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
I tried to achieve maximum realism, but the anime style the model essentially forgot it.
FAQ
Comments (20)
You really do amazing work. This model is fantastic, and of course, the other SamRealistic model is another favorite of mine.
Everything after 1.9.5 lost details. Skin is becoming plasticky.
Everyone has their own tastes; some like one thing and others like another. In 1.9 and 1.9.5 the colors and contrast were over-saturated; in subsequent versions I toned down the colors or balance the lighting.
@toya_san That is true. I provided images for comparison in the galleries below. I prefer the results of 1.9.5. Im sure you can see why.
You just have to play with the sampler settings.
er_sde + beta57 gives good skin.
CFG 4-5 and 20-30 steps.
Please continue to release non-turbo versions. I tried to use 2.0 model to add a bit of extra "semi-realism" during highresfix pass and it worked well without destroying artist style.
This ! Non turbo 2.0 is a blast !
I will try to release versions without turbo mode, but I accidentally ruined and deleted the non-turbo version of 2.1, so this version will only be turbo.
@toya_san What improvements/changes were made for 2.1?
@AI_DK Except for a lean towards greater realism, nothing else.
@toya_san it's interesting how your Turbo and non-Turbo don't differ that much. Non Turbo can be used with 12 steps and except for very fine detail it's fine. Certainly good enough for seed hunting. Turbo can be used with CFG 2 and negative, and it benefits from larger step count. So after some testing I don't even miss non turbo 2.1 .. but I don't intend to train.
If there is some things I would focus on, it's skin textures but especially hair textures. The hair is too perfect, too smooth, too fine.
I also want to second the request to continue releasing the base version. I'm not sure what you put in this model but it stabilizes the crap out of my loras and I get ultra fantastic characters and backgrounds.
An amazing model and a lot of recent improvements. Thanks a lot for making it. You might be making ZIT and F9K obsolete ;)
This model is really good, blows everything Z-image has to offer straight out of the water. I've been training LORAs on it (well, on Ultrarealfinetune, then I use it with this) and it handles those really well, too.
Chiming in on the request to please continue releasing base variants. I'm using a workflow where I do the first part of the generation in Base and then finish it with Turbo and it works really well. I can't keep doing that if you don't release base models. :)
Do you have a prompting guide? was the model trained with tags or natural language?
using tags seems to be hit or miss but sometimes produces much better results than pure natural language.
very nice results with that model. nice. thank you
I really liked turbo v2.1
Thanks a bunch for sharing! ♥
Great model. Realism aspect is unmatched, especially at 2b params range. Though, it feels like the model has slightly degraded natural language comprehension than base Anima.
My only wish is better support for Asian face features.
works well love it for non realistic
pretty solid checkpoint for real/semi-real character lora loading. it doesn't try to mess with the faces.
though there seems to be a huge lack in pose variety? comparing exact seeds between anima base and this checkpoint (v2 non turbo) it completely kills all pose variety. i can generate 10 images with a prompt and get nearly 1:1 results save for small detail changes.