Half-Life 2 Concept Art XL版本004 (ID: 548507)

---

If you enjoy the models I make, consider making a post on this model's page with images you've generated with this model so I know that the time I spent training it wasn't entirely for nothing.

---

This is an SDXL version of my Half-Life 2 Concept Art LoRA.
Considering all the talk about Stable Diffusion 3, this LoRA may become obsolete mere days after its release, but oh well.

As with that one, the dataset is fifty images from the concept art drawn for Half-Life 2, except this dataset is 1024x1024, removes several images to add several new, higher-quality ones and this model is a LoCon, so it's more generally usable.

Images produced by this LoRA can look lower-resolution than they actually are. This is more pronounced if you use a weight of 1.5, but will happen sometimes even if you use a weight of 1, so you may as well suck it up and use 1.5, then use img2img with a small denoising scale and no use of the LoRA at all for a dollar store denoising effect.

The dataset for the original model already had some images that were smaller than 512x512, so getting them to be 1024x1024 was quite the challenge.

I tried the naive approach of simply sizing them up, and the results were horrible.

I tried upscaling them, and the results were horrible.

Then, @rickenhausen834 recommended https://www.upscayl.org/ to me, which was much better than the native ComfyUI upscaling I was doing.
This created a pretty decent result, but I noticed that I had to do a lot of fiddling around with the prompting to get half-decent results, and anything outside the dataset of the LoRA had no chance of working well.

So I retrained the dataset but as a LoCon, and the results were much better.

I wanted to recommend Artium v2 for use with this model, but since that model's post is flagged as NSFW, the glowies have hid it from the recommended resources search gg

描述:

训练词语: Half-Life 2 beta,HL2 concept art