After half a year, I have finally completed my base model, DHXL. I can confidently say that I have invested a significant amount of time, money, and effort into it. DucHaitenXL is trained from IL 0.1, meaning it is based on the SDXL architecture but features a higher image resolution (1250x1250 at a 1:1 ratio). DHXL has been trained on a massive dataset. Drawing from my previous experience in training PONY models, I have separated different styles into 49 distinct keywords, ranging from style_0 to style_48. However, unlike PONY, these are not quality-based keywords like scores; they are simply priority markers for a specific artistic style. You do not need to use them if you find them unnecessary.
With 49 style trigger keywords, they can be used individually in prompts or combined to create new, unique styles. However, it is best to combine no more than six styles at a time, as stability decreases with more combinations.
Example:
"style_12, style_9, style_22, style_8, style_24, style_37, BREAK"
You can also adjust the influence strength of each style, similar to how you would with LoRA or other keywords, by modifying the weight of the style directly. For example: style_1:0.6 or style_1:1.2.
Note: When combining styles, commas and the keyword BREAK are crucial—do not omit the commas.
Some might ask, "How is this different from LoRA?" The key difference is that LoRA models cannot be used excessively at the same time. They force the checkpoint to conform to their structure, whereas this method functions as a general compromise, allowing for a broader understanding between the base model and user-defined styles.
One issue with this model is that while it has been trained using both natural language and Danbooru keywords, Danbooru is still prioritized. Some character names, poses, and specific actions require exact Danbooru keywords to work correctly. Additionally, the model leans towards generating Asian features. If you aim to create Western-looking characters but keep getting Asian facial features, try specifying distinct Western facial traits more clearly in the prompt.
Recommended Settings:
Sampler: DPM++ 2M
Scheduler: SGM Uniform
Steps: 50
CFG Scale: 8.5
Negative Prompts
For 2D-2.5D images:
(worst quality:1.5), (low quality:1.5), bad anatomy, bad or heavily deformed hands and fingers, extra limbs, missing limbs, conjoined fingers, signature, username, artist name
For 3D-Realistic images:
(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, multiple eyebrows, cropped, extra limb, missing limbs, deformed hands, long neck, long body, signature, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, painting by bad-artist
Detailed Style Breakdown
style_0, style_1, style_2, style_3, style_4 are styles geared towards hyper-realistic 3D and Unreal Engine 5 aesthetics.
style_0 is a highly detailed 3D Unreal style resembling real people, excelling in ultra-close-up facial depictions (though additional keywords are needed in prompts) and offering a variety of angles, especially from behind.
style_4 inherits characteristics from style_0 in creating 3D illusionary characters but leans more towards sensual content with special poses and actions.
style_2 and style_3 both follow a realistic approach. However, style_2 focuses more on RAW quality and performs better when generating multiple people in a single image, whereas style_3 is more photography-oriented, typically featuring a single subject with improved pose execution.
style_1 serves as a balanced fusion of styles 0, 2, 3, and 4, avoiding a full commitment to either realism or 3D aesthetics. Instead, it excels in overall structural detailing.
style_5 leans towards a more distinct 3D material rendering, similar to Blender.
style_6 follows a plastic-like texture style, resembling figures and collectible models.
style_7 is a type of 3D style that emerged from SD1.5.
style_8 adopts an Arcane-inspired movie style, a hybrid 3D-2D aesthetic.
style_9 is a 2.5D style that can also generate both 2D and 3D images, trained using Niji-style data.
style_10 and style_11 both follow a painted style, but they differ in color tones and brushwork. I can’t fully describe their differences in words.
Next, there are multiple styles that I cannot describe clearly, so I will provide images for better comparison and reference:
style_12
style_13
style_14
style_15
style_16
style_17
style_18
style_19
style_20
There are a total of 49 styles, so posting all of them in a single article would be too long. For now, I will only post styles 0 to 20.