Magi, The Manga Whisperer

TODO

[ ] Release PopCaptions dataset

Table of Contents

1. Magiv1
2. Magiv2
3. Magiv3
4. Datasets

Magiv1

• The model is available at 🤗 HuggingFace Model Hub.
• Try it out for yourself using this 🤗 HuggingFace Spaces Demo (no GPU, so slow).
• Basic model usage is provided below. Inspect this file for more info.

v1 Usage

from transformers import AutoModel
import numpy as np
from PIL import Image
import torch
import os

images = [
        "path_to_image1.jpg",
        "path_to_image2.png",
    ]

def read_image_as_np_array(image_path):
    with open(image_path, "rb") as file:
        image = Image.open(file).convert("L").convert("RGB")
        image = np.array(image)
    return image

images = [read_image_as_np_array(image) for image in images]

model = AutoModel.from_pretrained("ragavsachdeva/magi", trust_remote_code=True).cuda()
with torch.no_grad():
    results = model.predict_detections_and_associations(images)
    text_bboxes_for_all_images = [x["texts"] for x in results]
    ocr_results = model.predict_ocr(images, text_bboxes_for_all_images)

for i in range(len(images)):
    model.visualise_single_image_prediction(images[i], results[i], filename=f"image_{i}.png")
    model.generate_transcript_for_single_image(results[i], ocr_results[i], filename=f"transcript_{i}.txt")

Magiv2

• The model is available at 🤗 HuggingFace Model Hub.
• Try it out for yourself using this 🤗 HuggingFace Spaces Demo (with GPU, thanks HF Team!).
• Basic model usage is provided below. Inspect this file for more info.

v2 Usage

from PIL import Image
import numpy as np
from transformers import AutoModel
import torch

model = AutoModel.from_pretrained("ragavsachdeva/magiv2", trust_remote_code=True).cuda().eval()


def read_image(path_to_image):
    with open(path_to_image, "rb") as file:
        image = Image.open(file).convert("L").convert("RGB")
        image = np.array(image)
    return image

chapter_pages = ["page1.png", "page2.png", "page3.png" ...]
character_bank = {
    "images": ["char1.png", "char2.png", "char3.png", "char4.png" ...],
    "names": ["Luffy", "Sanji", "Zoro", "Ussop" ...]
}

chapter_pages = [read_image(x) for x in chapter_pages]
character_bank["images"] = [read_image(x) for x in character_bank["images"]]

with torch.no_grad():
    per_page_results = model.do_chapter_wide_prediction(chapter_pages, character_bank, use_tqdm=True, do_ocr=True)

transcript = []
for i, (image, page_result) in enumerate(zip(chapter_pages, per_page_results)):
    model.visualise_single_image_prediction(image, page_result, f"page_{i}.png")
    speaker_name = {
        text_idx: page_result["character_names"][char_idx] for text_idx, char_idx in page_result["text_character_associations"]
    }
    for j in range(len(page_result["ocr"])):
        if not page_result["is_essential_text"][j]:
            continue
        name = speaker_name.get(j, "unsure") 
        transcript.append(f"<{name}>: {page_result['ocr'][j]}")
with open(f"transcript.txt", "w") as fh:
    for line in transcript:
        fh.write(line + "\n")

Magiv3

• The model is available at 🤗 HuggingFace Model Hub.
• Basic model usage is provided below. Inspect this file for more info.

<img width="1080" alt="Screenshot 2025-06-26 at 22 52 10" src="https://github.com/user-attachments/assets/3281bd6a-7d65-439c-ad33-7e8c775bca39" />

v3 Usage

model = AutoModelForCausalLM.from_pretrained("ragavsachdeva/magiv3", torch_dtype=torch.float16, trust_remote_code=True).cuda().eval()
processor = AutoProcessor.from_pretrained("ragavsachdeva/magiv3", trust_remote_code=True)

model.predict_detections_and_associations(images, processor)
model.predict_ocr(images, processor)
model.predict_character_grounding(images, captions, processor)

Datasets

Disclaimer: In adherence to copyright regulations, we are unable to publicly distribute the manga images that we've collected. The test images, however, are available freely, publicly and officially on Manga Plus by Shueisha.

Other notes

• Request to download Manga109 dataset here.
• Download a large scale dataset from Mangadex using this tool.
• The Manga109 test splits are available here: detection, character clustering. Be careful that some background characters have the same label even though they are not the same character, see.

License and Citation

The provided models and datasets are available for academic research purposes only.

@InProceedings{magiv1,
    author    = {Sachdeva, Ragav and Zisserman, Andrew},
    title     = {The Manga Whisperer: Automatically Generating Transcriptions for Comics},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {12967-12976}
}

@InProceedings{Sachdeva_2024_ACCV,
    author    = {Sachdeva, Ragav and Shin, Gyungin and Zisserman, Andrew},
    title     = {Tails Tell Tales: Chapter-wide Manga Transcriptions with Character Names},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {December},
    year      = {2024},
    pages     = {2053-2069}
}

@misc{sachdeva2025panelsprosegeneratingliterary,
      title={From Panels to Prose: Generating Literary Narratives from Comics}, 
      author={Ragav Sachdeva and Andrew Zisserman},
      year={2025},
      eprint={2503.23344},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.23344}, 
}