MAIN GOAL
This LoRa is intended to make the anatomy correctness stable. Especially for hands and fingers. PLEASE help me test this LoRa performance and leave me a constructive feedback.
Additional goal is to make it possible to generate hires images directly without upscalers and without getting duplicate objects. This LoRa is trained with Max Resuolution set to 1216px. I tested it in resolution up to 1600x1200.
Description
To train this LoRa I'm using the carefully selected photographs (mostly female models) with hands playing an important role in composition / story telling of the picture. It is supposed that this LoRa should improve the anatomical correctness of rendered images. Also, during tests I noticed that this LoRa adds additional intricate color accents and features making the whole picture more interesting.
Tips
Try opening my images in ComfyUI to see some of my prompts demonstrating the LoRa performance. Please pay attention that I also posted some screenshots from the ComfyUI to visually demonstrate the effect of the LoRa. These screenshot usually don't have a ComfyUI workflow embedded. Os please don't request the prompts for screenshots please.
Tests
I'm testing this LoRa with CinEro XL v1.4 Bokelicious mostly. But it should work with other models as well.
Hosted on Tensor.art
Description
Pony version of CinEro Lora for Feminine Portraits improvement (anatomy, light, depth, colors).
FAQ
Comments (9)
doesn't work with CinEro Pony, it kills the beautiful cinematographic lighting you've created....
can you fix it?
1. This LoRa already integrated into CinEro Pony 9d with weight < 0.35. So, probably this LoRa won't add a much into the CinEro Pony 9d.
2. You mean that THIS LoRa kills cinematographic flavor of CinEro Pony or vice versa, cinematic style of this LoRa works on other checkpoints and gets destroyed by my CinEro Pony model?
I need more examples so I see precisely what to add.
I can use this LoRa as a base to train a new LoRa, but I need some examples of what style do you want to prepare a corresponding dataset.
@homoludens ah, ok,
I didn't know it was integrated into the model, so I won't be using it.
what are the key words to improve the hands?
Cinero pony is a very good model, I've been looking for one like that for a long time, to make artistic NSFW images.
I love the lighting and colour management which seems pretty easy and maybe the depth of field is still a bit strong, it's sometimes interesting but it tends to make images a bit too blurry sometimes.
Are chiaroscuro and Rembrandt lighting also integrated, as I can't really see the effect?
I'm going to continue testing and I'll come back to you later.
@Suzanne I didn't add a hands / fingers related triggers into captions. After thinking a lot about fingers, I assumed that less than nobody wants to get a bad hands by default and almost more than everybody wants to get good fingers no matter what.
So I decided that the golden path is to train and mix models with good hands and verify during training and merges that fingers don't get damaged.
This is why my models don't usually have hands or fingers triggers.
@Suzanne Did you try adding the words "perfect hands, five fingers" into the prompt?
@homoludens no, but you said there were no keywords for the hands.
Have you tested negative embedding for this?
@Suzanne I just tested it as positive. Seems working. No time yet to make in depth testing with and without these words. The fact that I didn't add these words intentionally to captions, don't exclude a chance that hands was mentioned by the basic auto-captioning tools.
Hey @homoludens
I added "perfect fingers" and "perfect eyes" to my test prompt. It is quite helping.
@dioxidin thank you, I'll try