Data generation

Every value goes through the same pipeline: a type provider produces something of the right type, field-name inference can override it with something semantically richer, and constraint decoration clamps the result to the column.

Field-type providers (MRO dispatch)

Each Django field class has a registered generator. Dispatch walks the field's method resolution order and the most specific registered class wins — so a SlugField (a subclass of CharField) resolves to the slug generator, not the char one. This fixes a whole class of bugs by construction, and means a custom field inherits its parent's generator for free.

Supported field types include:

Group	Fields
Text	CharField, TextField, SlugField, EmailField, URLField
Identity	UUIDField
Boolean	BooleanField (nullable too)
Integers	SmallIntegerField, IntegerField, BigIntegerField, and the Positive* variants
Numbers	FloatField, DecimalField
Temporal	DateField, DateTimeField, TimeField, DurationField
Network	GenericIPAddressField (IPv4/IPv6/both)
Structured	JSONField, BinaryField
Files	FileField, ImageField, FilePathField — see File & media fields

Django 5's GeneratedField is recognised and skipped (the database computes it). auto_now / auto_now_add columns are left to Django.

Field-name inference

A value that's the right type isn't always the right thing. Inference reads the field's name and, when it recognises a common semantic, draws from the matching provider:

Column name (examples)	Produces
email, email_address	a real email address
first_name, last_name, full_name, username	names from one persona
city, country, state, postal_code, street	a coherent address
phone, mobile	a phone number
company, job, occupation	company / job title
title, description, summary, bio	prose of the right length
slug	a hyphenated slug
price, amount, salary, total	a realistic amount (log-normal)
latitude, longitude	valid coordinates
age, year, quantity, rating, percentage	sensible numbers

Inference respects the field type: numeric rules only fire on numeric fields, text rules only on text fields, and fields with their own strong semantics (EmailField, URLField, SlugField, …) aren't second-guessed.

Why it matters

A CharField(name="city") becomes "Rotterdam" instead of "Lorem ipsum dolor". This one feature is what makes a seeded demo screenshot look like a real product.

Constraint decoration

Whatever the provider or inference produced is then made to fit the column:

• max_length — strings truncated at the exact boundary.
• max_digits / decimal_places — decimals quantized to fit.
• Integer ranges — respects the field type's storage range and any MinValueValidator / MaxValueValidator.
• choices — a value is picked from the choices, weighted (real status columns are 90% completed, not a uniform third each), and grouped choices are flattened correctly.
• GenericIPAddressField protocol — IPv4, IPv6, or a mix, per the field.
• USE_TZ — datetimes come out timezone-aware or naive to match your project.
• NULL injection — nullable, non-required columns are left NULL a configurable fraction of the time (--null-probability, default 0.1) so your optional code paths get exercised. Columns with a default are never nulled.

Realism

By default (--realism smart) numbers and dates follow believable shapes rather than random.randint:

• log-normal for prices/amounts,
• weighted choices for status-like fields,
• business-hour-weighted times,
• ordered lifecycle dates (see Row coherence).

Set --realism uniform to turn this off for dumb-fast, uniformly-random values.