Anima progression?
Hello everyone, I understand this may sound very selfish and narcissistic, but I was wondering if it would be possible to create a mini-blog, even on Hugging Face, where the developer could share problems and progress on the model's development?
I just think that with the combined efforts of the community, we could help create a wonderful model that's no worse than Illustrious and KREA 2...
If my post seems offensive, I apologize, but I just don't want to see a very promising model for Anime Illustrations go to waste.
There is not much to say. I am continuing to work on various post-training techniques such as RL and distillation. I'm doing this on my local 4x4090 machine, so it's basically free. But any kind of large-scale training, either a continued training of Anima, or training on top of a new base model, requires expensive rented cloud compute and I can't currently afford it. I do make decent money from the commercial licensing, but it's not yet enough to make this a sustainable full-time job where I can self-fund future Anima versions.
So in the meanwhile, I'm working on things like the Turbo variant. From everything I've seen, I believe it's possible to make Turbo as good or even better overall than Base. RL requires sampling from the model during training, so this is another reason to make a really good Turbo version, since it make it much faster to iterate and test various RL methods.
There is not much to say. I am continuing to work on various post-training techniques such as RL and distillation. I'm doing this on my local 4x4090 machine, so it's basically free. But any kind of large-scale training, either a continued training of Anima, or training on top of a new base model, requires expensive rented cloud compute and I can't currently afford it. I do make decent money from the commercial licensing, but it's not yet enough to make this a sustainable full-time job where I can self-fund future Anima versions.
So in the meanwhile, I'm working on things like the Turbo variant. From everything I've seen, I believe it's possible to make Turbo as good or even better overall than Base. RL requires sampling from the model during training, so this is another reason to make a really good Turbo version, since it make it much faster to iterate and test various RL methods.
Thank you for your hard work and for not abandoning the project, I will hope that you can do what you have planned and we will see a new era of anime Checkpoint 😄
There is not much to say. I am continuing to work on various post-training techniques such as RL and distillation. I'm doing this on my local 4x4090 machine, so it's basically free. But any kind of large-scale training, either a continued training of Anima, or training on top of a new base model, requires expensive rented cloud compute and I can't currently afford it. I do make decent money from the commercial licensing, but it's not yet enough to make this a sustainable full-time job where I can self-fund future Anima versions.
So in the meanwhile, I'm working on things like the Turbo variant. From everything I've seen, I believe it's possible to make Turbo as good or even better overall than Base. RL requires sampling from the model during training, so this is another reason to make a really good Turbo version, since it make it much faster to iterate and test various RL methods.
God speed. RL is absolute black magic that is incredibly complicated and compute expensive. But when you finally get it right the improvements are incredible for a model.
Really love Anima and with some of the efforts made by others to expand it, they have shown how much better it could be and how good Anima Base and Turbo are to ecosystem. I genuinely find myself taking from Anima and using its outputs to create Lora for Krea 2 and Illustrious and vice versa.
I really love Anima, thank you!