Custom model training, without owning a GPU
This is the feature that separates Leonardo from most hosted competitors, and the reason to choose it.
Upload a set of images and it trains a model on that subject or style, which you then generate from repeatedly. It is the hosted equivalent of training a LoRA in Stable Diffusion, without the hardware, the environment or the learning curve.
What decides the result is the consistency of the source set, not its size. Twenty images of the same character from varied angles, in varied lighting, against varied backgrounds will produce a better model than two hundred inconsistent ones. Common failures: every image shot in the same lighting (the model learns the lighting as part of the subject), backgrounds that repeat (it learns the background), and mixed subjects (it learns an average of both).
Training takes time and consumes credits, so getting the source set right before starting is worth an hour of preparation.
Choosing a model before writing a prompt
Leonardo offers a catalogue of fine-tuned models rather than one house style — separate models for realism, anime, illustration, concept art, product photography and game assets.
Selecting the right one is most of the styling decision, and it is a different skill from prompt writing. A photorealism prompt sent to an illustration model produces an illustration of your description, and users who fight this are usually fighting the wrong layer.
Production tooling around the generation
Beyond generation there is a canvas for inpainting and outpainting, upscaling, background removal, and generation of texture and tileable assets.
Tileable textures deserve a mention: for game and 3D work, a texture that repeats without a visible seam is a specific technical requirement that general image generators do not address at all, and having it in the same tool as the concept art is a real convenience.
Tokens, and how fast they go
- A browser and an account. A free tier with a daily token allowance that renews, which makes it genuinely evaluable — unusual at this capability level.
- An internet connection. Nothing local, no weights.
- For custom training, a consistent source set — typically dozens of images.
- Check current terms on commercial use and on whether free-tier generations are public before using output commercially.
The daily renewal is the part worth planning around. It suits regular light use better than an occasional heavy session, and running out means waiting rather than paying a small top-up.
Where it sits on quality
Peak image quality trails Midjourney and the leading models. Nobody switches to Leonardo because its images are the best.
They switch because they need forty consistent items rather than one exceptional one, and because custom training is available without hardware. Judged as a single-image generator it loses; judged as an asset pipeline it is well-shaped.
Ownership changed in 2024
Canva acquired Leonardo in 2024. The product continues to operate independently, and day-to-day use is unaffected.
It matters for planning. Strategic direction now belongs to Canva, and the long-term relationship between the two products — whether Leonardo remains standalone, becomes a Canva feature, or is repositioned — is not settled. For a hobbyist that is irrelevant; for a studio building a pipeline around custom-trained models, it is worth knowing that those models live in a product whose future is another company’s decision.
Who it serves
Game developers and indie studios producing sprites, items, environments and concept art at volume, where consistency across a set matters more than any single image.
Anyone who wants custom training without hardware — a recurring character, a product, a brand style — and would rather not learn a local pipeline.
People who want more control than Midjourney offers while staying hosted: model choice, negative prompts, guidance strength and seeds are all exposed.
It is a weaker fit for one-off atmospheric images, for confidential work on free tiers, and for anyone needing local generation.
Why studios pick it
- Custom model training with no GPU — the most practical route to a consistent recurring subject.
- A model catalogue, so style selection precedes prompt writing.
- A renewing free tier, rare at this capability.
- Production tooling — canvas, upscaling, background removal, tileable textures.
- Exposed parameters including seeds, for repeatable results.
Its limitations
- Peak quality trails the leaders for a single striking image.
- Token pricing that iteration consumes quickly; the free allowance goes faster than expected.
- Custom training depends on source-set discipline — a poor set yields a poor model, and the credits are spent either way.
- Ownership changed in 2024, so long-term direction is Canva’s.
- Hosted only, with no weights and no offline use.
Your trained models do not leave with you
This is the important one. Images you generated are yours to download; the custom models you trained are not exportable. They exist in Leonardo’s system and stop being available if you stop using it.
For anyone who has invested in training several models on characters or products, that is real lock-in — and the clearest practical argument for the local equivalent, where a trained LoRA is a file on your disk that works in every compatible tool.
Mitigation, if you use it seriously: keep the source image sets you trained from. Retraining elsewhere is possible if you still have the inputs, and impossible if you do not.
Comparable asset tools
- Midjourney — better images, no custom training, no free tier.
- Stable Diffusion — the local equivalent, with far more control and portable trained models if you have the hardware.
- Playground AI — comparable hosted generation with a stronger editing canvas.
- Recraft — the better choice when assets need to be vectors.
Compiled from Leonardo AI’s documentation and public sources. We have not hands-on tested this tool. Last reviewed 16 August 2026.
