Anime-Background-Finetuning (10143 manually curated by hand images from danbooru and reddit collections)
The dataset contain roughly 2k of anime Screencap data and 8k of scrapped danbooru illustration data.
This is the proccessed version of the dataset meant to be used for my personal finetuning practice project, please visit my RicemanT/Background-Finetuning repo for the raw unprocessed data that you can process yourself.
The dataset have two minor type of processing being done to it: converting all pngs to lossless webp, and downscaling images with absurd resolution down to 2000x pixel limit using cv2's inter_area for efficient training. It has also been cleanly deduped and wiped of images with wide/weird resolution ratio.
There was a watermark/logo removal phase being planned using a YOLO detection -> Iopaint workflow but after some thinking i've decided agaisnt, due to time constraint and current open source detection model not being reliable enough to nail the job for every image consistently.
The dataset have been tagged by the SOTA anime tagger convnextv2_huge.dbv4-full by animetimm/DeepGHS on the free Google Colab T4 GPU combined with Danbooru scrapped tags from the original image, then used as grounding to caption with Qwen 3.6 27B FP8 (1k image with thinking, 9k image on non-thinking due to compute constraint) using LightningAI free H100 credit.