The golden rule

Ten images that show the full range of your brand's visual world will always outperform fifty images of the same thing. Diversity of scene, composition, and context matters more than volume.

How many images to upload

Minimum: 10 images Optimal range: 12 to 20 images Maximum: 50 images

Uploading more than 50 images rarely improves results and can sometimes dilute the style's precision by introducing visual noise. Focus on selecting the right images rather than uploading everything you have.

What makes a good training image

1. It represents your visual identity at its peak Only upload images that you would use in a live campaign or on your website today. If an image looks outdated, off-brand, or substandard, leave it out. The model will learn from everything you upload — including the images you are not proud of.

2. It shows a variety of scenes and compositions If all ten images feature the same type of scene from the same angle, the model will learn a narrow interpretation of your style. A good dataset includes a range of compositional types — wide shots, close-ups, environmental scenes, product details, human presence, abstract moments.

3. It has consistent visual qualities across the set Your training images do not all need to look identical, but they should share common visual threads — the same light quality, the same colour relationships, the same tonal warmth. If some images are warm and editorial and others are cold and corporate, the model will be confused about what your brand actually looks like.

4. It is high resolution Upload images at the highest resolution you have. Minimum recommended resolution is 1024 x 1024 pixels. Higher is better. Low-resolution, compressed, or pixelated images produce lower-quality style outputs.

5. It has a clean subject Images with heavy text overlays, logos, watermarks, or complex graphic design elements make it harder for the model to isolate the visual style. Where possible, upload images where the subject and environment are the primary visual element.

What to avoid uploading

— Images from other brands or campaigns that do not represent your current identity — Screenshots of websites, apps, or documents — Images with heavy text, watermarks, or logo lockups — Low-resolution or heavily compressed images — Behind-the-scenes or reference images that are not final, polished outputs — Stock photography you do not own the rights to train on — Images taken in very different visual styles that contradict each other

Dataset types and specific advice

Photography datasets Focus on images that share the same lighting quality — colour temperature, direction, and intensity. If your brand photography is warm and editorial, all training images should reflect this. Include a range of subjects — people, products, environments, details.

Illustration datasets Upload illustrations that span the full range of your illustrative world — characters, environmental scenes, abstract moments, brand patterns. Include examples that show the full range of linework weight, colour palette, and compositional density.

3D render datasets Ensure all training images share the same render quality, lighting setup, and material aesthetic. Inconsistent render styles will produce inconsistent outputs. Include renders from multiple angles and contexts.

Mixed media datasets If your brand uses a combination of photography and illustration, you can train a style that blends both. Upload a balanced mix — aim for roughly equal proportions — and select examples where the two modes feel visually cohesive rather than contradictory.

How to evaluate your dataset before training

Before you upload, lay all your candidate training images out together — on a table, in a Figma frame, or in a folder view. Ask yourself:

— Do these images feel like they came from the same world? — Is there a consistent quality of light across most of them? — Do they show a range of scene types and compositional approaches? — Would I be proud to show all of these to a client today?

If the answer to all four is yes, your dataset is ready. If not, remove the weakest images and replace them with stronger ones before uploading.

After training — evaluating your results

Your first training output will show you how well the model has learned your style. Run 20 to 30 test prompts across a range of scene types before concluding that the style is production-ready.

If the outputs feel close but not quite right, the most common fix is refining the dataset — removing the weakest training images and replacing them with stronger, more representative ones — and retraining. Most style quality issues are resolved in the second training.