This is the latest checkpoint of miso diffusion series. An anime fine tune based on sd3.5m. This version is trained with 1.08m images for 5 epoch. I might do some little adjustment before putting it to training again because as the training images increases, it adds a lot of overhead which slows down the training speed. And as usual I merged only the clip into the model as t5 does little change to the appearance and only adds more censorship to it.You can find out more info on my huggingface: <a target="_blank" rel="ugc" href="https://huggingface.co/suzushi">https://huggingface.co/suzushi</a>where I will be updating regularly. A lot of effort was also spent on cleaning and processing the dataset, I have a danbooru 2025 collection which contains images from start of 2025 till may 25. Fell free to utilize it. Including my own toolkit to process the danbooru data: <a target="_blank" rel="ugc" href="https://github.com/suzushi-tw/wd14-toolkit">https://github.com/suzushi-tw/wd14-toolkit</a> This model has some substantial improvements, images are more clear and sharp. Though the hand issue persists and I don't think this can easily be fixed unless we train it with 3m images for multiple epoch. This time it also has a better understanding of characters as well. In 1.0 the output does not look relevant to the character at all. But this time it can replicate majority of the details.At the moment training an epoch cost around 130 dollars. If you like my project please consider supporting me: <a target="_blank" rel="ugc" href="https://ko-fi.com/suzushi2024">https://ko-fi.com/suzushi2024</a>Lastly, huge thanks to meg who has been supporting this project, without him this project would not have been possible !

Files with SHA256: 6199c6796bca6874ad1a303a9343c93dc9889083e649984fb01e31e679c80848