Why fine-tune an image model?

If you’re using Midjourney, FLUX, or Dalle to generate images, you may struggle to create consistent photos of your actual products. That’s because these models are general-purpose*,* and don’t know anything about your brand.

EverArt makes it extremely easy to train AI on product or lifestyle photography. All you have to do is upload a few images of your product or visual style and wait 10-20 minutes for the training to finish. From there, you can submit simple prompts to generate stunning visual content for immediate use in marketing, advertising and editorial*.*

Even simple prompts will do.

Creating your first model

1. Tap “New Model.”
2. Select the type of model to train (Aesthetic, Product, or Face).
3. Upload your images into EverArt. Be extremely selective about what you upload (tips below)

Model types:

Product Models:

Choose this option when you want to train on a consistent physical or digital product. The thing that’s consistent in your image uploads (for training) should be the product, everything else can variate.

• When choosing images, here are some tips:

  • ONLY upload images of the product facing forward, so the logo and primary details are clearly visible
  • AVOID images of the product facing backwards, facing sideways, overhead, or leaning back
  • Ideally, your product is the only object in the frame, shown consistently (same SKU, color, size, angle, etc)
  • Upload at least 5 images of the product (the best models have 15+ images)
  • Ideally the photos are high resolution (can be iPhone) and are of your product in the real world
  • Avoid images where the product is too far away or too close, this can confuse the model’s sense of its proportions
  • Works best for furniture, apparel, fashion, luxury goods, accessories, digital goods, lifestyle photos, and more

• GOOD DATASET: clearly shows logo, perfect angle, consistent

• BAD DATASET: facing backwards, overhead shots, bad angles, inconsistent

(example dataset of a model trained on a can of pepsi)

Aesthetic Models:

Choose this option if there’s a photography style or mood board that you want to capture the aesthetic of. The thing that should be consistent in the data you upload for training is the visual “vibe” of your uploads (e.g. same lighting, coloring, or visual style). If you upload images that have too different of an aesthetic, it won’t work very well. The objects/subject can vary between the images.

• Aesthetic models allow you to replicate a certain photography style or brand aesthetic (e.g. illustration style), or even a mood board.

• Example: A model trained on Super 8 film stock.

  

Face Model:

Train a “face” model if you want to generate images of a specific person. Just upload 6-15 images of them, making sure the person you want to train on is the only person in each image. Try incorporating a mix of closer photos to reveal facial details and medium-distance photos to capture physique, don’t upload images where the person is too far away.

• Example: You are training a model on the woman below, with photos from varying distances.