Type | |
Stats | 1,649 |
Reviews | (185) |
Published | Aug 2, 2023 |
Base Model | |
Training | Steps: 240 Epochs: 80 |
Trigger Words | filmic |
Hash | AutoV2 9CE214ABBE |
Version 1.1:
Version 1.1 is an overtrained version of 1.0. People are more likely to be wearing sci-fi clothes and backgrounds change shape a lot more.
Versions 1.0:
This is an early style lora based on stills from sci fi episodics. Currently, it works well at fixing 21:9 double characters** and adding fog/edge/blur to everything.
Resolution: 1568x672
Trigger: Filmic
Latent Resolution: See Notes
ComfyUI Workflow: Sytan's workflow without the refiner
Sampler: DDIM (DDIM best sampler, fite me)
Notes:
** Latent Image sizes is still an "in-research" topic, so the information surrounding the best settings is uncertain. IF you start with a 1024x1024 latent size and generate a 21:9 image, you will likely get double characters. The solution to this problem is unknown (There was this whole debate on the SAI discord about it), but there are a few proposed solutions:
4096x4096 latent image. This looks to fix it, but may make a "cropped looking" generation?
Latent Image = Image Size. This results in a decent initial generation, but falters when using upscaling algorithms. Leading to over complexity.
Latent Image = 2x Image Size. Results inconclusive, but didn't look great imo.
Just use a Lora trained on 21:9 images. Simple, easy, no technical mumbo jumbo needed!
And it is with Option #4 that I state this fixes double characters. Is it the "Correct" way? no, but there does not look to be a correct way at the moment. So this is a proposed solution to an existing problem.
* This is an early style lora, in some cases it will do some large adjustments to your output.
* Because it uses shots from films, real person likeness could potentially appear. It has not happened yet, but just in case, it can be best to avoid NSFW.