Articles on: Seelab models

Preparing training images

The quality of a Seelab model depends first on the images you give it. A few well-chosen photos are better than a long series: the model needs to see your subject from several angles, always identical, in varied contexts.


How many images?


The creation form accepts 1 to 8 images, whatever the model type. The Minimum / Uploaded / Maximum counter shows where you stand.


Quality matters more than quantity: a few sharp, varied images are better than images that look very much alike.


The common rules


  • Different angles: vary the viewpoints (front, three-quarter, profile, slightly from above).
  • Different formats: mix portrait, square and landscape images.
  • Subject centered and whole: the subject is fully visible, at different distances.
  • 1024 px minimum: HD images, sharp and well exposed, in JPEG or PNG.
  • A single subject: always the same product, the same person or the same style.


Avoid: the back of the product, which is often useless; distorted perspectives (very close wide-angle shots), which can degrade the model; blurry images, group shots or partly hidden subjects; duplicates.


By model type


Products & Objects


  • A single product, always in the same version (same label, same cap).
  • Varied backgrounds: plain, in context, on a material.
  • Varied lighting, so the model understands the shape and the materials.


Portraits & Avatars


  • Always the same person or mascot, in the same or a similar outfit.
  • Varied poses and expressions: front, three-quarter, profile; serious, smiling.
  • Different backgrounds, so the background isn't learned as part of the character.


Styles & Photography


  • Images that share the same mood: palette, lighting, textures, graphic treatment.
  • Varied subjects (people, objects, landscapes), as long as the style stays identical.
  • No text, watermark or logo in the images.


Mark the 3 key images


Once your images are imported, click the 👍 icon on 3 images: the ones that best define the look to reproduce. They tell Seelab what the model should learn first.


The creation form with the imported images


Write the brief


With Import a brief, add a text, a PDF, or both. The brief tells the model what the images can't:


  • The text brief: what you want to do with the visuals (e-commerce, social media, campaigns), points of attention on the product (logo, label, color, signature detail) and the style you're after. Without a PDF, be more specific: materials, colors, dimensions, product texts.
  • The PDF brief: what you already have, such as the product sheet, the technical sheet or your brand guidelines. Composition, ingredients, dimensions: the model knows them, so you no longer have to repeat them in your prompts.


Example of a text brief: "Nord Audio Monolithe, portable wireless speaker. Use: e-commerce packshots, lifestyle scenes in Scandinavian living rooms, social media visuals. Points of attention: keep the grey woven fabric, the cork base and the small NORD AUDIO logo exactly as in the reference images. Style: soft natural light, warm minimalist interiors."


The more complete the brief, the more faithful the model is to your product and your brand. Examples and tips: The training brief: text, PDF or both.


Tips


  • Change the setting, not the subject: this is the golden rule of a good dataset.
  • Start from your best existing product photos rather than a new photo shoot.
  • Not sure? As part of its onboarding and support, the Seelab team can prepare the dataset with you: Contact us in the Custom models space, or the chat.


Glossary


  • Dataset: the set of a model's training images.
  • Key images: the 3 images marked 👍, which best define the expected look.
  • Brief: the text and/or PDF that go with the images: planned uses, points of attention, style, product sheet, brand guidelines.

Updated on: 12/11/2026

Was this article helpful?

Share your feedback

Cancel

Thank you!