Convnextv2 Huge.dbv4 Full

Anime Tagger convnextv2_huge.dbv4-full

Model Details

• Model Type: Multilabel Image classification / feature backbone
• Model Stats:
  • Params: 692.6M
  • FLOPs / MACs: 1.2T / 600.4G
  • Image size: train = 512 x 512, test = 512 x 512
• Dataset: animetimm/danbooru-wdtagger-v4-w640-ws-full
  • Tags Count: 12476
    • General (#0) Tags Count: 9225
    • Character (#4) Tags Count: 3247
    • Rating (#9) Tags Count: 4

Results

#	Macro@0.40 (F1/MCC/P/R)	Micro@0.40 (F1/MCC/P/R)	Macro@Best (F1/P/R)
Validation	0.580 / 0.584 / 0.626 / 0.556	0.697 / 0.696 / 0.692 / 0.701	---
Test	0.580 / 0.584 / 0.627 / 0.556	0.697 / 0.696 / 0.693 / 0.702	0.611 / 0.612 / 0.630

• Macro/Micro@0.40 means the metrics on the threshold 0.40.
• Macro@Best means the mean metrics on the tag-level thresholds on each tags, which should have the best F1 scores.

Thresholds

Category	Name	Alpha	Threshold	Micro@Thr (F1/P/R)	Macro@0.40 (F1/P/R)	Macro@Best (F1/P/R)
0	general	1	0.38	0.685 / 0.673 / 0.697	0.457 / 0.514 / 0.430	0.494 / 0.490 / 0.524
4	character	1	0.51	0.946 / 0.962 / 0.930	0.930 / 0.948 / 0.915	0.943 / 0.959 / 0.930
9	rating	1	0.24	0.828 / 0.790 / 0.871	0.833 / 0.823 / 0.843	0.835 / 0.812 / 0.861

• Micro@Thr means the metrics on the category-level suggested thresholds, which are listed in the table above.
• Macro@0.40 means the metrics on the threshold 0.40.
• Macro@Best means the metrics on the tag-level thresholds on each tags, which should have the best F1 scores.

For tag-level thresholds, you can find them in selected_tags.csv.

How to Use

We provided a sample image for our code samples, you can find it here.

Use TIMM And Torch

Install dghs-imgutils, timm and other necessary requirements with the following command

pip install 'dghs-imgutils>=0.19.0' torch huggingface_hub timm pillow pandas

After that you can load this model with timm library, and use it for train, validation and test, with the following code

import json

import pandas as pd
import torch
from huggingface_hub import hf_hub_download
from imgutils.data import load_image
from imgutils.preprocess import create_torchvision_transforms
from timm import create_model

repo_id = 'animetimm/convnextv2_huge.dbv4-full'
model = create_model(f'hf-hub:{repo_id}', pretrained=True)
model.eval()

with open(hf_hub_download(repo_id=repo_id, repo_type='model', filename='preprocess.json'), 'r') as f:
    preprocessor = create_torchvision_transforms(json.load(f)['test'])
# Compose(
#     PadToSize(size=(512, 512), interpolation=bilinear, background_color=white)
#     Resize(size=(512, 512), interpolation=bicubic, max_size=None, antialias=True)
#     CenterCrop(size=[512, 512])
#     MaybeToTensor()
#     Normalize(mean=tensor([0.4850, 0.4560, 0.4060]), std=tensor([0.2290, 0.2240, 0.2250]))
# )

image = load_image('https://huggingface.co/animetimm/convnextv2_huge.dbv4-full/resolve/main/sample.webp')
input_ = preprocessor(image).unsqueeze(0)
# input_, shape: torch.Size([1, 3, 512, 512]), dtype: torch.float32
with torch.no_grad():
    output = model(input_)
    prediction = torch.sigmoid(output)[0]
# output, shape: torch.Size([1, 12476]), dtype: torch.float32
# prediction, shape: torch.Size([12476]), dtype: torch.float32

df_tags = pd.read_csv(
    hf_hub_download(repo_id=repo_id, repo_type='model', filename='selected_tags.csv'),
    keep_default_na=False
)
tags = df_tags['name']
mask = prediction.numpy() >= df_tags['best_threshold']
print(dict(zip(tags[mask].tolist(), prediction[mask].tolist())))
# {'sensitive': 0.9900546073913574,
#  '1girl': 0.9986221790313721,
#  'solo': 0.9894072413444519,
#  'looking_at_viewer': 0.8689708113670349,
#  'blush': 0.8729097843170166,
#  'smile': 0.9395995736122131,
#  'short_hair': 0.6831153631210327,
#  'long_sleeves': 0.6779903173446655,
#  'brown_hair': 0.802174985408783,
#  'holding': 0.3276722729206085,
#  'dress': 0.6280677318572998,
#  'sitting': 0.6450996994972229,
#  'purple_eyes': 0.8072393536567688,
#  'flower': 0.9524818062782288,
#  'braid': 0.8764650225639343,
#  'outdoors': 0.47000938653945923,
#  'tears': 0.9879008531570435,
#  'floral_print': 0.5994200706481934,
#  'crying': 0.34614139795303345,
#  'plant': 0.3870095908641815,
#  'crown_braid': 0.7048561573028564,
#  'happy_tears': 0.759681224822998,
#  'pavement': 0.2870482802391052,
#  'wiping_tears': 0.9898664951324463,
#  'brick_floor': 0.5737900137901306}

Citation

@misc{convnextv2_huge_dbv4_full,
  title        = {Anime Tagger convnextv2_huge.dbv4-full},
  author       = {narugo1992 and Deep Generative anime Hobbyist Syndicate (DeepGHS)},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/animetimm/convnextv2_huge.dbv4-full}},
  note         = {A large-scale anime-style image classification model based on convnextv2_huge architecture for multi-label tagging with 12476 tags, trained on anime dataset dbv4-full (\url{https://huggingface.co/datasets/animetimm/danbooru-wdtagger-v4-w640-ws-full}). Model parameters: 692.6M, FLOPs: 1.2T, input resolution: 512×512.},
  license      = {gpl-3.0}
}