Free transformation | Customize any anime character | Library exclusive for a limited time

A new Q group has been created: 522940538, everyone is welcome to join and have fun~

The main features of this large model are: Strong character lora integration ability, Good stability, Nice basic art style, and Extremely strong text control capabilities (which is usually considered a unique performance of SDXL).

A new Youkengi 3D anime has been released, everyone is welcome to play.

https://www.liblib.ai/modelinfo/ad6942b5acbc4301a2a2b30a2b4591ae?mine=1

In terms of text control capabilities, this model will be slightly weaker than Youkengi delicate anime, with different focuses.

https://www.liblib.ai/modelinfo/cdb5d856b52d49329a481f552d48c4e6

In terms of text control abilities, this model is the same as Youkengi clear anime, with different focuses.

https://www.liblib.ai/modelinfo/e5be47b63e5b4408be94942374ebd57c?mine=1

Replaced the pruned and merged VAE version.

Model features:

1. Strong character lora integration ability, can integrate well with most character loras and change character facial shape and features based on lora's style, referring to examples starting from the third image. Because there is no over-translation in the face, when the facial resolution is insufficient, it may appear blurry, in such cases, it is recommended to use tile map resampling for enlargement.

2. Good stability, not prone to body collapse, naturally good with hands.

3. Nice basic art style, even withou* ***ng character lora, the basic art style is good, the model's original art style is based on the first two example images.

4. Extremely strong text control capabilities, reflected in its strong understanding of words, manifested in the following points:

(1) Nearly always produces full-body images with the addition of "fullbody" prompt (even for horizontal images), adding "back to the viewer" prompt almost always results in a back view image (the model can largely disregard whether the output is horizontal or vertical, but in cases that deviate from the standard output rules, may increase anomaly rate).

(2) With this model, the effect of prompts almost matches some loras' effects.

(3) If your word descriptions are accurate and stable enough, due to the strong text control ability, even switching seeds, the compositional output of the model may be very close. In such cases, try using terms like dynamic pose, girl's pose, which have some generalization, to demonstrate the practical effect of these terms and increase compositional variety.

5. Also created a hand-drawn style lined lora, direct downloads available in the right file column without any trigger words. However, it may slightly alter the art style, the side effect being simplification of the background, similar to example image 15, use as needed. Referred to as exclusive because the effect may be unstable on some models and could even lead to breakdowns.

Usage:

1. No specific sampling method recommended: Examples were created using DPM++ 2M SDE Karras;

2. VAE usage is required: High saturation recommended, such as vae-ft-mse-840000-ema-pruned.safetensors;

3. Recommended HD restoration and redraw intensity of 0.4~0.5; Base steps recommended are 25~50 steps, HD redraw 20 steps.

4. For nuanced art styles: 4X UltraSharp upscaling algorithm recommended.

5. More versatile R-ESRGAN 4x+ can be used, but the model's stability is high, so not much selection is needed.

6. The only upscaling algorithm not recommended is latent, which may easily blur the image, only try for special effects by advanced users.

7. Recommended embeddings: negative_hand Negative Embedding, verybadimagenegative_v1.3

If you are satisfied, please share your feedback with me~

Also, following and bookmarking will keep you informed of updates and new models~