Type | |
Stats | 36 0 |
Reviews | (4) |
Published | Jul 13, 2024 |
Base Model | |
Training | Steps: 4,725 Epochs: 15 |
Usage Tips | Clip Skip: 2 Strength: 0.7 |
Trigger Words | jd_fenix |
Hash | AutoV2 BECBC4346F |
I made JD Fenix (his complete name is James Dominic Fenix which shared name from Marcus's deceased friend, Dominic Santiago) from Gears of War. He's son of Marcus Fenix, a well-known legendary Gear, and his mother Anya Stroud, also a well-known Gear too. JD appear in Gears of War 4 and 5. He's a member of Delta Squad, alongside with Kait and Del. He's trained by Clayton Carmine during his early day as a COG Gear. JD on Gears 4 has different appeareance from Gears 5--he's bald and has beard. In this version, I made JD from Gears of War 4 which he has short hair and no beard.
Trigger Word
jd_fenix
Supporting Trigger Word
In this model there's some trigger word that you can use for creating JD Fenix
//WORDS
jd_fenix --> main prompt
armor --> his armor
pants OR short --> his lower clothing
very short hair --> his short brown hair
//NOTES
Since this LoRA initially scrapped because of innacuracies and retested and has quite good result, this model is stil not 100% to it's full potential or TLDR; Not perfectly 100% create JD and I will update this model later. Here's some author notes regarding this model:
1. You may need use negative prompt eith words that I listed here, esp. armor, to create some flexibility. You can lower weight of LoRA to 0.7 - 0.75 with a caveat of sacrificing facial feature accuracy.
2. Not all SD Checkpoint behave same. In OhMen Checkpoint by jwolf, this model has high accuracy of JD, while CatalystCitron by CitronLegacy didn't meet requirement at all.
Settings I used:
Sampling Method:
DPM++ 2M SDE Karras
DPM++ 2M Karras
CFG Steps: 5 - 8
Sampling Steps: 20
Resolution: 768x920 or 640x768
All of sample image that I use is based on 640x768 with HiRES + Fix applied and upscaled by 1.55x and denoising strength of 0.55 - 0.63 using sampler Latent. Then I upscaled the final image 4x by using ESRGAN-4x, R-ESRGAN 4x, and R-ESRGAN 4x Anime 6B upscaler depending what kind image that I handled.
No inpainting, regional prompter, controlNet I use in all my sample image. All of them were built in from SD Checkpoint + Text Embedding.
Weight: 0.7 - 0.9
NSFW is best suited on 0.7 - 0.75. You can try > 0.75 like 0.8. But you need some work and specific checkpoint in order to work it properly