All of my other favorite models can do everything in 4-8 inference steps, so I decided my Anima should do so, also. I made this using the fp8-quantized official release of Anima Base 1.0, the standard Qwen-Image VAE, the turbo LoRA cited on the official Anima model card at Huggingface, and the text encoder as well, of course. Everything is merged into one unit, so you can use a checkpoint loader for it, and all together it's smaller than the original diffusion model alone.

Made in ComfyUI for use in ComfyUI, so YMMV in other interfaces. Works fine with either euler/simple or euler_ancestral/beta57, CFG ~1.0-1.5, and about 4 to 8 inference steps (12 is okay, too). Can handle image sizes in the 512x512 range, 1024x1024, and now even 1536x1536, because Anima's out of the beta training stage, finally.

Despite beliefs to the contrary, this model can easily produce photorealistic images without any need for special LoRAs, and it was trained on tags from Danbooru/Gelbooru, so please, prompt responsibly. And, despite other beliefs to the contrary, you can sell your outputs and stuff you make with them. The non-commercial license for Anima is about the model, not the outputs. they just don't want people hiring out or selling their model or derivatives like this one. It's also why I haven't set the model to allow on-site inference, because that would probably be frowned upon.

On that note, I don't expect any form of remuneration for this. I did it because I felt like sharing something that I've found useful and fun and convenient. As a resource scavenger in good standing, I do this to pay it forward. People here have given me so much, I'd be a total jerk if I didn't share as well. Feel free to re-share, re-upload, whatever you wish, just respect the original creators' license terms, please.