Models · Classification

Falconsai NSFW ViT

Falconsai's NSFW detector: a ViT-B/16 that classifies an image as normal or nsfw. Apache-2.0 weights.

stable Fork of https://huggingface.co/Falconsai/nsfw_image_detection@96cb0d0 · backends: portable, kernels

Checkpoint falconsai-nsfw-vit

Paper 2023 Added in 0.2.0

Kernels, RTX 5060 Ti 16 GB 1.61 ms 224x224, fp16
vs the faster original 3.6× original fp16: 5.74 ms
Peak memory 198 MiB original: 200 MiB
Accuracy vs original fp32 – no test-set run

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[timm]"

from fastforks import classification

model = classification.load("falconsai-nsfw-vit")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench falconsai-nsfw-vit

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

Input: synthetic, 224x224 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 6.760.85×367 MiB/388 MiB3.40.23
Original fp16.half() on 1 modules, fp16 inputsfp16 5.74–200 MiB/226 MiB3.00.27
Portable fp32fp32not measured ––––
Portablefp16 2.832.0×201 MiB/226 MiB1.80.09
Kernelsfp16 1.613.6×198 MiB/228 MiB2.40.06
Portable + torch.compile 2.022.8×236 MiB/266 MiB8 (build)–
TensorRT fp16fp16 1.125.1×196 MiBenginenot comparable27 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp3218.8 1,121 MiB process–224x224
Portablefp1615.9 1,244 MiB processmatches224x224
Portablefp3217.0 1,084 MiB processmatches224x224
Kernelsfp16not implemented–224x224

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

falconsai-nsfw-vit.safetensors 343 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download falconsai-nsfw-vit).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

Accuracy not measured for falconsai-nsfw-vit.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 1e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.009 and p99 <= 0.011 (absolute error of the output logits).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.00001 not recorded
GPU default dtypes, mean |diff|≤ 0.009 0.00599
GPU default dtypes, p99 |diff|≤ 0.011 0.00719

Recorded for falconsai-nsfw-vit on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

Falconsai's NSFW image detector: ViT-B/16 (ImageNet-21k pretrained) fine-tuned on about 80 000 images, classes normal / nsfw.

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 224) and centre-cropped to it; tested up to 224x224 (211 MiB allocated)
Outputs
logits
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fvm_convfused convolutions for image classifiers
  • fvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)
  • fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)
  • fvm_r3further fused classifier blocks
Apple (MPS)
fp32, fp16

Limitations

  • No accuracy measurement: the weights' training set (Falconsai's private labelled NSFW images) is not published and no public labelled NSFW test set is part of the evaluation data here; tiers are compared by output parity and speed only.
  • The image is stretched to 224 x 224 without keeping its aspect ratio, as the original processor does.
Full reference, generated from the repository

Falconsai's NSFW detector: a ViT-B/16 that classifies an image as normal or nsfw. Apache-2.0 weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
falconsai-nsfw-vitsynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.76
5.74
2.83
1.61
2.02
1.12
367 / 388
200 / 226
201 / 226
198 / 228
236 / 266
393 / 420
Details
  • Original code: https://huggingface.co/Falconsai/nsfw_image_detection at 96cb0d0342c7
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: Apache-2.0
  • Licence evidence: https://huggingface.co/Falconsai/nsfw_image_detection/blob/96cb0d0342c7afb80cab76ecc58b265fa44da256/README.md
  • Training data: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), Falconsai proprietary NSFW data (about 80 000 images, normal / nsfw)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: falconsai-nsfw-vit
  • Outputs: logits; label_set: nsfw-2
  • Input size: fixed 224x224: the input is resized (short side 224) and centre-cropped to it
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): falconsai-nsfw-vit 224x224 (0.2 GiB peak allocated, 0.3 reserved)
  • falconsai-nsfw-vit 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 27 s (ONNX + engine).
  • falconsai-nsfw-vit original 224x224: .half() on 1 modules, fp16 inputs
  • falconsai-nsfw-vit load / first call: original 2.95 s / 0.273 s, portable 2.63 s / 0.094 s, kernels 2.42 s / 0.085 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Parity

CPU, float32: max |diff| <= 1e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.009 and p99 <= 0.011 (absolute error of the output logits).

variantkernels error vs original, mean / p99GPU, torchkernels used
falconsai-nsfw-vit0.00599 / 0.00719RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120

Limitations

  • No accuracy measurement: the weights' training set (Falconsai's private labelled NSFW images) is not published and no public labelled NSFW test set is part of the evaluation data here; tiers are compared by output parity and speed only.
  • The image is stretched to 224 x 224 without keeping its aspect ratio, as the original processor does.

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. normal 100%
  2. nsfw <0.1%
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).