PyTorch Image Models

• What's New
• Introduction
• Models
• Features
• Results
• Getting Started (Documentation)
• Train, Validation, Inference Scripts
• Awesome PyTorch Resources
• Licenses
• Citing

What's New

October 24, 2026

• forward_intermediates() and prune_intermediate_layers() are now supported by every model. Newly added for CSPNet, NFNet, DLA, DenseNet (#2870, thanks ghaith-mq), CoaT, ConViT, ConvMixer, CrossViT, HRNet, Inception-v3, Inception-ResNet-v2, NASNet, PNASNet, SelecSLS, SENet, Sequencer2d, VGG, Visformer, Xception and Aligned Xception, with prune_intermediate_layers() for NaFlexViT and MobileNetV5 encoders. CoaT, ConViT, ConvMixer, CrossViT and Visformer gain features_only support.
• prune_intermediate_layers() fixes
  • prune_head=True removes the whole head (head norm, pre-logits layers, attention pooling) so the pruned model leaves no unused params (DDP w/o find_unused_parameters), except CoaT, CrossViT and HRNet when only some of their parallel scales / branches are taken. Pass prune_head=False to keep it.
  • Pruning a hierarchical model to an earlier stage removes the final norm and head sized for the last stage, forward_features() ends at the last kept feature and num_features / head_hidden_size match it. Pruning to the stem feature removes the stem pooling after it.
  • Fixes for grad checkpointing, stop_early, class token / post network removal, and TorchScript of pruned / feature_cls='getter' models across many families.
• head_hidden_size is the actual pre-logits (pooled) width in all models, including concat pooling (catavgmax) and after reset_classifier(). Pool types a model can't apply now raise at creation / reset instead of being silently ignored or failing in forward.
• Other fixes: VOLO num_classes=0 output is the class token embedding (matching forward_head(pre_logits=True)), vit_pe_core reset_classifier() w/ a non map pool removes the attention pool, pool type export via save_for_hf reads the head's current pooling, and headless (num_classes=0) Swin, SwinV2, DaViT and NextViT checkpoints reload correctly.

October 13, 2026

• Hyperparameter sweeps via a new sweep.py: grid / random search over any train.py argument, one effective --config per job, run on a pool of local GPUs and resumable. WSD sweeps train one stable trunk per hyperparameter combination and branch independent cooldowns from its milestone checkpoints, comparing annealing at several budgets for a fraction of the cost of full runs. See the training docs.
• train.py: --resume auto resumes in place from the experiment's own last.pth.tar (checkpoint history, model_best and summary.csv stay consistent), --stop-epoch N stops without changing the schedule, --save-epochs N [N ...] keeps milestone-N.pth.tar snapshots, and YAML --config values are validated against the script's arguments.
• Deprecate PlateauLRScheduler (--sched plateau). timm schedulers are stateless functions of the epoch or update count, the metric-driven plateau schedule never fit that and its state is not checkpointed, so resume was never exact. Use cosine or wsd.
• Docs: new Model API page, generated model family pages for every model module replacing the old per-model pages, and expanded data / loss / models reference pages.
• Model API consistency fixes: reset_classifier() keeps the current pooling in all models, unknown constructor args raise TypeError everywhere, Swin family dynamic_img_size, LeViT layer-wise lr decay, no_weight_decay names that matched no params (NesT, SwinV2, XCiT, FocalNet), resize state recorded for all set_input_size families.
• Fixes: GCViT global query mixed images within a batch, gradient checkpointing for RDNet and MobileNetV5, save_for_hf stale test input size, EfficientViT (MSRA) rectangular resize, ImageDataset image conversion errors (#2846). reparameterize_model moved to timm.models.

October 9, 2026

• Add CSPNeXt (RTMDet backbone) cspnext_{tiny,s,m,l,x} with ImageNet-1k classification weights converted from OpenMMLab. Thanks Munehiro Kobayashi.
• Optimizer improvements
  • Foreach (multi-tensor) impls for RMSpropTF, LaProp, LAMB, LARS, AdaBelief. Foreach steps are chunked by device L2 cache size (TIMM_FOREACH_CHUNK_SIZE to override).
  • RMS update clipping (clipping_threshold) for AdamWLegacy / NAdamW, new decoupled / corrected weight decay NovoGrad variants (novogradw, novogradc).
  • Correctness fixes across many optimizers, including corrected weight decay w/ in-place updated tensor lr, FP16 update stalls at small lr, low precision state handling (Kron, AdaBelief, NovoGrad), and no more in-place p.grad modification.
  • torch.compile / CUDA graph fixes for Adopt, SGDW, AdafactorBV.
  • New optimizer docs covering usage, naming conventions, and families.
• Add WSD (warmup-stable-decay) LR scheduler (--sched wsd, --lr-decay-fn, --lr-stable-timescale).
• Model EMA fixes for bfloat16 / float16 models. ModelEmaV3 bfloat16 EMAs use stochastic rounding by default, float16 models keep a float32 EMA.
• Model resizing via set_input_size()
  • Expanded support for windowed models (MaxViT, CoAtNet, BEiT, EfficientViT-MSRA, relpos ViT, Swin family).
  • Swin / SwinV2 / SwinV2-CR keep each stage's window to grid ratio on resize, a call at the current size is now a no-op. Per-stage window_size / window_ratio supported. Fixes Swin S3 and SwinV2 window16 / window12 models changing outputs on a same-size call.
  • Fix SwinV2 NaN w/ a window size of 1, odd stage feature sizes, and hybrid ViT resize on CUDA / low precision.
• Fix grad accumulation remainder handling (#2821), distributed sampling / repeated augmentation on small datasets, pretrained source overrides and distillation teacher setup, train.py w/ --sched none, and eval center crop interpolation for non-square inputs.
• ONNX export compatibility fixes across PyTorch versions, dropout consistency fixes, and more meta-device init fixes.

September 25, 2026

• Multi-label classification support
  • Add --task multilabel to train.py / validate.py, trained with BCE (MultiLabelClassificationTask). See the training docs for details.
  • Targets can be lists of class indices, dense multi-hot vectors (--target-format multihot), or one binary field per class (--target-key a,b,c), with / paths into nested fields. Supported by the hfds, hfids, tfds (multi-field keys), and wds readers.
  • Mixup / CutMix and NaFlex loaders support multi-label targets. Label smoothing moves each target towards 0.5 and defaults to 0 for multi-label.
  • Validation reports mAP (the default --eval-metric) plus micro, macro, and sample F1, with an F1 threshold set by --multilabel-threshold.
  • New dataset-level evaluators compute metrics over the full validation set across distributed ranks, excluding distributed sampler padding.
  • See the timm multi-label image datasets collection on the Hub for example / benchmark datasets that work out of the box.
• New loss options via --loss
  • Poly-1, Two-way, ZLPR, and Distribution-Balanced losses.
  • Class-frequency weighting for BCE / ASL, computed from label counts gathered by class_weights.py.
  • ASL is now fp16-safe and defaults to batchmean reduction
• RoPE Refactor
  • Extend training-only RoPE coordinate augmentation (shift, jitter, rescale) to axial, mixed, and MRoPE embeddings, sharing helpers with DINOv3 while preserving its existing augmentation behavior.
  • Add grid_type and normalize_coords options to the Fourier / rotary builders (build_fourier_pos_embed, build_rotary_pos_embed) and axial RoPE modules, and configurable MRoPE sections in EVA.
  • Support rotate_half in the tuple-output RotaryEmbedding and the base factory variant.
  • Correct rotation-layout forwarding in EVA, NaFlexViT, and MRoPE attention pooling.
  • Keep non-learned RoPE frequency and coordinate buffers at least float32 in half/bfloat16 models, and compute mixed RoPE phases in float32 under autocast. These precision fixes can change low-precision model outputs.

September 22, 2026

• Add Qwen-Drive-1.0-4B vision tower weights.
• Fix prefetch loader stream reuse/synchronization and AugMix for non-square and single-band images.
• Fix AdaMuon conv LR scaling, AdafactorBigVision clipping, MARS gradient history, and Kron checkpoint loading.
• Fix stale eval attention bias caches, pruned model feature metadata, dilated max-pool padding, and asymmetric loss gradient mode.
• Release 1.0.30

September 10-11, 2026

• Add Qwen3-VL / Qwen3.5 / Qwen3.8 ViT classifier and encoder variants, including classifiers w/ and w/o the native spatial merger.
• Add Sapiens2 ViTs with EVA and NaFlexViT support. Both model additions include native timm Hub weights. Thanks Yonghye Kwon.
• Add iFormer with native timm Hub weights. Thanks Ryan Hou.
• Add EfficientViM with native timm Hub weights. Thanks smellslikeml.
• Add DeepSeek-V4 / V4.1 ViT classifiers and encoders with native timm Hub weights. Thanks Yonghye Kwon.
• Switch default NaFlex SigLIP position interpolation, inference.py input-size selection, and improve --model-kwargs parsing.
• Fix meta-device construction and device/dtype factory support for a few models; move config tests to meta for faster CI, meta context coverage.

August 27, 2026

• Add LowFormer model and weights
• Fix or merge a host of real but small or not commonly encountered bugs surfaced by agents
• Release 1.0.29

August 11, 2026

• Add model defs and pretrained weights for
  • CPUBone, an efficient vision backbone for devices with low parallelization capabilities.
  • PP-LCNetV2.
  • LingBot-Vision.
• Add per-batch image and batch size scheduling for non-NaFlex training, including progressive small-to-large resolution schedules.
• Improve NaFlexViT with key-only attention masks, compile-friendly and memory-efficient position sampling, and cached variable-patch interpolation.
• Improve model factory path and checkpoint handling; fix equal-token positional embedding resampling.

July 10, 2026

• Improve optimizer torch.compile and tensor learning-rate support.
• Extend NaFlexViT patch-layout (for NaFlex-CLAP), and forward_intermediates (NaFlex dict input) support.
• Harden pickle loading and improve custom-label inference.
• Release 1.0.28

May 27, 2026

• Add model defs and pretrained weights for EUPE ViT (DINOv3-style) and ConvNeXt models. See the Efficient Universal Perception Encoder paper.
• Add TIPSv2 model defs and pretrained weights for (DINOv2-style) ViTs. See the TIPSv2 paper.

May 8, 2026

• Release 1.0.27

April 23, 2026

• Add Gemma4 ViT encoders w/ NaFlex pipeline support (variable aspect/size per image). Thanks Yonghye Kwon
• Support DINOv3 weights in NaFlexVit. Thanks Yonghye Kwon
• Some improvements to Muon fallback (AdamW/NadamW) lr behavior

March 23, 2026

• Improve pickle checkpoint handling security. Default all loading to weights_only=True, add safe_global for ArgParse.
• Improve attention mask handling for core ViT/EVA models & layers. Resolve bool masks, pass is_causal through for SSL tasks.
• Fix class & register token uses with ViT and no pos embed enabled.
• Add Patch Representation Refinement (PRR) as a pooling option in ViT. Thanks Sina (https://github.com/sinahmr).
• Improve consistency of output projection / MLP dimensions for attention pooling layers.
• Hiera model F.SDPA optimization to allow Flash Attention kernel use.
• Caution added to SGDP optimizer.
• Release 1.0.26. First maintenance release since my departure from Hugging Face.

Feb 23, 2026

• Add token distillation training support to distillation task wrappers
• Remove some torch.jit usage in prep for official deprecation
• Caution added to AdamP optimizer
• Call reset_parameters() even if meta-device init so that buffers get init w/ hacks like init_empty_weights
• Tweak Muon optimizer to work with DTensor/FSDP2 (clamp_ instead of clamp_min_, alternate NS branch for DTensor)
• Release 1.0.25

Jan 21, 2026

• Compat Break: Fix oversight w/ QKV vs MLP bias in ParallelScalingBlock (& DiffParallelScalingBlock)
  • Does not impact any trained timm models but could impact downstream use.

Jan 5 & 6, 2026

• Release 1.0.24
• Add new benchmark result csv files for inference timing on all models w/ RTX Pro 6000, 5090, and 4090 cards w/ PyTorch 2.9.1
• Fix moved module error in deprecated timm.models.layers import path that impacts legacy imports
• Release 1.0.23

Dec 30, 2025

• Add better NAdaMuon trained dpwee, dwee, dlittle (differential) ViTs with a small boost over previous runs
  • https://huggingface.co/timm/vit_dlittle_patch16_reg1_gap_256.sbb_nadamuon_in1k (83.24% top-1)
  • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.80% top-1)
  • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.67% top-1)
• Add a ~21M param timm variant of the CSATv2 model at 512x512 & 640x640
  • https://huggingface.co/timm/csatv2_21m.sw_r640_in1k (83.13% top-1)
  • https://huggingface.co/timm/csatv2_21m.sw_r512_in1k (82.58% top-1)
• Factor non-persistent param init out of __init__ into a common method that can be externally called via init_non_persistent_buffers() after meta-device init.

Dec 12, 2025

• Add CSATV2 model (thanks https://github.com/gusdlf93) -- a lightweight but high res model with DCT stem & spatial attention. https://huggingface.co/Hyunil/CSATv2
• Add AdaMuon and NAdaMuon optimizer support to existing timm Muon impl. Appears more competitive vs AdamW with familiar hparams for image tasks.
• End of year PR cleanup, merge aspects of several long open PR
  • Merge differential attention (DiffAttention), add corresponding DiffParallelScalingBlock (for ViT), train some wee vits
    • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_in1k
    • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_in1k
  • Add a few pooling modules, LsePlus and SimPool
  • Cleanup, optimize DropBlock2d (also add support to ByobNet based models)
• Bump unit tests to PyTorch 2.9.1 + Python 3.13 on upper end, lower still PyTorch 1.13 + Python 3.10

Dec 1, 2025

• Add lightweight task abstraction, add logits and feature distillation support to train script via new tasks.
• Remove old APEX AMP support

Nov 4, 2025

• Fix LayerScale / LayerScale2d init bug (init values ignored), introduced in 1.0.21. Thanks https://github.com/Ilya-Fradlin
• Release 1.0.22

Oct 31, 2025 🎃

• Update imagenet & OOD variant result csv files to include a few new models and verify correctness over several torch & timm versions
• EfficientNet-X and EfficientNet-H B5 model weights added as part of a hparam search for AdamW vs Muon (still iterating on Muon runs)

Oct 16-20, 2025

• Add an impl of the Muon optimizer (based on https://github.com/KellerJordan/Muon) with customizations
  • extra flexibility and improved handling for conv weights and fallbacks for weight shapes not suited for orthogonalization
  • small speedup for NS iterations by reducing allocs and using fused (b)add(b)mm ops
  • by default uses AdamW (or NAdamW if nesterov=True) updates if muon not suitable for parameter shape (or excluded via param group flag)
  • like torch impl, select from several LR scale adjustment fns via adjust_lr_fn
  • select from several NS coefficient presets or specify your own via ns_coefficients
• First 2 steps of 'meta' device model initialization supported
  • Fix several ops that were breaking creation under 'meta' device context
  • Add device & dtype factory kwarg support to all models and modules (anything inherting from nn.Module) in timm
• License fields added to pretrained cfgs in code
• Release 1.0.21

Sept 21, 2025

• Remap DINOv3 ViT weight tags from lvd_1689m -> lvd1689m to match (same for sat_493m -> sat493m)
• Release 1.0.20

Introduction

PyTorch Image Models (timm) is a collection of image models, layers, utilities, optimizers, schedulers, data-loaders / augmentations, and reference training / validation scripts that aim to pull together a wide variety of SOTA models with ability to reproduce ImageNet training results.

The work of many others is present here. I've tried to make sure all source material is acknowledged via links to github, arxiv papers, etc in the README, documentation, and code docstrings. Please let me know if I missed anything.

Features

Models

All model architecture families include variants with pretrained weights. There are specific model variants without any weights, it is NOT a bug. Help training new or better weights is always appreciated.

• Aggregating Nested Transformers - https://arxiv.org/abs/2105.12723
• BEiT - https://arxiv.org/abs/2106.08254
• BEiT-V2 - https://arxiv.org/abs/2208.06366
• Big Transfer ResNetV2 (BiT) - https://arxiv.org/abs/1912.11370
• Bottleneck Transformers - https://arxiv.org/abs/2101.11605
• CaiT (Class-Attention in Image Transformers) - https://arxiv.org/abs/2103.17239
• CoaT (Co-Scale Conv-Attentional Image Transformers) - https://arxiv.org/abs/2104.06399
• CoAtNet (Convolution and Attention) - https://arxiv.org/abs/2106.04803
• ConvNeXt - https://arxiv.org/abs/2201.03545
• ConvNeXt-V2 - http://arxiv.org/abs/2301.00808
• ConViT (Soft Convolutional Inductive Biases Vision Transformers)- https://arxiv.org/abs/2103.10697
• ConvMixer - https://arxiv.org/abs/2201.09792
• CPUBone - https://arxiv.org/abs/2603.26425
• CrossViT - https://arxiv.org/abs/2103.14899
• CSATv2 - https://huggingface.co/Hyunil/CSATv2

Optimizers

To see full list of optimizers w/ descriptions: timm.optim.list_optimizers(with_description=True)

Included optimizers available via timm.optim.create_optimizer_v2 factory method:

• adabelief an implementation of AdaBelief adapted from https://github.com/juntang-zhuang/Adabelief-Optimizer - https://arxiv.org/abs/2010.07468
• adafactor adapted from FAIRSeq impl - https://arxiv.org/abs/1804.04235
• adafactorbv adapted from Big Vision - https://arxiv.org/abs/2106.04560
• adahessian by David Samuel - https://arxiv.org/abs/2006.00719
• adamp and sgdp by Naver ClovAI - https://arxiv.org/abs/2006.08217
• adamuon and nadamuon as per https://github.com/Chongjie-Si/AdaMuon - https://arxiv.org/abs/2507.11005
• adan an implementation of Adan adapted from https://github.com/sail-sg/Adan - https://arxiv.org/abs/2208.06677
• adopt ADOPT adapted from https://github.com/iShohei220/adopt -

Augmentations

• Random Erasing from Zhun Zhong - https://arxiv.org/abs/1708.04896)
• Mixup - https://arxiv.org/abs/1710.09412
• CutMix - https://arxiv.org/abs/1905.04899
• AutoAugment (https://arxiv.org/abs/1805.09501) and RandAugment (https://arxiv.org/abs/1909.13719) ImageNet configurations modeled after impl for EfficientNet training (https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/autoaugment.py)
• AugMix w/ JSD loss, JSD w/ clean + augmented mixing support works with AutoAugment and RandAugment as well - https://arxiv.org/abs/1912.02781
• SplitBachNorm - allows splitting batch norm layers between clean and augmented (auxiliary batch norm) data

Regularization

• DropPath aka "Stochastic Depth" - https://arxiv.org/abs/1603.09382
• DropBlock - https://arxiv.org/abs/1810.12890
• Blur Pooling - https://arxiv.org/abs/1904.11486

Other

Several (less common) features that I often utilize in my projects are included. Many of their additions are the reason why I maintain my own set of models, instead of using others' via PIP:

• All models have a common default configuration interface and API for
  • accessing/changing the classifier - get_classifier and reset_classifier
  • doing a forward pass on just the features - forward_features (see documentation)
  • these makes it easy to write consistent network wrappers that work with any of the models
• All models support multi-scale feature map extraction (feature pyramids) via create_model (see documentation)
  • create_model(name, features_only=True, out_indices=..., output_stride=...)
  • out_indices creation arg specifies which feature maps to return, these indices are 0 based and generally correspond to the C(i + 1) feature level.
  • output_stride creation arg controls output stride of the network by using dilated convolutions. Most networks are stride 32 by default. Not all networks support this.
  • feature map channel counts, reduction level (stride) can be queried AFTER model creation via the .feature_info member
• All models have a consistent pretrained weight loader that adapts last linear if necessary, and from 3 to 1 channel input if desired
• High performance reference training, validation, and inference scripts that work in several process/GPU modes:
  • NVIDIA DDP w/ a single GPU per process, multiple processes with APEX present (AMP mixed-precision optional)
  • PyTorch DistributedDataParallel w/ multi-gpu, single process (AMP disabled as it crashes when enabled)
  • PyTorch w/ single GPU single process (AMP optional)
• A dynamic global pool implementation that allows selecting from average pooling, max pooling, average + max, or concat([average, max]) at model creation. All global pooling is adaptive average by default and compatible with pretrained weights.
• A 'Test Time Pool' wrapper that can wrap any of the included models and usually provides improved performance doing inference with input images larger than the training size. Idea adapted from original DPN implementation when I ported (https://github.com/cypw/DPNs)
• Learning rate schedulers

Results

Model validation results can be found in the results tables

Getting Started (Documentation)

The official documentation can be found at https://huggingface.co/docs/hub/timm. Documentation contributions are welcome.

Getting Started with PyTorch Image Models (timm): A Practitioner’s Guide by Chris Hughes is an extensive blog post covering many aspects of timm in detail.

timmdocs is an alternate set of documentation for timm. A big thanks to Aman Arora for his efforts creating timmdocs.

paperswithcode is a good resource for browsing the models within timm.

Train, Validation, Inference Scripts

The root folder of the repository contains reference train, validation, and inference scripts that work with the included models and other features of this repository. They are adaptable for other datasets and use cases with a little hacking. See documentation.

Awesome PyTorch Resources

One of the greatest assets of PyTorch is the community and their contributions. A few of my favourite resources that pair well with the models and components here are listed below.

Object Detection, Instance and Semantic Segmentation

• Detectron2 - https://github.com/facebookresearch/detectron2
• Segmentation Models (Semantic) - https://github.com/qubvel/segmentation_models.pytorch
• EfficientDet (Obj Det, Semantic soon) - https://github.com/rwightman/efficientdet-pytorch

Computer Vision / Image Augmentation

• Albumentations - https://github.com/albumentations-team/albumentations
• Kornia - https://github.com/kornia/kornia

Knowledge Distillation

• RepDistiller - https://github.com/HobbitLong/RepDistiller
• torchdistill - https://github.com/yoshitomo-matsubara/torchdistill

Metric Learning

• PyTorch Metric Learning - https://github.com/KevinMusgrave/pytorch-metric-learning

Training / Frameworks

• fastai - https://github.com/fastai/fastai
• lightly_train - https://github.com/lightly-ai/lightly-train

Deployment

• timmx (Export timm models to ONNX, CoreML, LiteRT, TensorRT, and more) - https://github.com/Boulaouaney/timmx

Licenses

Code

The code here is licensed Apache 2.0. I've taken care to make sure any third party code included or adapted has compatible (permissive) licenses such as MIT, BSD, etc. I've made an effort to avoid any GPL / LGPL conflicts. That said, it is your responsibility to ensure you comply with licenses here and conditions of any dependent licenses. Where applicable, I've linked the sources/references for various components in docstrings. If you think I've missed anything please create an issue.

Pretrained Weights

So far all of the pretrained weights available here are pretrained on ImageNet with a select few that have some additional pretraining (see extra note below). ImageNet was released for non-commercial research purposes only (https://image-net.org/download). It's not clear what the implications of that are for the use of pretrained weights from that dataset. Any models I have trained with ImageNet are done for research purposes and one should assume that the original dataset license applies to the weights. It's best to seek legal advice if you intend to use the pretrained weights in a commercial product.

Pretrained on more than ImageNet

Several weights included or references here were pretrained with proprietary datasets that I do not have access to. These include the Facebook WSL, SSL, SWSL ResNe(Xt) and the Google Noisy Student EfficientNet models. The Facebook models have an explicit non-commercial license (CC-BY-NC 4.0, https://github.com/facebookresearch/semi-supervised-ImageNet1K-models, https://github.com/facebookresearch/WSL-Images). The Google models do not appear to have any restriction beyond the Apache 2.0 license (and ImageNet concerns). In either case, you should contact Facebook or Google with any questions.

Citing

BibTeX

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
}

Latest DOI