GenshinCLIP

A simple and small-size open-sourced CLIP model fine-tuned on Genshin Impact's image-text pairs.

Visit the github for case study and data pair examples.

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

Model	Checkpoint Size	Val Loss
GenshinImpact-CLIP-ViT-B-16-laion2B-s34B-b88K	0.59 GB	1.152
GenshinImpact-ViT-SO400M-14-SigLIP-384	3.51 GB	0.362

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-B-16"
# 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-CLIP-ViT-B-16-laion2B-s34B-b88K')
tokenizer = get_tokenizer('hf-hub:mrzjy/GenshinImpact-CLIP-ViT-B-16-laion2B-s34B-b88K')

# 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 /= image_features.norm(dim=-1, keepdim=True)
        text_features /= text_features.norm(dim=-1, keepdim=True)
        text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
        print(text_probs) # [0.0319, 0.0062, 0.0012, 0.9608]

Model Card

CLIP for GenshinImpact

CLIP-ViT-B-16-laion2B-s34B-b88K model further fine-tuned on 17k Genshin Impact English text-image pairs at resolution 384x384.

Training data description

There're currently 17,428 (train) and 918 (validation) text-image pairs used for model training.

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.

Data Distribution:

Validation Loss Curve