An attempt to create a realistic ANIMA model.
Version 2.0
Version 2.0 was primarily tuned for the "er_sde" sampler, steps 12-16.
Run: Euler/simple (Beta), CFG 1, Steps 8-16
Text encoder: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
VAE: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
Description
It is an refined base model, and all example images shown were generated using a turbo LoRA.
FAQ
Comments (11)
Thank you, it's good to have non-turbo checkpoints (as we can add the lora ourselves, or an updated one later) !
more consistent and detailed realism, and mostly fixed posed variety compared to v2.0. nice work.
I think Toya knows what he's doing hehe :)
Why am I not able to train a lora with this model? Works fine using Anima base.
I’ve tested version 2.2.5 with Turbo LoRA, and it’s really good; many of my complex prompts are executed well. Then there’s the realism, color, and contrast—they suit my taste perfectly and can be fine-tuned with the right prompting. The anatomy is incredibly good, too. The model can even handle a bit of text generation. Plus, 10 steps using Turbo LoRA and the er_sdesgm_uniform sampler are sufficient for 1920x1440 images. It also works well with the LoRAs I use. Thanks for it :)
amazing model, but hopefully you can work on addressing the same face syndrome!
Unfortunately, this seems to be an issue with every single photorealistic lora/model related to Anima. I think the number of parameters are just too small.
@Jellai no, that's training issue. The model way smarter than SDXL. It must be something in the training. IMHO mainly the anime base. Anime faces are same. The character name is mostly the haircut. It needs bunch of name tagged faces, same people with different haircuts. Or tagging of face features, like large nose, small nose etc. That actually somewhat works, I suggest experimenting with it.
IMHO it would be best to throw like 100 top celebrities on it. No need to tune it till they are recognizable .. but to get some named archetypes, and especially to show to the model that the name is not the haircut.
@Jellai definitely not the case, as TheodorSid said. I think it's mostly just the anime base training, where faces are nearly all identical in anime, so it probably needs a lot of diversity added. Partly why in my model I'm not trying to train photorealistic, but just semi 'realism', I tried to see what I could do with realism and it just... takes far too much effort. It's really impressive what's been done with SAM!
To get around limits of same face, and poses that you find in lots of "realistic models" you can use an img2img workflow. Create the image in an anime style and then transfer it over to a realistic style: https://civitai.com/articles/31980/sam-anima-anime-to-realistic-image-transformer-with-auto-captioning-and-pure-high-res-fix
Is there some info ? What's supposed to be better ? If anything I find it somewhat more steps hungry and overall less detailed compared to 2.0 non-turbo.