Because the site refused to let me upload two LoRA files, HL2Beta002 is zipped and categorised as training data.

The dataset for this one is fifty images taken from the concept art drawn for Half-Life 2.

I created this LoRA because I'm a huge fan of 2000s art with colourful lighting and complex backgrounds, and I hate how the average AI image has bright white lighting with solid colours passing for walls in the background.

Features:

• Great skies - especially cloudy and non-blue skies

• Great lighting

• Great pipes and sci-fi machines

• Good "sense of space"

• Good at low-brightness scenes

• Occasionally uses blur for an authentic illustrative effect (mostly around distant windows)

Limitations:

• Not good at bright blue skies since the dataset basically had none - rely on your base model for quality there

• Can kind of be "crowded out" by a strong character prompt and have no effect - in that case, you need to add a strong background prompt as well

Usage Instructions:

This download has two files - HL2Beta001 and HL2Beta002.

They're mostly identical, but HL2Beta002 is purposely a little more deepfried than HL2Beta001, since I was disappointed with 001's results.

What I like to do is use both of them in combination - HL2Beta001 with a weight of 0.5 and HL2Beta002 with a weight of 0.5.

If you want a stronger result, use only HL2Beta002 with a weight of 1.

If you want to mix it with another LoRA, use only HL2Beta002 with a weight of 0.5.

Nature images like to place whatever you want to generate on top of really flat, horrible cliffs. To mitigate this, you might like to put ((quarry, pyramid)) in your negative prompt.

Common keywords:

• Half-Life 2 Beta - shared with all images, good "trigger keyword" for the LoRA

• HL2 concept art - same as above

• city 17 - shared by all images of City 17 - this means green cloudy skies and skyscrapers

• industrial - the more industrial parts of City 17 - use "industrial window" and "brick factory" for a factory-like area

• technology - shared by all images with pipes, Combine technology, cables, etc... I find it more effective to just write "complicated network of pipes"

• air exchange - shared by all images of the Air Exchange - red skies and sunlight with blue shadows, lots of pipes

• Citadel - tall metal tower, seems to be unreliable in practice

• Combine - shared by all images of Combine technology, seems unreliable

• palace - tries to generate tile floors and blue light, seems unreliable

• depot/wasteland - tries to generate railroad bridge and big metal building on top of a cliff

• quarrytown - shared by all images of beta Ravenholm, which at the time was just a sunny mountain town

• seafloor - tries to generate a rough, bumpy wasteland - too small sample size

• nighttime, midnight, dark clouds - gives you a darker sky

• grey-blue sky - gives you a faded blue sky with lots of tiny clouds

• grey-green sky - gives you an apocalyptic green sky with lots of tiny clouds and sometimes a splash of yellow

• grey-red sky - gives you an apocalyptic red sky with lots of tiny clouds

• grey-yellow sky - can give you a nice sunset or a horrible piss-yellow sky at random - use "clouds lit from above by evening sunlight" or "clouds lit from above by early morning sunlight" in your prompts instead

• grey-orange sky - same as "grey-yellow sky"

About the preview images:

The preview images are kind of disingenuous. They're img2img outputs, and basically 1024x1024 recreations of images I generated at 512x512 with txt2img and a different seed. So, it'll be very hard to get an exact recreation. I'll try to leave a comment down below with the 512x512 originals, so you can at least get the seed for both stages. Any images that don't have an equivalent image in the comment already are 512x512.

I'll also try to leave a few comments with different models, since all of the preview images were generated with dreamshaper_4BakedVAE.safetensors.

From my limited testing so far, it seems like other models have similar results, but with different building shapes.