Questions on usage

[TomLucidor]

Q1: What if all you have are antique anime video files? How would you manually separate frame and deduplicate? (and maybe upscaling as well) https://github.com/AnyiRao/SceneSeg https://github.com/chaiNNer-org/chaiNNer
Q2: How do you separate different characters given that some characters are just too niche to be in the classifier dataset?
Q3: If the capturing tool mainly focuses on capturing heads, how can this be used to also make an outfit LoRA?
Q4: How do you capture the backside of a character's outfit? Would the detector be able to catch that partially for manual tagging?

[Maximax67]

1. At the moment, my dataset maker downloads a ready-made screencaps for an anime/film/series from fancaps.net and works with it. It does not allow you to make a screencaps from a video. But you can make a storyboard in any way you like and upload your set of pictures to the scrapped folder in your project folder (will be created when you launch dataset maker and choose your project and character name). Then you can skip stage 2 of downloading pictures. If you use Google Colab, it will be more convenient to save the project to Google Drive and upload the images to a folder there.

2. When training the model, I used a hand-collected set of anime faces. The main idea was for the model to learn to recognize different characters in the same anime. Most of the time it does it well. Problems were noticed in "Death Note" where some characters are very similar due to such drawing. I used triplet loss for training and pretrained on ImageNet efficientnet-b0 as model. I did not use any special techniques for separating characters that were too similar but different.

3. It is not yet possible to make outfit LoRA using my dataset maker. Although if your character wears one specific type of clothing, it may be suitable. You can use dataset maker to select pictures of characters who frequently wore the type of clothing you want, and manually remove the wrong pictures.

4. The dataset maker will not be able to select the back of a character’s clothing for you if his face is not visible, since he selects by the faces you need, and not by clothes.

In general, I had an idea to do a similar project, but for outfit LoRA where there would be a selection based on the character’s clothes, and not on his face. But to do this, you need to create a model that would be able to distinguish between the types of clothing the character wears. I don't have time for this now. I'm currently working on a project to create a model that would cluster images by characters. This way I can get a dataset for each character. Not just for one at a time.

If you still have questions, please write.

[TomLucidor]

Although if your character wears one specific type of clothing, it may be suitable. You can use dataset maker to select pictures of characters who frequently wore the type of clothing you want, and manually remove the wrong pictures

Assuming that the character is consistent in wearing N specific outfits, how can this be automated? Clustering and removing outliers? What about cropping of bodies? (since most of the YOLO anime models are there to detect faces)
AniSeg, Inupedia, CartoonSegmentation, and "Segment Anything" being notable exceptions

• https://github.com/jerryli27/AniSeg
• https://github.com/Inupedia/anime-character-image-preprocessing?tab=readme-ov-file#background-removal
• https://github.com/continue-revolution/sd-webui-segment-anything
• https://github.com/CartoonSegmentation/CartoonSegmentation