistic variants of it.

If you set Max to 3 with a library of 8 images, only the first 3 are sent to every prompt. Reorder the library tiles to control which ones get used.

Run it

After the library and matching are set, the rest is the standard loop:

1. Type prompts in the Prompts textarea, blank-line separated.
2. Check Refine for model (Speed / Quality) and aspect.
3. Click Run →.

Each prompt row in the live progress list shows tiny thumbnails of the reference images it received, so you can sanity-check pairing without leaving the panel.

Working with the live progress list

While a batch is running, the prompt list above the Run button doubles as a status display. For each prompt:

• The text of the prompt, with any matched image filenames subtly highlighted.
• Thumbnail strip showing which library images were attached.
• Status: queued → generating · N% → done or failed.
• An unmatched warning chip if auto-match found nothing.

After the batch ends, this status sticks around. Edit the prompt textarea (any keystroke) and the list flips back to authoring mode for the next run.

What you’ll want eventually

• Name your library carefully. 001-hero-frontal.png matches “hero frontal” but the leading 001- doesn’t hurt. Underscores and hyphens are both fine.
• Reorder the library to control manual-mode fanout. Drag the most-canonical image to the top.
• Use Chain prompts for iteration loops. With chain on, each output becomes the next prompt’s source — perfect for stylistic ladders. See Chain prompts .