circlestone-labs

/

Anima

Like 2.37k

Follow

CircleStone Labs 1.07k

I don't know how ugly you got in your rendered output. But I have no problems with Base 1.0

Positive:

masterpiece, best quality, score_7, highres, absurdres, anime screenshot, 1girl

Negative:

worst quality, low quality, score_1, score_2, score_3, score_9

(832x1216, Euler/Euler A)

Increase anime screenshot (or anime coloring) weight to 1.5 or 2.0 if you see too "waifu" style images, and you can also put score_9 in the negative.

Try using artist tags as well so you don't get the default style which can look a bit bland just make sure when you name the artist and put a @ in front of the artist tag. E.g. @reakaakasky,
@ big chungus, etc.

This helps make the image more stable regarding the overall style that you want even on turbo. Definitely check out the different artists as it can help make an image look less AI sloppy.

Just to note of you do use the turbo model The artist tags don't look as good as the base model which is expected as it's a distilled version of the base model that only requires 8 to 12 steps and CFG of 1.

I don't know how ugly you got in your rendered output. But I have no problems with Base 1.0

Positive:

masterpiece, best quality, score_7, highres, absurdres, anime screenshot, 1girl

Negative:

worst quality, low quality, score_1, score_2, score_3, score_9

(832x1216, Euler/Euler A)

Increase anime screenshot (or anime coloring) weight to 1.5 or 2.0 if you see too "waifu" style images, and you can also put score_9 in the negative.

I already know some stuff, but thanks for give these tips ;)

Try using artist tags as well so you don't get the default style which can look a bit bland just make sure when you name the artist and put a @ in front of the artist tag. E.g. @reakaakasky,
@ big chungus, etc.

This helps make the image more stable regarding the overall style that you want even on turbo. Definitely check out the different artists as it can help make an image look less AI sloppy.

Just to note of you do use the turbo model The artist tags don't look as good as the base model which is expected as it's a distilled version of the base model that only requires 8 to 12 steps and CFG of 1.

After some tests, I can tell that anima turbo works better with cfg scale set to "2.0". The model follows the prompt better with this cfg scale strength, and plus, also solve some of the artificial looking issue :)

With Turbo we can't use the negative prompt as the CFG is 1.0 that means the negative prompt is ignored, mathematically. I'm using Turbo too.

However, I have the exact opposite issue. I do love the semirealistic style Anima is doing, instead of the anime style, but I can't control when this comes out, it seems to depend both on the quality tags and the image resolution. When it comes out it is a beautiful work of semirealistic art, resembling a photoreal image. Sure I can use realistic artist tags and they work well enough overall, but it's not the same thing. Tags as "realistic shading, realistic skin, studio lights, high quality" help a lot bringing out the semirealistic style, but don't always work. Any advice please ?

Below there's an example ComfyUI workflow where the semirealistic style comes out. As usual drag the png into Comfy.

update. Rising the CFG to 2.0 does help, some scenes that didn't come out in realistic style now come out fine. This also doubles the rendering time so it comes with a tradeoff.

update. In my experiments I'm doing right now, it also helps a lot to insert a direction for the AI, at the very start of the prompt, in this case we exploit the chatting capabilities of Lumina 2. For this to work we have to bypass the lumina clip, which would inject different directions, and use the standard clip instead. I myself don't use the lumina clip anyway, as it causes artifacts at low resolutions, below 1024. This way we don't have to increase the CFG and we get fast renderings.

This direction below seems to be very effective.
"You are a photographer who will take a realistic picture of the scene with a real camera and realistic shading."

update. I think I'm good, after one day of tests I managed to get all my images in realistic style. Chatting with the AI is good but using the right tags is better, for most scenes it's enough to set a couple tags if the scene description recalls a professional photographic environment. The tremendous speed of the Turbo version certainly helped doing quick tests and fast learning how Anima behaves.

Honestly I'm quite disappointed here, I see the community interaction is quite low, I'm accustomed with much better contributions. Still useful reading comments from others, so thank you.