Smilesheep

DNF Spectre Preview (IL) (Patreon)

Published:

2025-11-18 10:16:05

Imported:

2025-11

Tags:

DNF Spectre

Content

This character is a fan-made creation and the first Illustrious (IL) model. Here, I'll share my creation process.

First, I've deployed four environments locally:

1. sd1.5: The earliest model, the beginning of everything. Currently, it's practically unused, and LoRa updates are infrequent. This model is based on a web UI.

2. sdxl-pony: One of the most commonly used large-scale models, with a relatively complete ecosystem. Compared to IL, it has its own advantages and disadvantages. Its advantage is 2.5D+ or realistic. It supports web UI or ComfyUI; I mainly use web UI for my projects.

3. sdxl-Illustrious: Another commonly used large-scale model, with a relatively complete ecosystem. Compared to Pony, it has its own advantages and disadvantages. Its advantage is 2D or 2.5D. Built on ComfyUI, the LoRa community is huge, significantly surpassing Pony.

4. wan2.2: Currently the newest and most popular video model. While it can also generate images from text, it's primarily used for generating videos from images. This model has great potential; AI has developed so rapidly this year!

This IL model took a lot of time to develop, mainly in the following areas:

1. Building a new workflow, including text-to-image and image-to-video generation. Referring to Pony's NSFW parameters, most poses were replicated in IL, with good results. Moreover, it doesn't require NSFW's LoRa, which is very powerful! The advantage of this is that the style of NSFW-generated images tends to be more consistent.

2. Adjusting the secondary enlargement rules for text-to-image generation in the workflow, i.e., a 1280800 image becomes 19201200 after secondary enlargement.

3. Face restoration. Initially, I didn't understand how to do this in text-to-image generation, but thankfully, that's now solved. The next question is...

4. Redrawing parameters. The CFG and redrawing values of the IL model and Pony's model are completely different, so after extensive experimentation, a relatively stable and reasonable value was found. For example, a redraw value of around 3 works best for Pony. However, in IL, a value of around 1.5 is better. For CFG, Pony's value is generally 6-7, while IL's is 1-1.5.

Finally, this work underwent many adjustments during production, so you might notice slight inconsistencies in the images. I redid over 200 retouched images of the characters' faces (using a combination of raw images and face swapping in Photoshop) to resolve the previous issues with the faces' aesthetics.

Later investigation revealed that the problem stemmed from not loading LoRa during the secondary redraw, enlargement, and facial retouching.

In short, I learned a lot from this work. Creating IL images in the future will be much easier; as they say, the beginning is always the hardest.

Recently, subscriptions have increased significantly, resulting in a noticeable rise in revenue. I will continue to work hard to present more and better quality works to repay your support and encouragement!

------------------------------------------------------------------------------------------------------

这期人物是fans定制，也是第一期Illustrious(IL)模型。在这里分享下我的制作历程吧。

首先，我本地已经部署了四套环境，分别是：

1、sd1.5，最早的模型，一切的开始，目前基本上没什么人用了，Lora新增更新也很少，这个模型是基于webui的。

2、sdxl-pony，目前常用的大模型之一，生态比较完善，和IL相比各有优劣势，其优势是2.5d+或真人。支持webui或comfyui，我的主要作品还是习惯用webui。

3、sdxl-Illustrious，目前常用的大模型之一，生态比较完善，和pony相比各有优劣势，其优势是2d或2.5d。依托于comfyui制作，社区lora数量庞大，大幅度领先pony。

4、wan2.2，目前最新最火的视频模型。虽然也可以文生图，但主要还是用于图生视频，这个模型潜力很大，今年AI发展的太快了！

这次的IL模型花了很多时间，主要有以下几个点：

1、一是搭建新的工作流，包括文生图、图生图。参考pony的nsfw参数，在IL里面复刻了大部分的姿势，效果还不错。而且不需要nsfw的lora，很强！这样的优势是，nsfw出图的画风趋于统一。

2、调整工作流中，文生图的二次放大规则，也就是1280800的图，通过二次放大后，变为19201200。

3、人脸修复，前期文生图时，一直没搞明白怎么弄，索性现在是解决了。随之而来的是下个问题。

4、重绘的参数，IL模型和Pony模型的CFG和重绘数值是完全不同的，所以经过大量的实践后，找到了一个相对稳定合理的值。比如pony的重绘值在3左右效果最好。但到了IL中，这个值在1.5左右比较好。而CFG的值pony一般在6-7。而IL的值为1-1.5。