Fantasy role-playing game Youkengi_realistic(fantasy).

Youke youkengi_realistic(fantasy)_SDXL is the first realistic style large-scale model in the Youke series, also an SDXL large-scale model.

Its core abilities include:

1. Strong affinity to fantasy style: Drawing inspiration from some anime images to go beyond staged poses and real-world constraints in image composition;

2. Appealing 1girl: 1girl has a pure and sweet face, providing a fresh feel for those tired of traditional AI faces, although it strives for flawlessness which is characteristic of AI faces;

3. Not relying on overly flattering terms: Avoiding excessive use of flattering terms that aim solely at enhancing visual quality, except for necessary lighting and shadow terms;

4. Strong textual comprehension: Capable of translating prompt descriptions into images accurately, except for concepts unfamiliar to the model, ensuring most mentioned elements are portrayed in the image;

5. High drawing accuracy: Drawing with minimal errors, significantly reducing overall detail inaccuracies (excluding hands), achieving near-perfect replication even with sample images;

6. Ability to draw individuals of all ages and various ethnicities: Default model represents Asians, while other ethnicities require specific adjustments;

7. Easy to use: Utilizes a hybrid tagging system, similar to SD1.5, providing strong support for both short and long phrases;

8. Doesn't get stuck on faces and has a strong lora integration: A common feature in my second-generation models, achieved by ensuring pollution control during refinement.

9. Naturally strong body performance: Shows good body performance with minimal distortion;

Cons: The model tends to be overexposed, resulting in a natural skin-smoothing effect that may give a 2.9D appearance. Drawing full-body images may lose some realism due to insufficient resolution. Excessive reliance on certain flattering terms or their accumulation can lead to image breakdowns. In such cases, removing some flattering terms is recommended. Under specific tag combinations, there is a probability of image breakdowns (both positive and negative outcomes, but with very low probability). Any minor modifications can resolve such issues. Compared to my other anime models, the hand quality may not be as good since realistic style is less common; thus its proficiency in realistic style is uncertain for now.

Since this is my first realistic style SDXL model, there is no reference rating available, and previous use of other realistic style models has been limited. It’s best to evaluate its performance based on actual usage.

Below are XYZ images:

My other models:

Youke youkengi_anime_SDXL_base:https://www.liblib.ai/modelinfo/4a340e48e0d34d698d0d54d57715b631

Youke youkengi cute:https://www.liblib.ai/modelinfo/ac8fb054451245778b5ea3452104c731

Youke youkengi chibi:https://www.liblib.art/modelinfo/33253de83fcb4465a63cd3c421723cf4?mine=1

Youke youkengi 3D:https://www.liblib.ai/modelinfo/ad6942b5acbc4301a2a2b30a2b4591ae?mine=1

Youke youkengi delicate:https://www.liblib.ai/modelinfo/cdb5d856b52d49329a481f552d48c4e6?mine=1

Youke youkengi anything:https://www.liblib.ai/modelinfo/b0466fd9edd34ea7a67dab412db47fb3

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

Youke youkengi delicate if: https://www.liblib.ai/modelinfo/19c85fcbdf424520b0fb70613da88ff6?mine=1

Regarding usage:

1. VAE has been integrated: The current VAE offers slightly higher saturation and contrast than the default SDXL VAE, providing the option to override the model's VAE effects;

2. Sampling methods are versatile with common options yielding satisfactory results. Testing was conducted in the DPM++ 2M SDE Karras environment, known for its superior sampling performance.

3. Recommended high-resolution restoration (hires. fix) enlargement, with redraw magnitude around 0.4; the latent enlargement algorithm is not recommended as it may lead to image blurring.

4. Recommended base steps: 25-50 steps, additional steps for high-resolution restoration enlargement: 10-20 steps;

5. Recommended resolutions: 832*1152*2, 1152*832*2, 1024*1024*2, 1152*1152*2. Slight variations are acceptable but avoid using SD1.5 resolutions for drawing, as the training data primarily consists of 4K images, too low base resolution may result in various limb errors.

6. Common negative terms: worst quality, low quality

XL models exhibit strong comprehension, hence simple negative terms are preferred. Complex terms may significantly impact the image quality.

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

Additionally, stay tuned for updates and new models by following and bookmarking this page~