[2023-9-12] Open Requests for Character LoRA

Update 2023.2.26: My discord server is ready now, you can make requests here: https://discord.gg/eHQd5SVfmJ

After a series of technical iterations, the capabilities of the automated character LoRA training pipeline in version v1.4 have been confirmed. Therefore, to assist more individuals and collect more material for further iterations, with the hope of discovering and addressing potential issues with existing technology, we have decided to officially open requests for character LoRA.

What Characters Are Eligible for Training Requests?

First, let's clarify the scope of characters we are accepting for training:

• Characters should have clear human-like visual features

  • Acceptable characters include: regular human females, females with animal ears, horns, or heterochromia, and other human-like female fantasy characters.

  • Non-acceptable characters include: those without a human-like form, such as monsters, pokemon, robots, animals, etc.

  • We are not currently accepting requests for male characters due to technical complexities, which require further exploration. We will provide updates when we are ready to accept such requests.

• Characters should have a reasonable amount of data sources, such as character tags on image websites and anime videos. Details will be provided later.

What Essential Information Should Requesters Provide?

Under these conditions, we accept LoRA requests from two different sources:

• Web-based Request

  • For individual character (e.g. misaka mikoto, tohsaka rin, klee in genshin impact, etc) requests, requesters need to provide the following information:

    • Official character details, including:

      • The character's English name (or romanized name).

      • An official character image for reference in assessing the quality of LoRA-generated images.

    • Sources of character images, including at least one of the following:

      • Character tags on major image websites (e.g., Danbooru, Zerochan, Pixiv).

      • A self-provided dataset of individual character images (not too few in number).

  • For requests related to an entire series (e.g., Fate Series, Genshin Impact, Touhou, Idol Master, etc), requesters need to provide:

    • The official English name of the series and its common English abbreviation (if available).

    • A wiki site for the series that includes the following information (all are required):

      • A list or index of all characters in the series.

      • The English names (or romanized names) of all characters, along with at least one official character image.

      • The site should be publicly accessible on the internet without too many anti-web-crawling mechanisms.

• Anime-based Request

  • For requests related to anime series (e.g. EVA, A Certain Scientific Railgun, sukasuka, etc), requesters should provide:

    • The official English name of the anime and its common English abbreviation (if available).

    • The official website or wiki page for the anime, which must include the following information:

      • Full English names (or romanized names) of all major characters.

      • Official character images for all major characters.

      • Downloadable magnet links for anime resources, which should ideally meet the following criteria:

        • High-resolution video (1080p is recommended) for improving LoRA quality.

        • Better to avoid using videos with subtitles.

        • Provide the entire series rather than individual episodes, as this ensures an adequate training dataset and improved overall LoRA performance.

Requests will be processed once you provide the necessary information and post your request as a comment of this article. If your request is accepted, you will receive a reply stating "Request Accepted". For requests that are rejected, we will provide an explanation of the reasons. For requests requiring additional information or confirmation, we will reply with specific requirements.

Anything Else to Note?

Here are some important notes:

• For all types of requests:

  • The quantity of character data is crucial. We aim to use no fewer than 200 images for training. We strongly recommend that, for data sources provided by requesters, the final number of collected images after cleaning is strongly recommended to be no less than 100 images. If the number of images is less than 60, the quality of the resulting LoRA cannot be guaranteed, and there is a probability of very poor LoRA performance. The current v1.4 automated training process has been observed to probably produce lower-quality results with limited image quantities. We are actively working to address this issue and anticipate significant improvement in future iterations.

  • The quality of character images is equally important, so requesters should strive to provide high-quality data sources.

  • In the description of the LoRA model eventually published on Civitai, we will open the dataset used for actual training (e.g. example of dataset) and the files of the model at each step of the training process (e.g. example of model).

• For anime-based requests:

  • Requesters do not need to specify which characters to train. We will select all suitable characters for training based on the extraction of the video and then batch train them into LoRA before uploading.

  • In general, main characters and frequently appearing supporting characters with sufficient data will be trained. Characters with not enough screen time will not be trained.

  • We will also release a Hugging Face dataset that clusters characters by visual features for all characters appearing in the anime. This dataset will include preview images and compressed files organized by character, like this example. If you can manually train using a smaller number of images (e.g., 20 images), you can consider using our processed dataset. However, this dataset is not guaranteed to be 100% accurate and may require additional manual adjustments.

  • For anime with fewer than 2 trainable characters, we generally won't conduct training due to economic considerations. However, we will still make the preliminary processed dataset available to assist other trainers in completing LoRA training.

• To avoid potential copyright disputes, we do not accept requests for training that involve original characters created by artists or any other requests that may carry public or legal risks.

• Our team's work is entirely non-commercial, and our GPU resources are limited, with most resources dedicated to computer vision research tasks. Therefore:

  • We will not charge any fees for any accepted requests, and the process is entirely free.

  • Our training data and models will be made fully public and available to the internet community. So we will never accept any form of private requests.

  • We can only commit to completing requests as soon as possible and cannot provide any specific deadlines.

If you initiate a request, it will be considered that you are fully aware of and accept all of the above. Our team will not be held responsible for any potential impacts or consequences arising from this. Your understanding is greatly appreciated.

Furthermore, we extend our sincere gratitude to all those who contribute to the AIGC technology, including all the requesters. Please continue to support our technical endeavors in the future!