Moodboard

The Moodboard is an infinite, zoomable canvas for building visual boards: character looks, location references, tone and palette studies, or a whole film's visual bible. Boards are saved as sessions you can switch between; the most recent opens automatically.

Working the canvasCopy Anchor Link

• Add images: Add Image lets you upload from disk or pick from the built-in Inspirations library.
• Generate: prompt from the bottom bar and images land directly on the canvas, using the same model catalog as the Image Generator.
• Prompt library: the Select a Prompt button on the left toolbar drops a curated prompt into the prompt box.
• Text: add labels and notes with font, size, color, and style controls.
• Arrange: drag, resize, multi-select with a drag-rectangle, reorder, delete. Zoom from 0.1× to 5×; hold Space and drag to pan.
• Crop: trim any image to a drawn rectangle.
• Upscale: enhance a selected image in place (20 credits).
• Download: one image or everything selected.
• Export as PNG: in the board's title menu, saves the whole board as a single image.

Uploading, arranging, text, cropping, and exporting are free. Only generation, upscaling, and training cost credits.

References from the canvasCopy Anchor Link

When your selected model supports reference images, any images you have selected on the canvas are used as references automatically. They appear as chips above the prompt. This makes iterating on a look natural: select the frames that work, prompt the next variation.

Custom model training (LoRA)Copy Anchor Link

Train a custom model on your own image pairs, for example a consistent character, product, or style edit.

1. Click Train LoRA on the bottom bar. The Train Model dialog opens. It's the same dialog on this tool and on an AI Studio project's Moodboard.
2. Pick the Model (Qwen Image Edit Plus — LoRA trainer) and enter a Training Name.
3. Upload Training Data (ZIP) of image pairs: photo_start.jpg + photo_end.jpg (before/after). Add extra references per pair as photo_start2.jpg, photo_end2.jpg, and so on, and set Reference Image Count (1 to 3) to the number of references per pair.
4. Give each pair its edit instruction: a caption file such as photo.txt, or type one instruction in the optional Default caption field to cover every pair without a file. Pairs with neither fall back to a generic caption, so give every pair one or the other.
5. Click Start Training (80 credits). It runs in the background; you can close the dialog.

When training completes, it's listed under Select Trained LoRAs. A full walkthrough is in Consistent characters.

Generating with your LoRAsCopy Anchor Link

1. Click Select LoRA on the bottom bar and pick up to 3 trained LoRAs. They appear under Selected LoRAs.
2. Select at least one image on the canvas. LoRAs run on Qwen Image Edit Plus LoRA, an edit model, so your selected images are the Base images it edits. Generating without one is blocked.
3. Prompt and generate as usual.

While LoRAs are selected, the model picker shows Qwen LoRA with its price and is locked. The price per generation is also shown under Selected LoRAs. Remove the LoRAs to pick another model again.

SharingCopy Anchor Link

Boards are private by default. In a workspace, the board owner can open Share from the board's title menu to make a board visible to the whole workspace or to specific members.