Type | |
Stats | 877 3,684 |
Reviews | (107) |
Published | Jun 21, 2024 |
Base Model | |
Training | Steps: 8,800 Epochs: 20 |
Usage Tips | Clip Skip: 1 Strength: 1 |
Hash | AutoV2 255C2838A3 |
I really wanted just some nice Portrait / face generator for Cyberpunk style female characters, After 2 lora that I didn't publish came out way better than I thought I decided to take some time and train and see what can be done on PONY, and I really think this came out better than I thought,
Now the training asset was 440 4K in game photomode captures of various characters to train on variety, training set was all female and mostly close up face shots, thats what I was going for, I dont believe this is a general Cyberpunk type lora but when it comes to making Portraits I think it really does well. to train on Civit I took the 4k images and converted them to 2048 dimension since thats the limit and only trained on batch of 1, I think that really enhanced the detail that is available. It seems it can do whatever PONY does do and Cyberpunks it up a little but do understand the training set only had photo mode portraits, mostly face and some torse and full body shots.
Sample images are all original generation, no upscale or inpainting, I think maybe 1 or 2 was inpainted face, but I wanted to show just raw generation, it really does better upscaled or inpainted to improve face. I use foooocus so its very simple to upscale and inpaint.
You can use tags like Cyberpunk, Cyberpunk 2077, in game render or unreal engine to get more render / game like images if default is not doing the best as you might be using a different checkpoint, I think adding 3d, cartoon, anime to the negative improves the quality,
best way to change style is using the fooocus styles they give consistent results imo.