Create: Generate Images

The Create tab generates new images from a written description. You describe the image, pick a model, and the result is placed into your document as a new layer.

How to generate an image

1. 1
   Choose a model in the dropdown at the top of the tab.
2. 2
   Type a description into the text field. This description is called a prompt. Plain language works, you do not need special wording.
3. 3
   Optional: add reference images (see below).
4. 4
   Optional: set the number of variations to receive several results to compare.
5. 5
   Click Generate. Each result arrives as its own layer.

Reference images

Most models accept reference images. They guide the result, for example toward a certain style, product, or person.

• Add images with the image button (PNG, JPG, or WebP files) or take the content of the current layer with the layer button.
• In the Create tab every image you add is a reference. Create never edits one of your images. If you want to change an existing image, use the Edit tab.
• Drag the thumbnails to change their order. The order is passed on to the model.
• Each model has its own limit for the number of images (see the list below). Images beyond the limit are dropped.

A selection in your document does not become an editing target here either. On Create it only sets the aspect ratio, the automatic resolution, and where the result is placed.

Models in the Create tab

The list is served by Seam and can change without a plugin update. In the order the picker shows them:

• Nano Banana 2.1 (default): General purpose model by Google. Resolutions from 1K to 4K, fourteen aspect ratios including extreme panoramas such as 8:1, up to 14 reference images. Supports Preview and Commit, a thinking level, and an optional web search that lets the model look up current information.
• Nano Banana 2 Lite: The fast, cheap model of the same family. Renders at 1K only, so there is no resolution setting. Up to 14 reference images, thinking level available.
• Nano Banana Pro: The larger model of the same family. Resolutions 1K to 4K, up to 14 reference images, web search, Preview and Commit.
• Imagen 4 Ultra: Text only model by Google. Does not accept reference images. Five aspect ratios, up to 4 variations per run.
• GPT-Image-2: Image model by OpenAI. Quality from auto to high, three aspect ratios, up to 10 reference images, up to 4 variations per run. Supports Preview and Commit.
• Reve 2.1: Seventeen aspect ratios, more than any other model, from 4:1 panorama to 1:4 column. Up to 8 reference images, up to 4 variations per run. Seam picks the right Reve endpoint automatically from the number of images you supply.
• FLUX.2 LoRA: Text only model that can load your own trained models (see "My models" below). Guidance, steps, and an acceleration setting. Up to 4 variations per run.
• Seedream 5.0 Pro: Text only. Eight aspect ratios, up to 6 variations per run, the highest of any model.

Models that only change an existing image (FLUX.2 Max Edit, FLUX.2 LoRA Edit, Seedream 5.0 Pro Edit, Grok Imagine 2.0, Background Removal) live in the Edit tab.

Model settings

The settings panel shows only the options the selected model supports. Values you set are remembered per tab and per model for the rest of the session.

• Aspect ratio: The shape of the image, for example square, landscape, or portrait. On "auto" Seam derives it from your selection or canvas.
• Resolution: The output size. On "auto" Seam picks the smallest preset that covers the target area.
• Variations: How many images are generated in one run. Each variation is charged as its own image. The maximum depends on the model, from 4 up to 10.
• Quality (GPT-Image-2): Auto, low, medium, or high. This is the largest single cost lever in the plugin, low is about twenty times cheaper than high.
• Thinking level (Nano Banana 2.1 and Lite): How much the model reasons before it generates: minimal, medium, or high on Nano Banana 2.1, minimal or high on Lite. Higher takes longer.
• Web search (Nano Banana 2.1 and Pro): Lets the model look up current information while generating.
• Seed: A number that fixes the random part of a generation. The same seed with the same settings repeats a result. The shuffle button next to the label picks a new random number, an empty field means random.
• Guidance: How strictly the model follows your description. Lower values give the model more freedom.
• Steps: How many refinement passes the model runs. More steps take longer.
• Acceleration: None, regular, or high. Generates faster at some cost in detail.
• My models: Your own trained models, shown for FLUX.2 LoRA and FLUX.2 LoRA Edit. You train and manage them under Custom Models in the dashboard; the plugin only picks one from the list. Each session starts at "None".