Sign In

Gigadiffusion LoRA

70
413
2
Type
LoRA
Stats
413
Reviews
Published
Mar 15, 2023
Base Model
SD 1.5
Training
Steps: 1,500
Epochs: 2
Trigger Words
gigapixel upscale of
Hash
AutoV2
12EF2D51F6
default creator card background decoration
jpohhhh's Avatar
jpohhhh

Trained on gigapixel scans of paintings, this LoRA enables poster-sized prints indistinguishable from real art when used with the Gigadiffusion SD Web UI Plugin and Illuminati Diffusion.

Examples are the size of a thumbnail at poster size, and the "after" examples are still only 1/4 the detail of a real Gigadiffusion. Full size output is available here, along with the base image.

Here's some quick tips. The complete workflow is here.

  • I use a weight of 0.35.

  • Use this negative prompt to get well-behaved upscaling with Illuminati Diffusion: (double vision, twins, portrait, people, human, female, person, face, head, mouth, ears, nose:1.6) (blurry, smooth:1.0)*

  • Avoid nfixer/nartfixer/nrealfixer when upscaling, tiles won't look like good art.

* The negative weight is higher than anything I've seen, this snaps it out of digital image mode and into art mode and works around the issues with using it for upscaling. And yet, Illuminati still generates twins 50% of the time. Also, lots of skeletons given the ban on all the human stuff. :)