Help center

Go to Samsa.AI

AI Training Guide

A comprehensive guide that teaches you how to train effective AI models in Samsa, covering everything from dataset preparation to model configuration. The guide focuses on practical tips for curating high-quality training datasets, including detailed instructions for image selection, and angle considerations, making it essential reading for anyone looking to create custom AI models.

Introduction to Model Training

Model training is the foundation of creating custom AI-powered image generation capabilities. When you train a Samsa model, you're teaching the AI to understand and reproduce specific subjects, whether they're people, objects, or artistic styles. The model will learn from the images you provide and can later be used to generate new images based on this learning.

Getting Started

To begin training a model in Samsa:

  1. Navigate to the Models section

  2. Select "Create New Model"

  3. Upload your training dataset

  4. Configure your model parameters

What Makes a Good Dataset?

1. Diverse Backgrounds

Background variation in your training images is crucial for several reasons:

  • Different backgrounds help the model distinguish between the subject and its environment

  • Varied settings teach the model what features are constant across contexts

  • Natural backgrounds are often better than artificial studio settings

  • Include both simple and complex backgrounds to improve model flexibility

For example, if you're training a model of a coffee mug, include images of it on a desk, in a kitchen, outdoors, and in different lighting conditions. This teaches the model to focus on the mug's constant features while understanding how it appears in different contexts.

2. Different Angles and Perspectives

The angles included in your training dataset directly determine what viewpoints your model can generate. This is a critical decision point in dataset creation.

info icon
Important: Only include angles that you want your model to generate. Adding unnecessary angles can "unsettle" the model and reduce the quality of your results.

If you want your model to generate from multiple angles:

  • Start with cardinal angles (front, side, back, 45-degree views)

  • Include both eye-level and elevated perspectives

  • For products, demonstrate any functional angles

  • Capture any unique features from multiple viewpoints

  • Consider the most common use cases for your generated images

However, if you only want to generate images from the front, include only front-view training images. For example:

  • Portrait models intended for front-facing headshots only

  • Product photos that should always show a specific view

  • Brand assets that need to maintain a consistent perspective

  • Marketing materials with standardized viewing angles

Your choice of angles should align with your specific generation needs. Every angle you include tells the model "this is a valid way to show this subject." Make sure that matches your intentions.

3. Varied Lighting Conditions

Lighting dramatically affects how subjects appear, and your model needs to understand these variations:

  • Natural daylight (morning, noon, and evening)

  • Artificial indoor lighting (warm and cool)

  • Directional lighting that highlights specific features

  • Diffused lighting that shows overall form

  • Shadow interactions that demonstrate depth and dimension

Poor lighting in training images often results in models that struggle with shadow placement and realistic lighting in generated images.

4. Minimum Dataset Size

We recommend a minimum of 10 different training images per style, subject or product. Here's why this number matters:

Benefits of a 10+ image dataset:

  • Provides sufficient data for pattern recognition

  • Reduces model inflexibility and bias

  • Minimizes the impact of unwanted details

  • Allows for proper feature learning

Quality remains paramount - a smaller set of excellent images often outperforms a larger set of poor-quality ones.

5. Different Shot Types

To build a comprehensive model, include these essential shot types:

Full Shots:

  • Establish overall proportions and scale

  • Show complete subject in context

  • Demonstrate natural positioning and stance

  • Help the model understand full composition

Medium Shots:

  • Capture regular interaction distance

  • Show common viewing perspectives

  • Balance detail with context

  • Demonstrate typical use cases

Detail Shots:

  • Highlight specific features and textures

  • Show important small elements

  • Capture surface qualities and materials

  • Provide reference for fine details

6. Image Quality Requirements

All images should meet these technical specifications:

  • Minimum resolution: 1024x1024 px or 1920x1080 px

  • Sharp focus on the subject

  • Clear definition without blur

  • No pixelation or compression artifacts

  • Consistent quality across the dataset

Person-Specific Guidelines

For training models of people:

  • Maintain consistent appearance across images

  • Choose recent photos taken within a short timeframe

  • Include characteristic expressions and poses

  • Consider permanent features (like glasses) carefully and show them in all images if they should be trained on

Common Pitfalls to Avoid

  1. Inconsistent Data

    • Mixing different versions of products

    • Including temporary subject variations

    • Using inconsistent lighting or processing

  2. Limited Perspectives

    • Only showing front views when multiple angles are needed

    • Missing important angles

    • Insufficient environmental variety

  3. Poor Type Selection

    • Over-restrictive type definitions

    • Too general descriptions

    • Conflicting attributes

Professional Services

For enterprise clients requiring specialized model development, Samsa offers professional training services including:

  • Custom dataset curation

  • Optimized type configuration

  • Performance tuning

  • Integration support and success program

Contact enterprise@samsa.ai for enterprise solutions.

Best Practice Examples

Professional Portrait Model

Required elements:

  • Full face views (from intended angles only)

  • Various expressions

  • Different lighting conditions

  • Consistent personal style

  • Range of environmental contexts

Product Model

Required elements:

  • Required viewing angles (front-only or multi-angle as needed)

  • Detail shots of key features

  • Scale reference shots

  • Various use case scenarios

  • Different lighting conditions

Model Limitations and Expectations

Remember these key points:

  • Models can only generate based on learned patterns

  • Missing perspectives will be approximated

  • Consistent features in training will be persistent in output

  • Quality of training data directly influences output quality

Your model is ready for use when:

  • Generation attempts closely match training data

  • Consistent features appear reliably

  • Subject is recognizable from new angles

  • Style and quality remain stable across generations

Did this answer your question?
😞
😐
😁