Fresh and transparent | Youkengi clear anime

New Q group created: 522940538, welcome to join~

The characteristics of this large model are: fresh art style, stronger color affinity, simple backgrounds, and extreme text control capabilities (often considered a unique performance of SDXL).

In terms of text control, it is superior to youkengi anything anime at the same level, with different emphasis.

https://www.liblib.ai/modelinfo/b0466fd9edd34ea7a67dab412db47fb3

In terms of text control, this model is slightly weaker than youkengi delicate anime, with different emphasis.

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

Replaced the pruned and merged version of VAE.

1. Model Features:

1. Fresh art style, stronger color affinity, simple backgrounds as shown in the sample. It can produce a strong combined effect with watercolor and simple sketch styles.

（1）Recommendation: 【dailu】lora【watercolor sketch hybrid style】V1 version, clip1, lora weight 0.8 (too high weight may strain the hand, this model also may not handle it smoothly...)

Recommend using style words to enhance the fresh style: white background, simple background, watercolor sketch, inksketch,

https://www.liblib.ai/modelinfo/7760ac4845be4a9782cfd7982c6c45d1

Example of usage in sample 2, sample 6

（2）Recommendation: 【dailu】lora【simplified watercolor style】, can create a simplified style with macaron color tones

https://www.liblib.ai/modelinfo/71a3337330804883bd2f2c2378b99d2f

Usage example in sample 6

（3）New addition: 【youkengi sketch_line.safetensors】, can add a hand-drawn style, downloadable from the right sidebar.

Designed for clear and delicate2 hand-drawn styles, when used with ordinary lora without trigger words, effects shown in sample 9, with the side effect of simplifying the background.

It is specified for use because applying it to other models may be unstable, even leading to failure. Currently testing it on large sushi particles also yield good effects.

2. Extreme text control capabilities, demonstrated in strong word understanding, exemplified by the following points:

（1）Almost guaranteed full-body images if "fullbody" is included (even for landscape images), including "back to the viewer" almost always generates back view images (almost disregarding landscape or portrait orientations, but deviations from the regular image generation rules may increase distortion rates).

（2）With this model, the effect of prompts alone can almost match some lora outputs.

（3）With precise and stable word descriptions, due to the strong text-to-image control, even when switching seeds, the composition drawn by this model may be very close. In such cases, incorporating words like dynamic pose, girl's pose, which have a certain level of generalization, can effectively enhance compositional variations.

Usage instructions/Generation recommendations:

1. No specific recommended sampling method: not picky, the sample images were drawn using DPM++ 2M SDE Karras;

2. Requires the use of VAE: Recommend using a VAE with higher saturation, kl-f8-anime2.ckpt and vae-ft-mse-840000-ema-pruned.safetensors for drawing;

3. Recommend using high-definition repair redraw with 0.4 amplitude; basic steps recommended are 25~35, high-definition redraw steps 10~20.

4. For portraying delicate art styles: recommend using the 4X UltraSharp upscaling algorithm:

5. More versatile R-ESRGAN 4x+ is recommended, but the model's stability is very strong, not requiring much selection.

6. The only upscaling algorithm not recommended is "latent," as it easily blurs the image; experts using unique effects may try it out.

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

If you are satisfied, please share the results with me~

Lastly, follow and bookmark to receive updates on future versions and new models~