Modele dyfuzyjne

Diffusion models · id diffusion

Generowanie przez stopniowe odszumianie.

Generating by gradual denoising.

In one sentence

To modele tworzące dane przez stopniowe usuwanie szumu z początkowo losowego zapisu.

These models create data by gradually removing noise from an initially random representation.

Problem it solves

Chcesz tworzyć wiele różnych obrazów lub dźwięków, które zachowują poznane cechy rzeczywistych przykładów.

You want varied images or sounds that retain characteristics learned from real examples.

What it does

Nauka na zaszumionych przykładach pokazuje, jak je poprawiać; podczas generowania kolejne kroki kierują losowy zapis ku pasującej treści.

Learning from noisy examples teaches correction; during generation, successive steps guide a random representation towards suitable content.

Everyday example

Po opisaniu leśnej ścieżki widzisz, jak generator stopniowo wyłania obraz z losowych plam.

After describing a forest path, you see a generator gradually reveal an image from random patches.

This concerns you if you…

korzystasz z generatorów obrazów lub dźwięków
chcesz zrozumieć stopniowe tworzenie treści
use image or sound generators
want to understand gradual content generation

Examples

obraz · wideo
image · video

Subtopics

Dyfuzja w przestrzeni pikseli (Pixel-space diffusion): Diffusion trained and sampled directly on the pixel grid, without an autoencoder.
- Klepsydrowy transformer dyfuzyjny (Hourglass diffusion transformer): An hourglass-shaped transformer that scales linearly with pixel count.
- Przesunięcie harmonogramu szumu (Noise-schedule shifting for high-resolution pixels): An adjustment of log signal-to-noise ratio as the pixel count grows.
- Transformer pikselowy o dwóch poziomach (Dual-level pixel diffusion transformer): A single-stage diffusion transformer with a patch level and a pixel level.
Dyfuzja w przestrzeni ukrytej (Stable Diffusion) (Latent diffusion (Stable Diffusion)): Diffusion performed in an autoencoder's compressed latent space rather than directly in pixel space, reducing computation; text conditioning can be implemented through cross-attention. This design is used by Latent Diffusion models, including Stable Diffusion.
Dyfuzyjne modele językowe (Diffusion language models): A family of language models that learn to reverse token masking or noising, rather than only next-token prediction.
- Dyfuzja blokowa (Block diffusion language models): An autoregressive distribution over blocks, with discrete diffusion inside a block and causal attention across blocks.
- Dyfuzja ciągła na zanurzeniach słów (Continuous diffusion on word embeddings): Continuous diffusion in embedding space: Gaussian noise on vectors, then rounding back to a token.
- Dyskretna dyfuzja ze score entropy (Score entropy discrete diffusion): Discrete diffusion trained with score entropy, which matches ratios of token probabilities.
- Dyskretna dyfuzja ze strukturalną macierzą przejść (Structured discrete diffusion (D3PM)): Discrete diffusion with a transition matrix: uniform, neighbour-based, or with a mask as an absorbing state.
- Maskowana dyfuzja tekstu (Masked diffusion language models): Diffusion whose only corruption is a mask token: the objective is a weighted mixture of masked-language losses.
  - Próbkowanie z ponownym maskowaniem (Remasking samplers for masked diffusion): Masked-diffusion samplers that can mask previously predicted tokens again and revise them over subsequent steps.
Kaskadowe modele dyfuzyjne (Cascaded diffusion models): A chain of diffusion models at increasing resolution, from sketch to detail.
- Augmentacja obrazu warunkującego (Conditioning augmentation for cascaded diffusion): Corrupting the low-resolution input while training super-resolution models.
Sterowanie promptem bez klasyfikatora (classifier-free guidance) (Classifier-free guidance): A technique to strengthen adherence to the prompt: one model learns to generate with and without the condition (dropped randomly in training), and at sampling time the output is pushed toward the conditional version. The guidance scale is the familiar “prompt fidelity vs diversity” knob.
Szybkie generowanie w kilku krokach (Fast and few-step diffusion sampling): Methods for shortening diffusion sampling: better solvers without retraining and distillation into few-step and one-step models.
- Modele spójności (generowanie jednym krokiem) (Consistency models): Models trained to map any point on a denoising trajectory to its endpoint, enabling one- and few-step sampling; trained by distillation or standalone.
- Przeciwstawna destylacja dyfuzji (generowanie w 1–4 krokach) (Adversarial diffusion distillation): Distilling large diffusion models to 1–4 steps by combining a score-based teacher signal with an adversarial loss.
- Stopniowa destylacja kroków (Progressive distillation of diffusion samplers): Iterative distillation where a student learns to reproduce two steps of the teacher's deterministic sampler in one, halving the step count each round.
- Szybsze samplery bez ponownego treningu (DDIM, solvery ODE) (Fast training-free samplers (DDIM, ODE solvers)): Deterministic or less stochastic sampling procedures for trained diffusion models, based on non-Markovian processes and probability-flow ODE solvers.
Transformery w modelach dyfuzyjnych i przepływowych (DiT) (Transformers for diffusion and flow models (DiT)): Generative architectures in which a transformer processes patches of an image representation instead of a conventional U-Net backbone. DiT applied this design to latent diffusion, while newer variants also use transformers in flow-trained models.
Zaszumianie i odszumianie krok po kroku (DDPM) (Forward noising and learned denoising (DDPM)): The basic construction of denoising diffusion probabilistic models: a forward process gradually adds noise to data, while a model learns to reverse the corruption, often by predicting the added noise. Sampling-step count affects cost and quality, and separate fast-sampling or distillation techniques can reduce it.

Sources

DeepLearning.AI — course catalogue · course catalogue: 42 topic labels with counts, about 130 courses.