Articles on: Advanced guides

Style models: teach Seelab your way of seeing

Obsidienne: one way of seeing, three subjects


A product model learns an object. A style model learns a way of seeing: a light, a palette, a material, a frame. Once trained, it brings that eye to whatever you ask for: a cocktail, a watch, a face, your bottle.


For this guide, we built a fictional brand style from scratch, Obsidienne: night, a direct flash, blacks that swallow everything, chrome, and a single colour, oxblood red. Dataset, training, tests, limits: everything you see here was produced on Seelab.


What a style model keeps


A style model does not keep the subjects of its images. It keeps what they have in common:


  • The light: here, a hard frontal flash, sharp shadows, reflections that snap.
  • The palette: deep blacks, desaturated tones, a single red accent.
  • The material: lacquer, chrome, wet surfaces, film grain.
  • The framing: tight, off-centre, with plenty of negative space.
  • The attitude: high and low angles, subjects caught rather than posed.


That is why a good style dataset shows very different subjects with the same treatment: the model understands that the subject changes and the eye stays.


Style, product or character? A style model does not reproduce a specific object. For a bottle, a bag or an armchair identical to the original, you need a product model; for a recurring face, a character model. Both combine with your style, as you will see below.


1. Write the art direction


Before the first image, put your style into a few lines. This text will guide the dataset, then the training brief. Obsidienne's:


Night, direct flash, deep crushed blacks, chrome and lacquer. One single accent colour: oxblood red, used sparingly. 35 mm film grain, tight off-centre crops, generous negative space. Never daylight, pastels or a white studio background.


A few ideas for writing yours:


  • Name the light: its source, its hardness, its direction.
  • Count your colours: a strong style often fits in two or three tones.
  • Say what the style refuses: that is often what makes it recognisable.
  • Give it a name: you will find it faster in the model selector, and so will your team.


2. Build the dataset


The model accepts up to 8 images. For Obsidienne, 8 images, 8 subjects: a material, a silhouette, an object, a fashion detail, an architecture, a still life, a hand, a portrait.


The Obsidienne dataset


These images were generated in Seelab with Nano Banana Pro, in 3:4 and 2K: one prompt per subject, followed by the same style description, word for word. That is 8 credits per image, 64 credits for the dataset. You can also start from your own photos: a campaign, a shoot, the work of a photographer you collaborate with (with the necessary rights).


What makes a good style dataset:


  • Varied subjects, identical treatment: if all your images show bottles, the model may learn "bottle" rather than "night flash".
  • Flawless consistency: one off-style image, brighter or more colourful, blurs the message. Six accurate images beat eight approximate ones.
  • Resolution: 1024 px minimum, ideally more. Vary the formats if you can.
  • No text or logos in the images, unless they are part of the style.


3. Train the model


In Custom models, click Create new model, then choose Styles & Photography.


The three model types


The training form


  1. Name the model: Obsidienne - Night flash editorial.
  2. Check the type: Styles & Photography.
  3. Import your images and click 👍 on the 3 images that best sum up the style. For Obsidienne: the car, the rose and the portrait, three different subjects with the same flash.
  4. Add a brief: your art direction, as text or PDF. Say what the model is for, describe the look, and what it must avoid.
  5. Click Start training.


For Obsidienne, training took about thirty seconds. The model then appears in your library with the Style badge.


Obsidienne in the library


In detail: Create and train a model and The training brief: text, PDF or both.


4. Test it without giving away the answer


The real test of a style model: short, ordinary prompts, without a single style word. If the Obsidienne look shows up anyway, the model is doing the work, not the prompt.


In Creation, pick your model in the selector, under My custom models, and simply write what you want to see. On the left, a market model alone; on the right, Obsidienne. Same prompt, word for word.


Before, after


The cocktail leaves its wooden bar for a black marble counter. The sneakers move from an afternoon pavement to midnight asphalt. The portrait keeps its white shirt and gold earrings, but takes on the dataset's oxblood lip. The prompt asked for none of it.


Then subjects the dataset never showed: an armchair, a watch, a bottle.


Unseen subjects


The armchair arrives in red velvet against black: the model chose on its own where to put its single colour. One generation costs 4 credits in 2K.


To push your tests further:


  • Start neutral, then add an intention: "top view", "close-up", "in the rain". The style follows.
  • Test subjects far from your dataset: that is where you see what the model really understood.
  • Too tame? Add a few words from your art direction to the prompt to push the look.


5. Your product, in your style


A style model alone reinvents objects: ask for a perfume and it draws one. To make it your perfume, pair your product model with an image of the style.


The method that works: select the product model in the selector, then add an image from your style dataset as a reference image. They are available directly in the Resource Hub, Datasets tab. The prompt describes the scene and points to the style of the image.


Product model + style image


The perfume bottle standing on wet black marble at night, a red rose lying next to it, in the exact photographic style of image 1: harsh direct flash, crushed blacks, chrome reflections, film grain.


The Koval Blackout bottle is intact, logo and base included; everything else is Obsidienne.


For a pixel-perfect composition, the Composer also blends your product into a scene generated in your style.


6. Make the style live


A style model quickly becomes the foundation of a whole production:


  • Campaigns and social media: a consistent series, whatever the subjects and whoever produces it.
  • Moodboards and pitches: show an art direction before the shoot, on the campaign's real subjects.
  • Workflows: an Image node with your style model, one prompt per subject, and a whole series comes out with the same eye.
  • Several styles per brand: a campaign style, an e-commerce style, an events style. One model each.


Cheat sheet


  • Write the style before training it: light, palette, materials, framing, and what it refuses.
  • 8 images maximum, varied subjects, identical treatment.
  • 3 key images 👍 that sum up the style, and a brief that says what the model is for.
  • Test without style words: if the look appears, the model got it.
  • Exact product + style: the product model and a dataset image as reference.


With the Seelab team


Your brand already has a visual signature? As part of its onboarding and support, the Seelab team can build the dataset from your campaigns, train and refine your style models, combine them with your product models and train your teams. Talk to your Seelab contact or write to us from the chat.


Learn more


Updated on: 12/11/2026

Was this article helpful?

Share your feedback

Cancel

Thank you!