WD Tagger v3 (fastforks)

SmilingWolf's WD v3 anime image taggers: timm ViT / ConvNeXt / EVA-02 trained on Danbooru, 10 861 tags (multi-label).

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5.

Author weight repos: https://huggingface.co/SmilingWolf/wd-convnext-tagger-v3, https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3, https://huggingface.co/SmilingWolf/wd-vit-large-tagger-v3, https://huggingface.co/SmilingWolf/wd-vit-tagger-v3.

Task: classification. Tier: stable. Outputs: logits. Backends: portable, kernels.

Input: one RGB image. Output: ImageNet-1k class logits.

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
wd-convnext-tagger-v3@1 wd-convnext-tagger-v3.safetensors 77ad172eb1f7 Apache-2.0 yes
wd-eva02-large-tagger-v3@1 wd-eva02-large-tagger-v3.safetensors 74f05b0aad86 Apache-2.0 yes
wd-vit-large-tagger-v3@1 wd-vit-large-tagger-v3.safetensors 6aa088240416 Apache-2.0 yes
wd-vit-tagger-v3@1 wd-vit-tagger-v3.safetensors ee78db0cd9f1 Apache-2.0 yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0.

Training data: Danbooru (anime illustrations with community tags; WD taggers: image ids up to 7 220 105).

Licence evidence: https://huggingface.co/SmilingWolf/wd-vit-tagger-v3/blob/7f6b584d0bd3f55c4531f14ba3d4761b2bccdc0f/README.md

Input size

fixed 448x448: the input is resized (short side 448) and centre-cropped to it Tested up to (portable, fp16, RTX 5060 Ti 16 GB): wd-convnext-tagger-v3 448x448 (0.3 GiB peak allocated, 0.4 reserved); wd-eva02-large-tagger-v3 448x448 (0.7 GiB peak allocated, 0.9 reserved); wd-vit-large-tagger-v3 448x448 (0.7 GiB peak allocated, 0.7 reserved); wd-vit-tagger-v3 448x448 (0.2 GiB peak allocated, 0.3 reserved)

Speed and memory

Classification, synthetic

model input time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
wd-convnext-tagger-v3 448x448 15.3 9.60 5.59 4.46 2.2x 1.3x 259 / 292 260 / 292 238 / 276
wd-eva02-large-tagger-v3 448x448 84.4 26.6 18.3 14.1 1.9x 1.3x 668 / 684 669 / 684 673 / 806
wd-vit-large-tagger-v3 448x448 76.8 19.9 12.9 12.5 1.6x 1.0x 662 / 682 669 / 682 667 / 724
wd-vit-tagger-v3 448x448 19.0 5.78 3.98 4.05 1.4x 1.0x 231 / 242 235 / 242 234 / 284

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
wd-convnext-tagger-v3 2.92 (448x448) 2.34 (448x448) 2.50 (448x448) 0.48 (448x448) 0.24 (448x448) 0.07 (448x448)
wd-eva02-large-tagger-v3 4.82 (448x448) 5.38 (448x448) 3.61 (448x448) 0.35 (448x448) 0.06 (448x448) 0.08 (448x448)
wd-vit-large-tagger-v3 4.74 (448x448) 5.54 (448x448) 4.68 (448x448) 0.44 (448x448) 0.09 (448x448) 0.06 (448x448)
wd-vit-tagger-v3 2.33 (448x448) 2.44 (448x448) 2.28 (448x448) 0.32 (448x448) 0.09 (448x448) 0.09 (448x448)

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
wd-convnext-tagger-v3 448x448 5.59 ms 4.45 ms 3.10 ms 4.46 ms 12.2 s 39.7 s 260 / 292 380 / 392 238 / 276 521 / 548
wd-eva02-large-tagger-v3 448x448 18.3 ms 15.7 ms deferred (build >5 min) 14.1 ms 22.4 s – 669 / 684 714 / 724 673 / 806 –
wd-vit-large-tagger-v3 448x448 12.9 ms 13.3 ms 11.2 ms 12.5 ms 13.5 s 36.2 s 669 / 682 713 / 742 667 / 724 1266 / 1290
wd-vit-tagger-v3 448x448 3.98 ms 5.33 ms deferred (build >5 min) 4.05 ms 9.55 s – 235 / 242 282 / 302 234 / 284 –

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 1.5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.015 and p99 <= 0.052 (absolute error of the output logits).

Limitations

  • An image with an alpha channel is not composited on white first, as the reference script does: pass RGB.

Usage

from fastforks import classification

model = classification.load("wd-convnext-tagger-v3")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg")
result.top(5)  # top-5 classes
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: wd-convnext-tagger-v3, wd-eva02-large-tagger-v3, wd-vit-large-tagger-v3, wd-vit-tagger-v3.

Kernels backend

classification.load("wd-convnext-tagger-v3", mode="kernels") runs our CUDA kernels (cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including fastforks/wd-tagger