WARNING: Do not consider anything in this repo production ready.
Checkpoints
- siglip_swinv2_base_2025_02_22_18h56m54s
Text encoder trained on top of frozen SmilingWolf/wd-swinv2-tagger-v3, so pretty much SigLIT style. Compatible with existing DeepGHS indexes/embeddings. - siglip_swinv2_base_2025_05_02_22h02m36s
Based onsiglip_swinv2_base_2025_02_22_18h56m54s, with unfrozen image encoder. So SigLIP with warm start, I guess. - siglip_eva02_base_2025_05_02_21h53m54s
A test with a different architecture, trained from scratch using SigLIP.
Usage Example
See deepghs/search_image_by_image_or_text for example usage.
Compatibility
The checkpoints in this repo have been structured for compatibility with the dghs-imgutils package.
You can run it locally with the following 2 approaches
Infer with Code
pip install dghs-imgutils>=0.17.0
from imgutils.generic import siglip_predict
pred = siglip_predict(
images=[
'https://huggingface.co/datasets/narugo1992/nzb_files/resolve/main/eshuushuu_51.webp',
],
texts=[
# short tags
'1girl',
'1boy',
'orange_hair',
'blue_hair',
# long text
'1girl, solo, thighhighs, orange_shirt, twintails, brown_hair, hair_bun, long_hair, double_bun, '
'zettai_ryouiki, jar, sitting, bow, school_uniform, long_sleeves, smile, pink_bow, skirt, orange_skirt, '
'very_long_hair, black_thighhighs, orange_dress, miniskirt',
'food, halo, red_eyes, side_ponytail, skirt, macaron, pink_hair, sailor_collar, holding, '
'black_sailor_collar, cake, long_hair, drumsticks, black_skirt, pleated_skirt, pink_halo, 1girl, '
'ahoge, red_neckerchief, chibi, neckerchief, long_sleeves, holding_food, sash, blush, holding_drumsticks, '
'multiple_views, white_cardigan, looking_at_viewer,'
],
repo_id='deepghs/siglip_beta',
model_name='smilingwolf/siglip_swinv2_base_2025_02_22_18h56m54s'
)
print(pred)
# [[2.5059912e-02 1.7571157e-04 2.1646977e-03 1.8494057e-04 1.0000000e+00
# 3.8877626e-15]]
Launch a Gradio Demo
pip install dghs-imgutils[demo]>=0.17.0
from imgutils.generic import SigLIPModel
SigLIPModel(
repo_id='deepghs/siglip_beta',
).launch_demo(
default_model_name='smilingwolf/siglip_swinv2_base_2025_02_22_18h56m54s'
)