Leonardo.Ai in 2023: Generating Images, Remixing Models and Editing a Canvas

In February 2023, I am exploring Leonardo.Ai as an image-generation alternative to Midjourney. I like that it has a web interface and, in my experience with these examples, gives me fine-grained controls over the results. My preference is subjective and reflects the versions I tested at the time.

The home page shows featured models and images created by the community. Opening an image reveals its prompt, the model used and other settings, with options to remix it or use it as a starting point. The gallery is useful for seeing how different styles are made.

The featured models include photorealistic and paper-art styles. I select the paper-art model and enter a simple prompt for a Japanese village. The resulting images have the layered paper look I wanted, though a few examples cannot establish a reliable speed or quality comparison with another service.

The generation screen lets me select the number of images, dimensions and aspect ratio. Guidance scale controls how strongly the prompt influences generation, while step count affects the generation process and time. The interface also offers tiling and an image-to-image input; these are the controls shown in the 2023 version.

I use Stable Diffusion 1.5 for another prompt and add a negative prompt to discourage faces. The first images are reasonable, but I have seen better results and decide to build on a community creation. A negative prompt guides generation rather than guaranteeing that an unwanted feature will never appear.

I find a game icon set made with Leonardo Creative, choose remix and modify the prompt from a role-playing game toward a first-person game with a 3D style. The resulting grids contain multiple icons in each image. I like being able to generate a coordinated set together rather than assembling separately generated images.

The fine-tuned model browser includes platform models for different looks, including an isometric science-fiction building style. I choose it and ask for a tall industrial skyscraper, adding the isometric description to the prompt. Four examples appear, and I could continue adjusting the prompt and settings.

There are community-trained models as well, including sticker styles and cat portraits. The image pages let me follow creators, copy or remix prompts, use an image as an input or select the underlying model. The sticker examples already include a white border and shadow in the generated image.

I also show a model I trained earlier on 11 Pokémon images. The dataset contains a duplicate and some low-resolution pictures, and some of the generated Charizard examples look low resolution too. I suspect the source images contributed, although this single example does not establish the cause or guarantee that higher-resolution images will fix every output.

The prompt-generation feature can turn a basic idea into more detailed alternatives. I start with a robot and receive several descriptions, including a metallic figure illuminated by blue light. I choose one of the suggestions and use it to generate an image.

The AI Canvas was newly available in beta when I recorded this. It lets me select an area, generate content, expand the canvas and erase or add elements. In my attempts the canvas output sometimes looks stranger or blurrier than images produced through the regular generator, but the editing workflow is interesting.

Finally, I look at account settings, including interests and an NSFW content toggle, and show favorite creations. I called the service free and unlimited in the video, but plans and limits can change: Leonardo’s current pricing page lists a free tier with a daily token allowance. The interface and model availability shown here are a historical snapshot, not current instructions.