I love this pose, I hope you enjoy it.
【Checkpoint】YesMix
Precautions
(must read) Please disable 'face restoration during generation.
• Some preview images are generated using 'Hires Fix' and enhanced using 'SD Upscale'. Therefore, if you want to generate the preview images, please download the original preview images. •
The maximum generation resolution should not exceed 1024 in theory; Otherwise, extra or deformed heads/limbs will be waiting for you.
Donations: https://www.paypal.me/usuariontroficial
Description
I love this pose, I hope you enjoy it.
【Checkpoint】YesMix
Precautions
(must read) Please disable 'face restoration during generation.
• Some preview images are generated using 'Hires Fix' and enhanced using 'SD Upscale'. Therefore, if you want to generate the preview images, please download the original preview images. •
The maximum generation resolution should not exceed 1024 in theory; Otherwise, extra or deformed heads/limbs will be waiting for you.
Donations: https://www.paypal.me/usuari ontroficial
FAQ
Comments (4)
Sorry, but what's the difference between this and the "concept" one?
Hi, i'm trying to train a dogeza pose lora myself, but it isn't working well. The girls are deformed, have extra limbs or boobs on their backs. Stuff like that.
I hope you could give me some advice?
I have about 150x1 images. Training settings are:
Standard lora, batch size 4, epoch 80 -> 3k steps
cosine with 4 restarts, 5% warmup
TE rate 5e-5, unet 2e-4
128 ranks, 64 alpha
some dropout, clip 2
The images are automatically tagged with wd 1.4
I cleaned up any tags to do with poses (kneeling, sitting, bent over, etc), so the lora absorbs the new pose
Hi
From what I see you have correct parameters, I had that problem with a lora I had previously, but I put the following prompts: (face down), (top-down bottom-up), (naked dogeza).
Try them, and tell me if it works for you.
Another way would be for you to upload your lora, and I can test it to help you.
@Danes34
Thanks for your quick reply :)
You mean put those keywords in the generation prompt or into the dataset?
I will start a new training run soon. I will keep you updated if you like :)
Btw... Do you have any experience training a lora without a trigger word? As I have understood, you can ether use a trigger word and all the stuff you want your lora to learn is absorbed in it, but then you have to use it in your prompt together with the lora.
Or you can train without one and everything will be learned by the model itself and you don't have to remember to put in the right triggers in the prompt.