models
46 memberslucid.modelsLucid 3.0 model zoo — public surface.
Model Families
Vision Models
41 familieslucid.models.visionVision model families — chronological order of publication.
Text Models
5 familieslucid.models.textText model families — Phase 4 of the model zoo.
Generative Models
3 familieslucid.models.generativeImage-generative model families — Phase 5 of the model zoo.
Classes
AutoConfig2 methodsGeneric config loader — return the ModelConfig for any name.
AutoModelAuto-dispatching wrapper that loads the backbone of any family.
AutoModelForCausalLMAuto-dispatching wrapper that loads a causal (left-to-right) LM.
AutoModelForImageClassificationAuto-dispatching wrapper that loads a registered image classifier.
AutoModelForImageGenerationAuto-dispatching wrapper that loads an image-generation model.
AutoModelForMaskedLMAuto-dispatching wrapper that loads a masked-LM model.
AutoModelForObjectDetectionAuto-dispatching wrapper that loads a registered object detector.
AutoModelForQuestionAnsweringAuto-dispatching wrapper that loads a SQuAD-style QA head.
AutoModelForSemanticSegmentationAuto-dispatching wrapper that loads a semantic-segmentation model.
AutoModelForSeq2SeqLMAuto-dispatching wrapper that loads a seq2seq (encoder-decoder) model.
AutoModelForSequenceClassificationAuto-dispatching wrapper that loads a sentence-level classifier.
AutoModelForTokenClassificationAuto-dispatching wrapper that loads a per-token classifier.
ModelConfig4 methodsBase class for every model's configuration dataclass.
PretrainedModel6 methodsBase for every Lucid model that supports the pretrained-checkpoint flow.
BackboneMixin2 methodsMarker mixin for models usable as feature extractors.
ClassificationHeadMixin1 methodsStandard Linear classification head + transfer-learning helper.
DiffusionMixin1 methodsReverse-process sampling loop for diffusion-family generative models.
FeatureInfo1 methodsSpecification of one stage emitted by a feature-extracting backbone.
CausalLMMixin1 methodsDecoder-only autoregressive sampling for causal LM heads.
MaskedLMMixin1 methodsPer-token cross-entropy loss helper for masked-LM / token-class heads.
ModelOutput7 methodsBase for all model-forward output dataclasses.
BaseModelOutput1 methodsGeneric backbone forward output — sequence model variant.
BaseModelOutputWithPooling1 methodsBackbone output + pooled representation (e.g. CLS).
ImageClassificationOutput1 methodsOutput of any {Family}ForImageClassification model.
ObjectDetectionOutput1 methodsOutput of any {Family}ForObjectDetection model.
InstanceSegmentationOutput1 methodsOutput of any instance-segmentation model.
SemanticSegmentationOutput1 methodsOutput of any {Family}ForSemanticSegmentation model.
CausalLMOutput1 methodsOutput of any causal (left-to-right) language model head.
MaskedLMOutput1 methodsOutput of any masked-LM head (BERT, RoFormer, …).
Seq2SeqLMOutput1 methodsOutput of any encoder-decoder seq2seq language model.
DiffusionModelOutput1 methodsSingle-step output of a diffusion U-Net's forward pass.
VAEOutput1 methodsEnd-to-end VAE forward output.
GenerationOutput1 methodsFinal result of a generative model's sampling loop.
BackboneProtocolRefinement of PretrainedModelProtocol for the family's
direct (backbone) model — the class whose name does not
contain For<Task> (e.g. ResNet, BERTModel, ViT,
DDPMModel).
HasConfigClassBuilding-block protocol: model class points at its family's
Config via the config_class ClassVar.
HasFamilyMetaBuilding-block protocol: class has been decorated with
@model_family_meta and so carries the docs-facing metadata
triple at __model_family_meta__.
HasModelTypeBuilding-block protocol: class carries a model_type ClassVar.
ModelConfigProtocolComposite structural contract for a family Config class.
OutputDataclassProtocolForward-return dataclass shape — any dataclass whose name ends
with Output (e.g. ImageClassificationOutput,
MaskedLMOutput). Carries at least an logits /
family-specific tensor field plus an optional loss.
PretrainedModelProtocol2 methodsComposite structural contract for a family's model class —
either the backbone (ResNet, BERTModel) or a task wrapper
(ResNetForImageClassification, BERTForMaskedLM).
TaskWrapperProtocolRefinement of PretrainedModelProtocol for *For<Task>
head classes (e.g. ResNetForImageClassification,
BERTForMaskedLM).
Functions
create_model→ PretrainedModelLook up name and call its factory — the timm-style entry point.
is_model→ boolReturn whether name is registered (case / separator insensitive).
list_models→ list[str]List all registered model names, optionally filtered by task / family.
model_entrypoint→ ModelFactoryReturn the registered factory callable for name.
register_model→ Callable[[ModelFactory], ModelFactory]Decorator that registers a factory function under its __name__.