9 models · ranked by HuggingFace downloads
by timm
MobileNetV3 small model at 100% width multiplier, trained on ImageNet-1k using the LAMB optimizer via the timm library. At under 3M parameters, it targets image classification on mobile and edge hardware where latency and memory are primary constraints. Part of timm's standardized pretrained model zoo with consistent preprocessing and inference APIs.
by google
Google's ViT-Base (Vision Transformer base model) with 16×16 pixel patch size trained at 224px resolution on ImageNet-21k and fine-tuned on ImageNet-1k. The paper introducing ViTs demonstrated that pure transformer architectures without convolutional inductive bias can match CNNs on image classification when trained on sufficient data. Widely used as a starting backbone for image classification fine-tuning.
by Falconsai
Vision Transformer (ViT) fine-tuned for binary NSFW vs. safe image classification. Provides a single classifier for flagging potentially unsafe image content without category-level labeling. Built on ViT-base architecture and fine-tuned on a curated dataset of safe and unsafe images.
by timm
ViT-Base with 16×16 patches, pre-trained on ImageNet-21k using AugReg (augmentation and regularisation), producing a strong transfer learning backbone before ImageNet-1k fine-tuning. This checkpoint is the pre-fine-tuned version; it is typically fine-tuned downstream rather than used directly for classification. The AugReg recipe significantly improves transfer performance over vanilla ViT-B/16 pre-training.
by microsoft
Swin Transformer V2 Tiny at 4px patch size and 16-patch window for 256px input images. Swin V2 improves over V1 with log-spaced continuous position bias and cosine attention for better scale transfer. Apache-2.0 licensed and available via standard Transformers image-classification pipeline.
by Freepik
Freepik's NSFW image classifier built on a timm-wrapped backbone for binary or multi-class content safety detection. MIT-licensed for integration into content moderation pipelines. Trained by Freepik, a major stock media platform, likely on production-scale labeled data.
by timm
ConvNeXt-Femto is the smallest variant in the ConvNeXt family, pre-trained on ImageNet-1K using the timm library's distillation training (d1). At femto scale it's designed for extreme compute efficiency with minimal accuracy. Apache-2.0 licensed and available via timm's standard model registry.
by timm
EdgeNeXt-Small is a lightweight CNN-transformer hybrid architecture optimized for mobile and edge inference, pre-trained on ImageNet-1K with Universal Self-Attention Interaction (USI) training. MIT-licensed and available via timm's model registry.
by Marqo
A fine-tuned image classifier from Marqo that flags adult or explicit content in images at 384px input resolution. It outputs probability scores for NSFW versus safe content and is commonly used as a pre-filter in content moderation pipelines before storing or serving user uploads. Apache 2.0 licensed for commercial deployment.