Architectures de modèles
Guide des architectures de modèles
Architectures de modèles génératifs
Comprenez les composants des produits de texte, plongements, image, vidéo, audio et multimodaux, puis choisissez consciemment le réseau central, la représentation, le générateur et l’enveloppe système.
22 familles d’architectures · 5 groupes · une page par famille
La distinction essentielle
Une capacité n’est pas une architecture
« Chat », « image » et « multimodal » décrivent une interface. Décodeur seul, espace d’états, DiT et MoE décrivent le calcul interne. RAG et l’usage d’outils décrivent le système d’exécution autour du modèle. Un système en production peut légitimement porter ces trois types d’étiquette.
= capacité de chat
+ architecture à décodeur seul
+ recherche et reclassement
+ contrôleur d’outils typé
+ règles et transfert à un humain
Rédiger un passeport d’architecture
- 1Capacité: Entrées, sorties, tâche utilisateur et preuves requises
- 2Représentation: Jetons, vecteurs, pixels, codes de codec ou variables latentes continues
- 3Réseau central: Encodeur, décodeur, seq2seq, attention, SSM ou hybride
- 4Capacité du modèle: Dense ou routée vers des experts ; paramètres totaux et actifs
- 5Génération: Autorégressive, diffusion/flux, antagoniste ou déterministe
- 6Système: Recherche, outils, validateurs, permissions, supervision et humains
Parcourir les 22 familles
Chaque famille possède une page sur l’entraînement, l’inférence, les atouts, les compromis et les critères de sélection.
Product capabilities
What the product accepts and returns. These are deployment shapes, not mutually exclusive neural architectures.
Text & chat language models
Generate text one token at a time, usually with a causal decoder trained for next-token prediction and then adapted to follow instructions.
Embedding models
Map text, images, audio, users, or items into vectors whose geometry is trained to preserve a useful notion of similarity.
Image generation & editing
Synthesize or transform pixels, most often through a text or image conditioner, a latent generator, and an image decoder.
Video generation
Generate spatial and temporal structure together using frame, patch, or latent video representations with cross-frame computation.
Audio, speech & music models
A family of task-specific pipelines: speech recognition, speech synthesis, audio understanding, voice conversion, sound generation, and music generation are not one architecture.
Core backbones
How information moves through the trainable network. A deployed model may combine several of these choices.
Encoder-only Transformers
Let every input token attend to tokens on both sides, producing contextual representations rather than an open-ended generation loop.
Decoder-only Transformers
Use a causal mask so each position sees only earlier tokens, matching the left-to-right generation process used by most general-purpose chat models.
Encoder–decoder Transformers
Encode an input bidirectionally, then generate an output causally while cross-attending to the encoded source.
Mixture of Experts (MoE)
Replace selected dense sublayers with many expert networks and a learned router that activates only a small subset for each token.
State-space & recurrent models
Update a compact state as tokens arrive instead of retaining an explicit attention relationship between every pair of positions.
Representation & generation
How images, audio, video, and other high-dimensional outputs are represented, learned, and sampled.
Autoencoders, VAEs & learned tokenizers
Learn an encoder that compresses data and a decoder that reconstructs it; variational and quantized variants shape the latent space for sampling or downstream generation.
Generative Adversarial Networks (GANs)
Train a generator to fool a discriminator while the discriminator learns to distinguish generated samples from training data.
Normalizing flows
Transform a simple distribution into a complex data distribution through a sequence of invertible mappings with tractable Jacobians.
Diffusion & score-based generation
Learn to reverse a gradual noising process, producing data by repeatedly transforming noise toward a sample.
Latent diffusion
Run diffusion in an autoencoder’s lower-dimensional latent space, then decode the generated representation back to pixels or another signal.
Diffusion Transformers (DiT)
Use a Transformer over noisy image or video patches as the denoising network inside a diffusion or flow-based generative process.
Autoregressive visual & audio models
Convert images, video, or audio into discrete codes and predict those codes in an ordering, often with a causal Transformer.
Modality interfaces
How vision, text, audio, and other modalities are encoded, aligned, fused, or connected to a generator.
Vision encoders
Transform pixels into feature maps, patch tokens, or pooled vectors that downstream classifiers, retrievers, decoders, or multimodal models can consume.
Contrastive & dual encoders
Encode two inputs independently (such as a query and document or image and caption) and train matching pairs to land near each other.
Multimodal fusion models
Connect modality-specific encoders and generators through projection, cross-attention, shared token spaces, or combinations of these mechanisms.
AI system architectures
Runtime structures around a model that add knowledge, tools, control, and verifiable behavior.
Retrieval-augmented & hybrid systems
Retrieve evidence at request time and provide it to a generator, classifier, or agent instead of relying only on model parameters.
Tool-using & reasoning systems
Place a model inside a controller that can plan, call typed tools, inspect results, revise, verify, stop, and hand control to a person.