Hello: All
Thank you for sharing your work so that the community can benefit from it and push the boundaries of the model :)
Recommended use:
28 Steps / CFG: 5.0 / Sampler: Euler or Deis / Scheduler: Beta or Karras or normal
IMAGE SIZE : 1024X1024
Best Quality :
masterpiece, best quality, amazing quality, 3D animation, unreal engin-blender, CGI, Cel Shading, ultra-detailed.or
Masterpiece, best quality, amazing quality, Amazing perfect beautiful girl, blush, 8K, HDR, RAW, ISO, 35mm, dynamic angle shoot.Description
FAQ
Comments (11)
Congrats for this illustrious model, I don't find those latin names like "Exitium" everywhere :P , I saw the name in my notifications feedback & I expected it is another XT-404 ✌🏽.
I am still very impressed by the SkyNet Omega WAN model of Exitium, I wish this one is also so revolutionary in its results.
@hazzoom82659 : thank you very much my friend ♥ ;) :)
I am unable to generate a full-body image.
@1372595763568 : That's strange, I can do it perfectly. I suggest you use the following prompt (full body:1.4)
@XT_404 Thank you. I've sorted it out. It's a problem with a certain LoRa device.
This is the most pleasing model so far
@berezka_dude : Thx :)
Interesting Model. Would you share how you made it? There is a unique semirealism in here.
@Tyde001 :
Thank you! Achieving this specific semi-realistic look on SDXL Illustrious was a matter of finding the right balance between 2D structure and 3D texture during the fine-tuning process. Here is a breakdown of the workflow.
The core of this project was the dataset. I used a curated collection of approximately 2,500 high-resolution images, all 1024x1024 or larger. The secret lay in the stylistic mix rather than just the quantity. Specifically, the dataset consisted of 40% high-end digital illustrations to maintain clean lines and the Illustrious aesthetic, and 30% 3D renders from engines like C4D or Unreal Engine 5. This 3D component is crucial for semi-realism as it teaches the model how light interacts with materials like skin, latex, and jewelry. The remaining data included 20% semi-realistic anime art and 10% high-quality cosplay photography to inject real-world skin textures and natural lighting reflections into the latent space.
For the tagging strategy, I employed a hybrid method. Since Illustrious relies heavily on Danbooru-style tags, I retained those for character features. However, I added descriptive natural language to define the general vibe. Keywords such as subsurface scattering, cinematic lighting, photorealistic textures, and highly detailed skin were emphasized in the captions to override the flat 2D look of standard anime models.
Regarding the technical training parameters, I used Illustrious-XL as the base model with Adafactor or Prodigy as the optimizer for stable convergence. The training ran for roughly 15,000 to 20,000 steps with a low learning rate around 1e-6 to avoid deep-frying the images and losing the anime charm. I also trained with a high emphasis on the VAE to ensure the colors stayed vibrant while keeping the shadows deep and realistic. The goal was to create a bridge between flat anime styles and the heavy realism of SDXL Base, creating a unique 2.75D aesthetic.