class

RealNVPForImageGeneration

extendsImageGenerationModel

RealNVPForImageGeneration(config: RealNVPConfig)

source edit

RealNVP with the bits/dim training loss and .generate().

forward(x) returns a NormalizingFlowOutput whose loss is the mean negative log-likelihood in bits per dimension — the quantity the paper reports, and the scale-free form that keeps the gradient magnitude independent of image size. generate(n_samples) draws $z \sim \mathcal{N}(0, I)$ and returns $f^{-1}(z)$ in a single parallel pass, back in [0, 1] pixel space.

Parameters

configRealNVPConfig

Architecture, prior and preprocessing knobs.

Attributes

realnvpRealNVPModel

Underlying bijection, exposing encode / decode / log_prob / bits_per_dim.

Notes

Reference: Dinh, Sohl-Dickstein, and Bengio, "Density Estimation Using Real NVP", ICLR, 2017 (arXiv:1605.08803). Reported bits/dim: 3.49 (CIFAR-10), 4.28 / 3.98 (Imagenet 32 / 64), 2.72–3.08 (LSUN), 3.02 (CelebA).

Examples

>>> import lucid
>>> from lucid.models.generative.realnvp import (
...     RealNVPConfig, RealNVPForImageGeneration,
... )
>>> cfg = RealNVPConfig(sample_size=8, num_scales=2, residual_blocks=1,
...                     base_dim=8)
>>> model = RealNVPForImageGeneration(cfg).eval()
>>> out = model(lucid.rand((2, 3, 8, 8)))
>>> out.loss.shape                      # scalar bits/dim
()
>>> model.generate(n_samples=3).samples.shape
(3, 3, 8, 8)

Used by 2

Constructors

dunder

init

→None

__init__(config: RealNVPConfig)

source edit

Instance methods

forward

→NormalizingFlowOutput

forward(x: Tensor)

source edit

generate

→GenerationOutput

generate(n_samples: int = 1, temperature: float = 1.0, device: str | None = None)

source edit

Sample images by inverting a prior draw.

Parameters

n_samplesint= 1

Number of samples to draw.

temperaturefloat= 1.0

Multiplies the prior sample before inversion. Values below 1 trade diversity for typicality — a standard trick for flows, not part of the paper.

devicestr= None

Where to allocate the prior sample. Defaults to the device the model's parameters are on.

Returns

GenerationOutput

samples of shape (n_samples, C, H, W) in pixel space — the logit stage is undone on the way out. The squash is not saturating, so values land just outside [0, 1] (within ±(1 - c) / 2c); clip before display.

Examples

>>> import lucid
>>> from lucid.models.generative.realnvp import (
...     RealNVPConfig, RealNVPForImageGeneration,
... )
>>> cfg = RealNVPConfig(sample_size=8, num_scales=2, residual_blocks=1,
...                     base_dim=8)
>>> model = RealNVPForImageGeneration(cfg).eval()
>>> samples = model.generate(n_samples=3).samples
>>> samples.shape
(3, 3, 8, 8)
>>> c = cfg.data_constraint
>>> pad = (1 - c) / (2 * c)
>>> bool(((samples >= -pad) & (samples <= 1 + pad)).all().item())
True
>>> model.generate(temperature=0.0)
Traceback (most recent call last):
    ...
ValueError: temperature must be positive, got 0.0

>>> import lucid >>> from lucid.models.generative.realnvp import ( ... RealNVPConfig, RealNVPForImageGeneration, ... ) >>> cfg = RealNVPConfig(sample_size=8, num_scales=2, residual_blocks=1, ... base_dim=8) >>> model = RealNVPForImageGeneration(cfg).eval() >>> out = model(lucid.rand((2, 3, 8, 8))) >>> out.loss.shape # scalar bits/dim () >>> model.generate(n_samples=3).samples.shape (3, 3, 8, 8)

>>> import lucid >>> from lucid.models.generative.realnvp import ( ... RealNVPConfig, RealNVPForImageGeneration, ... ) >>> cfg = RealNVPConfig(sample_size=8, num_scales=2, residual_blocks=1, ... base_dim=8) >>> model = RealNVPForImageGeneration(cfg).eval() >>> samples = model.generate(n_samples=3).samples >>> samples.shape (3, 3, 8, 8) >>> c = cfg.data_constraint >>> pad = (1 - c) / (2 * c) >>> bool(((samples >= -pad) & (samples <= 1 + pad)).all().item()) True >>> model.generate(temperature=0.0) Traceback (most recent call last): ... ValueError: temperature must be positive, got 0.0