H.R. Giger Arh+版本v3 (ID: 26871)

H.R. Giger Arh+

[EDIT: Retrained and overhauled version 10.0 is up and should now work with most models and LoRAs out there.]

This LoRA model might work or it might not. I did this mainly for my own enjoyment as I wanted a model that specifically does high quality replication of the organic patterns in Giger's airbrush works. I went through my library and scanned in the highest quality images I had, some of them were digital photographs of the real works, so the source material was really high quality for the most part. When it works, it creates reaaally smooth and wide dynamic range renders.

But, as this was my very first time training anything, I probably messed something up as when the LoRA training rendered out test prompts, the images were 100% what I was aiming for - but when I tested the LoRA out on Automatic1111's WebUI, nothing seemed to work or give the wide contrast results - it's probably something in my settings or the implementation of the LoRA network models is not quite there yet. At some point, like magic, it just started working more and more.

Anyhoo, on WebUI, the best results come from somewhat lower CFG range, like 3-5 and with a base model that gives interesting stuff to paint on. I'll dive deeper into the training stuff and was thinking of making a SDv2 model with 768 sample images, but this seems to work ok so far. Test it out and show me what you create with it.

H.R. Giger (1940-2014)

描述:

This one's a BigBoi™. Trained at whopping 1024px using same size sample images, it can create some pretty mad details. Best suited for img2img (which I intended this for) but works really well with various checkpoints (be sure to try this with the Giger checkpoint by Mince >> https://civitai.com/models/12124/hans-rudi-giger-style).

Best is to render at 1024 x 1024 but it can stretch over it. I've been doing test renders at 896 x 1280 and results have been pretty amazing so far. Of course, best case scenario would be to have a checkpoint model underneath that is trained in 1024 from the get-go, but still, this was just a funny science experiment. I'm really tempted to try 1280 training, just for shts n giggles. But this one is already giving loads of details.

训练词语: