GenshinCLIP

Simple open-sourced CLIP models fine-tuned on Genshin Impact's image-text pairs.

The models are far from being perfect, but could still offer some better text-image matching performance in some Genshin Impact scenarios.

Model Link

Dataset	Model Name	Link	Checkpoint Size	Val Loss
Game Version 5.0	GenshinImpact-5.0-ViT-SO400M-14-SigLIP-384	Huggingface	3.51 GB	0.519
Game Version 4.x	GenshinImpact-CLIP-ViT-B-16-laion2B-s34B-b88K	Huggingface	0.59 GB	1.152
Game Version 4.x	GenshinImpact-ViT-SO400M-14-SigLIP-384	Huggingface	3.51 GB	0.362

Case Study

• For GenshinImpact-ViT-SO400M-14-SigLIP-384

Case	Image	Multiple Choices	CLIPScore (After Finetune)	CLIPScore (Before Finetune)
1		1) This is an image of Kuki Shinobu 2) This is an image of Ganyu 3) This is an image of Keqing 4) This is an image of Liyue	1) 0.000 2) 0.438 3) 0.000 4) 0.000	1) 1.207e-06 2) 4.144e-08 3) 1.201e-07 4) 2.212e-08
2		1) This is an area of Liyue 2) This is an area of Mondstadt 3) This is an area of Sumeru 4) This is Qingce Village	1) 0.016 2) 0.000 3) 0.001 4) 0.233	1) 0.015 2) 0.003 3) 0.009 4) 0.377
3		1) This is Andrius Wolf of the North 2) This is Stormterror Dvalin 3) This is Guardian of Apep's Oasis	1) 1.000 2) 0.000 3) 0.000	1) 0.001 2) 0.000 3) 0.000
4		1) This is Amakumo Fruit 2) This is Padisarahs 3) This is Naku Weeds	1) 0.000 2) 0.000 3) 0.024	1) 9.425e-07 2) 1.207e-06 3) 1.669e-05

Note: Case 4 is a bad case. (Correct answer is Amakumo Fruit)

Intended uses & limitations

You can use the raw model for tasks like zero-shot image classification and image-text retrieval.

How to use (With OpenCLIP)

Here is how to use this model to perform zero-shot image classification:

import torch
import torch.nn.functional as F
from PIL import Image
import requests
from open_clip import create_model_from_pretrained, get_tokenizer

def preprocess_text(string):
    return "Genshin Impact\n" + string

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# load checkpoint from local path
# model_path = "path/to/open_clip_pytorch_model.bin"
# model_name = "ViT-SO400M-14-SigLIP-384"
# model, preprocess = create_model_from_pretrained(model_name=model_name, pretrained=model_path, device=device)
# tokenizer = get_tokenizer(model_name)

# or load from hub
model, preprocess = create_model_from_pretrained('hf-hub:mrzjy/GenshinImpact-ViT-SO400M-14-SigLIP-384')
tokenizer = get_tokenizer('hf-hub:mrzjy/GenshinImpact-ViT-SO400M-14-SigLIP-384')

# image
image_url = "https://static.wikia.nocookie.net/gensin-impact/images/3/33/Qingce_Village.png"
image = Image.open(requests.get(image_url, stream=True).raw)
image = preprocess(image).unsqueeze(0).to(device)

# text choices
labels = [
    "This is an area of Liyue",
    "This is an area of Mondstadt",
    "This is an area of Sumeru",
    "This is Qingce Village"
]
labels = [preprocess_text(l) for l in labels]
text = tokenizer(labels, context_length=model.context_length).to(device)
with torch.autocast(device_type=device.type):
    with torch.no_grad():
        image_features = model.encode_image(image)
        text_features = model.encode_text(text)
        image_features = F.normalize(image_features, dim=-1)
        image_features = F.normalize(image_features, dim=-1)
        text_features = F.normalize(text_features, dim=-1)
        text_probs = torch.sigmoid(image_features @ text_features.T * model.logit_scale.exp() + model.logit_bias)
        scores = [f"{s:.3f}" for i, s in enumerate(text_probs.tolist()[0])]
        print(scores)  # [0.016, 0.000, 0.001, 0.233]

Model Card

SigLIP for GenshinImpact

SigLIP model further fine-tuned on 15k Genshin Impact English text-image pairs at resolution 384x384.

Training data description

All the images and texts are crawled from Genshin Fandom Wiki and are manually parsed to form text-image pairs.

Image Processing:

• Size: Resize all images to 384x384 pixels to match the original model training settings.
• Format: Accept images in PNG or GIF format. For GIFs, extract a random frame to create a static image for text-image pairs.

Text Processing:

• Source: Text can be from the simple caption attribute of an HTML <img> tag or specified web content.
• Format: Prepend all texts with "Genshin Impact" along with some simple template to form natural language sentences.

For example, here are some training image-text pairs:

Image	Text
	Genshin Impact This is Tainted Water-Splitting Phantasm. Tainted Hydro Phantasm Idle Waving 2
	Genshin Impact This is Annihilation Specialist Mek. Although the Fontaine Research Institute of Kinetic Energy Engineering... ..., state apparatus enforcing the monopoly on violence.
	Genshin Impact Collei Expressions a Smiley expression
	Genshin Impact This is the map of viewpoint Nine Pillars of Peace in Cuijue Slope, Minlin, Liyue Nine shackles of stone were said to have been laid down deep in the valleys of Cuijue Slope to drive off evil and cleanse the world.

Data Distribution:

5.0	4.x

Validation Loss Curve